Publications resulting from research conducted using Delta AI appear here. Check back to see how the list of exciting discoveries made using Delta grows.
If you have a publication that should be listed here and isn’t, please share your success with us!
5854943
3NXZNVBX
1
nature
50
default
4596
https://delta.ncsa.illinois.edu/wp-content/plugins/zotpress/
%7B%22status%22%3A%22success%22%2C%22updateneeded%22%3Afalse%2C%22instance%22%3Afalse%2C%22meta%22%3A%7B%22request_last%22%3A150%2C%22request_next%22%3A50%2C%22used_cache%22%3Atrue%7D%2C%22data%22%3A%5B%7B%22key%22%3A%22GCSLIGJ8%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Mohapatra%20et%20al.%22%2C%22parsedDate%22%3A%222026-09-25%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BMohapatra%2C%20R.%2C%20Lancaster%2C%20L.%2C%20Fielding%2C%20D.%2C%20Quataert%2C%20E.%20%26amp%3B%20Bryan%2C%20G.%20L.%20How%20Magnetic%20Fields%20Regulate%20Cooling%20and%20Mixing%20in%20Turbulent%20Radiative%20Mixing%20Layers.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2609.29782%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2609.29782%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22How%20Magnetic%20Fields%20Regulate%20Cooling%20and%20Mixing%20in%20Turbulent%20Radiative%20Mixing%20Layers%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Rajsekhar%22%2C%22lastName%22%3A%22Mohapatra%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Lachlan%22%2C%22lastName%22%3A%22Lancaster%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Drummond%22%2C%22lastName%22%3A%22Fielding%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Eliot%22%2C%22lastName%22%3A%22Quataert%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Greg%20L.%22%2C%22lastName%22%3A%22Bryan%22%7D%5D%2C%22abstractNote%22%3A%22Turbulent%20radiative%20mixing%20layers%20%28TRMLs%29%20are%20expected%20wherever%20hot%20and%20cold%20gas%20move%20past%20one%20another%2C%20including%20in%20the%20solar%20corona%2C%20galactic%20winds%2C%20and%20cold%20filaments%20in%20galaxy%20clusters.%20These%20environments%20are%20often%20magnetized%2C%20but%20magnetic%20effects%20on%20mixing%20and%20cooling%20remain%20less%20well%20understood%20than%20in%20the%20hydrodynamic%20%28HD%29%20case.%20We%20present%20magnetohydrodynamic%20%28MHD%29%20simulations%20of%20TRMLs%20at%20resolutions%20up%20to%20%241024%20%5C%5Ctimes%202048%5E2%24%2C%20spanning%20fields%20aligned%20with%20and%20transverse%20to%20the%20shear%2C%20and%20polarity-reversing%20configurations%20in%20which%20oppositely%20directed%20fields%20form%20current%20sheets%20at%20the%20interface.%20Even%20initially%20weak%20hot-phase%20fields%2C%20with%20%24%5C%5Cmathcal%7BM%7D_%7B%5C%5Crm%20A%2Cshear%7D%5C%5Cequiv%20v_%7B%5C%5Crm%20shear%7D%5C%2Fv_%7B%5C%5Crm%20A%7D%5C%5Csim14%24%2C%20reduce%20the%20mass%20and%20enthalpy%20flux%20and%20radiative%20cooling%20rate%20by%20up%20to%20an%20order%20of%20magnitude%20relative%20to%20HD.%20Magnetic%20tension%20weakens%20turbulent%20motions%2C%20reducing%20both%20the%20diffusion%20of%20hot%20gas%20into%20the%20layer%20and%20the%20folding%20of%20the%20cooling%20surface.%20Transverse%20fields%20suppress%20cooling%20somewhat%20more%20strongly%20than%20shear-aligned%20ones%2C%20although%20a%20transverse%20field%20exerts%20no%20tension%20against%20the%20initial%20linear%20instability.%20Polarity%20reversal%20changes%20the%20morphology%20of%20the%20cooling%20gas%20without%20restoring%20HD-like%20mixing.%20The%20dependence%20on%20the%20Damk%5Cu00f6hler%20number%20is%20similar%20to%20the%20HD%20case%2C%20but%20the%20degree%20of%20suppression%20is%20dependent%20on%20the%20initial%20field%20orientation.%20The%20net%20cooling%20rate%20appears%20resolution-independent%20in%20HD%20but%20declines%20with%20resolution%20in%20MHD%2C%20and%20has%20not%20converged%2C%20so%20our%20suppression%20factors%20are%20lower%20limits.%20Magnetic%20fields%20therefore%20strongly%20regulate%20cooling%20and%20mixing%20in%20multiphase%20gas%2C%20and%20quantitative%20predictions%20require%20careful%20treatment%20of%20transport%20processes.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-09-25T00%3A58%3A55Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2609.29782%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2609.29782%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-27T19%3A28%3A07Z%22%7D%7D%2C%7B%22key%22%3A%22QTYTWD9Z%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Gong%20et%20al.%22%2C%22parsedDate%22%3A%222026-09-21%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BGong%2C%20J.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20SAM-V%3A%20Geometry-Aware%20Segment%20Anything%20for%20Multi-View%20Instance%20Segmentation.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2609.25490%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2609.25490%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22SAM-V%3A%20Geometry-Aware%20Segment%20Anything%20for%20Multi-View%20Instance%20Segmentation%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jiangshan%22%2C%22lastName%22%3A%22Gong%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yuqun%22%2C%22lastName%22%3A%22Wu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Qiqian%22%2C%22lastName%22%3A%22Fu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yao%22%2C%22lastName%22%3A%22Xiao%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Chuhang%22%2C%22lastName%22%3A%22Zou%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shenlong%22%2C%22lastName%22%3A%22Wang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Derek%22%2C%22lastName%22%3A%22Hoiem%22%7D%5D%2C%22abstractNote%22%3A%22Consistent%20multi-view%20object%20segmentation%20is%20critical%20for%203D%20perception%20and%20robotics%2C%20yet%20remains%20challenging%20under%20severe%20viewpoint%20and%20occlusion%20changes.%20Existing%20methods%20typically%20perform%203D%20instance%20segmentation%20on%20point%20clouds%20or%20rely%20on%20offline%202D%20mask-matching%20pipelines.%20However%2C%203D%20instance%20segmentation%20is%20limited%20by%20scarce%203D%20annotations%2C%20while%20offline%202D%20matching%20suffers%20from%20object%20identity%20ambiguity%20across%20frames.%20To%20leverage%20strong%202D%20and%203D%20priors%20jointly%2C%20we%20propose%20SAM-V%20%28Geometry-Aware%20Segment%20Anything%20for%20Multi-View%20Instance%20Segmentation%29.%20Instead%20of%20combining%20the%20two%20priors%20through%20post-hoc%20matching%2C%20SAM-V%20directly%20integrates%20features%20from%20a%20feed-forward%20geometry%20model%20%28VGGT%29%20into%20a%202D%20segmentation%20foundation%20model%20%28SAM%29%2C%20trained%20end-to-end%20for%20cross-view%20instance%20prediction.%20SAM-V%20introduces%20a%20prompt-fusion%20mechanism%20that%20enriches%20sparse%20SAM%20prompt%20tokens%20with%20view-specific%20camera%20tokens%20and%20local%20VGGT%20features%2C%20making%20the%20prompt%20representation%20both%20view-aware%20and%20spatially%20grounded%2C%20together%20with%20a%20mask%20decoder%20that%20attends%20to%20dense%202D%20and%203D%20features.%20By%20conditioning%20the%20mask%20decoding%20directly%20on%20multi-view%20geometry%2C%20SAM-V%20produces%20consistent%20multi-view%20segmentation%20of%20a%20prompted%20object%20in%20a%20single%20forward%20pass%20without%20offline%20mask%20matching%20or%20explicit%203D%20reconstruction.%20On%20the%20IGGT%203D%20tracking%20benchmark%2C%20where%20consistent%20instance%20identity%20across%20frames%20directly%20determines%20performance%2C%20SAM-V%20improves%20overall%20IoU%20by%205%20points%20and%20frame-level%20recall%20by%2012%20points%20on%20the%20ScanNet%2B%2B%20split%20over%20the%20state-of-the-art%20multi-view%20instance%20segmentation%20baseline%20and%20leads%20on%20all%20metrics%20in%20the%20zero-shot%20ScanNet%20split.%20Our%20code%20and%20pretrained%20models%20are%20available%20at%20https%3A%5C%2F%5C%2Fgithub.com%5C%2Fgong208%5C%2FSAM-V.git.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22arXiv%3A2609.25490%22%2C%22date%22%3A%222026-09-21%22%2C%22DOI%22%3A%2210.48550%5C%2FarXiv.2609.25490%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22http%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2609.25490%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-27T19%3A18%3A52Z%22%7D%7D%2C%7B%22key%22%3A%2264XEXA59%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Hategan-Marandiuc%20et%20al.%22%2C%22parsedDate%22%3A%222026-09-16%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BHategan-Marandiuc%2C%20M.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20Toward%20Autonomous%20Radio%20Follow-up%20of%20Multi-messenger%20Transients%20with%20RADAR%3A%20From%20Alert%20Parsing%20to%20Inference%20and%20Observation%20Scheduling.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2609.18233%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2609.18233%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Toward%20Autonomous%20Radio%20Follow-up%20of%20Multi-messenger%20Transients%20with%20RADAR%3A%20From%20Alert%20Parsing%20to%20Inference%20and%20Observation%20Scheduling%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Mihael%22%2C%22lastName%22%3A%22Hategan-Marandiuc%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Tanner%22%2C%22lastName%22%3A%22O%27Dwyer%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Alessandra%22%2C%22lastName%22%3A%22Corsi%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Eliu%22%2C%22lastName%22%3A%22Huerta%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zilinghan%22%2C%22lastName%22%3A%22Li%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ian%20T.%22%2C%22lastName%22%3A%22Foster%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Kyle%22%2C%22lastName%22%3A%22Chard%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ryan%22%2C%22lastName%22%3A%22Chard%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Amal%22%2C%22lastName%22%3A%22Gueroudji%22%7D%5D%2C%22abstractNote%22%3A%22Multi-messenger%20astronomy%20%28MMA%29%2C%20the%20joint%20study%20of%20cosmic%20sources%20through%20gravitational%20waves%20%28GWs%29%2C%20electromagnetic%20%28EM%29%20radiation%2C%20neutrinos%2C%20and%20cosmic%20rays%2C%20is%20rapidly%20reshaping%20time-domain%20astrophysics.%20Realizing%20the%20promise%20of%20MMA%20will%20require%20coordinating%20heterogeneous%20observing%20resources%20and%20automating%20the%20chain%20from%20alert%20to%20analysis%20to%20follow-up.%20RADAR%20%28Radio%20Afterglow%20Detection%20and%20AI-driven%20Response%29%20is%20a%20federated%2C%20privacy-enhancing%20framework%20for%20the%20radio%20follow-up%20of%20GW%20events%2C%20previously%20validated%20on%20GW170817.%20Here%2C%20we%20extend%20it%20along%20three%20axes.%20First%2C%20we%20benchmark%20three%20large%20language%20models%20%28LLMs%3B%20GPT-5.5%2C%20Claude-Opus-4.7%2C%20and%20Gemini-3.5-Flash%29%20against%20the%20GW170817%20radio%20light%20curve%20dataset.%20GPT-5.5%20attains%20the%20highest%20event-level%20%24F_1%24%20score%2C%20the%20harmonic%20mean%20of%20precision%20and%20recall%2C%20at%20%240.893%20%5C%5Cpm%200.010%24%2C%20and%20the%20highest%20GCN-level%20recall%2C%20%240.794%20%5C%5Cpm%200.013%24%2C%20improving%20on%20previous%20GPT-4.1%20results%20by%2016%5C%5C%25%20and%2010%5C%5C%25%2C%20respectively%2C%20while%20Claude-Opus-4.7%20achieves%20the%20highest%20precision%2C%20%240.978%20%5C%5Cpm%200.014%24.%20Second%2C%20we%20introduce%20concurrent%20likelihood%20evaluation%2C%20which%20speeds%20up%20the%20MCMC%20computation%20by%20a%20factor%20of%20%2440%5C%5Ctimes%24%20over%20our%20previous%20results.%20Third%2C%20we%20present%20an%20LLM-driven%20scheme%20that%20converts%20natural-language%20observing%20requests%20into%20submission-ready%20scheduling%20blocks%20for%20the%20Karl%20G.%20Jansky%20Very%20Large%20Array.%20Together%2C%20these%20developments%20advance%20RADAR%20toward%20a%20scalable%2C%20largely%20autonomous%20system%20for%20GW%20radio%20follow%20up.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22arXiv%3A2609.18233%22%2C%22date%22%3A%222026-09-16%22%2C%22DOI%22%3A%2210.48550%5C%2FarXiv.2609.18233%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22http%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2609.18233%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-23T22%3A52%3A55Z%22%7D%7D%2C%7B%22key%22%3A%22XBQHAW68%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Trachu%20et%20al.%22%2C%22parsedDate%22%3A%222026-09-16%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BTrachu%2C%20T.%2C%20Cornell%2C%20S.%2C%20Chen%2C%20W.%20%26amp%3B%20Watanabe%2C%20S.%20LACE%3A%20Layer-Wise%20Compression%20for%20Dynamic%20Frame%20Rate%20Codecs.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2609.17509%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2609.17509%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22LACE%3A%20Layer-Wise%20Compression%20for%20Dynamic%20Frame%20Rate%20Codecs%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Thanapat%22%2C%22lastName%22%3A%22Trachu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Samuele%22%2C%22lastName%22%3A%22Cornell%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22William%22%2C%22lastName%22%3A%22Chen%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shinji%22%2C%22lastName%22%3A%22Watanabe%22%7D%5D%2C%22abstractNote%22%3A%22Neural%20audio%20codecs%20are%20a%20key%20component%20in%20speech%20language%20modeling.%20However%2C%20their%20high%20frame%20rates%20lead%20to%20long%20sequence%20lengths%2C%20increasing%20computational%20costs.%20Dynamic%20frame%20rate%20codecs%20mitigate%20this%20by%20reducing%20the%20effective%20frame%20rate%20using%20a%20compression%20step%20to%20merge%20multiple%20frames%20together.%20However%2C%20most%20prior%20methods%20either%20operate%20on%20single-codebook%20codecs%20or%20apply%20a%20single%20compression%20step%20before%20multi-layer%20quantization.%20This%20forces%20all%20quantization%20layers%20to%20share%20the%20same%20segmentation%20boundaries%2C%20despite%20the%20residual%20embeddings%20at%20different%20quantization%20layers%20exhibiting%20different%20rates%20of%20change%20over%20time.%20We%20propose%20LACE%20%28Layer-Adaptive%20Codec%20Encoding%29%2C%20a%20dynamic%20frame%20rate%20codec%20that%20applies%20an%20independent%20compression%20step%20at%20each%20quantization%20layer%2C%20enabling%20layer-specific%20segmentation%20boundaries.%20To%20use%20LACE%20tokens%20in%20downstream%20text-to-speech%20%28TTS%29%2C%20we%20further%20introduce%20union%20alignment%20and%20boundary%20anchor%20mechanisms%20to%20make%20durations%20consistent%20across%20layers%20while%20preserving%20compression%20benefits.%20Experiments%20on%20LibriTTS%20show%20that%20LACE%20offers%20a%20better%20rate-quality%20tradeoff%20than%20prior%20dynamic%20frame%20rate%20methods%20on%20the%20reconstruction%20task%20and%20improves%20TTS%20inference%20efficiency%20while%20maintaining%20competitive%20synthesis%20quality.%20Our%20code%20is%20released%20as%20part%20of%20the%20ESPnet3%20codec%20recipe.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-09-16T01%3A19%3A06Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2609.17509%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2609.17509%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-23T22%3A49%3A40Z%22%7D%7D%2C%7B%22key%22%3A%22687MJQ99%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Zhu%20et%20al.%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BZhu%2C%20Z.%2C%20Gong%2C%20Y.%2C%20Xiao%2C%20Y.%2C%20Liu%2C%20Y.%20%26amp%3B%20Hoiem%2C%20D.%20How%20to%20Teach%20Large%20Multimodal%20Models%20New%20Skills.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2510.08564%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2510.08564%26lt%3B%5C%2Fa%26gt%3B%20%282025%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22How%20to%20Teach%20Large%20Multimodal%20Models%20New%20Skills%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zhen%22%2C%22lastName%22%3A%22Zhu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yiming%22%2C%22lastName%22%3A%22Gong%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yao%22%2C%22lastName%22%3A%22Xiao%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yaoyao%22%2C%22lastName%22%3A%22Liu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Derek%22%2C%22lastName%22%3A%22Hoiem%22%7D%5D%2C%22abstractNote%22%3A%22How%20can%20we%20teach%20large%20multimodal%20models%20%28LMMs%29%20new%20skills%20without%20erasing%20prior%20abilities%3F%20We%20study%20sequential%20fine-tuning%20on%20five%20target%20skills%20while%20monitoring%20general%20ability%20on%20eight%20held-out%20benchmarks%20across%20three%20model%20families.%20Surprisingly%2C%20we%20find%20that%20performance%20lost%20on%20held-out%20tasks%20after%20fine-tuning%20on%20one%20skill%20can%20partly%20recover%20when%20the%20model%20is%20subsequently%20tuned%20on%20a%20different%20skill.%20We%20trace%20this%20behavior%20to%20a%20measurable%20shift%20in%20the%20output%20token%20distribution%2C%20manifested%20through%20a%20simple%20counting-bias%20probe%20that%20shows%20the%20shift%20co-varies%20with%20forgetting.%20Guided%20by%20this%20insight%2C%20we%20identify%20two%20simple%2C%20robust%20tuning%20recipes%20that%20learn%20strongly%20while%20limiting%20drift%3A%20%28i%29%20updating%20only%20the%20self-attention%20projection%20layers%20%28SA%20Proj.%2C%20%24%5Cu0394%24%20learning%20%2B24.9%20%5C%2F%20%24%5Cu0394%24%20held-out%20forgetting%20-0.6%29%2C%20and%20%28ii%29%20updating%20only%20the%20MLP%20Gate%26amp%3BUp%20while%20freezing%20the%20Down%20projection%20%28%2B30.5%20%5C%2F%20-2.1%29.%20Both%20substantially%20outperform%20full-LLM%20tuning%20%28%2B31.8%20%5C%2F%20-23.3%29%20in%20the%20learning-forgetting%20trade-off.%20We%20also%20compare%20against%20common%20forgetting%20mitigation%20methods%3A%20Learning%20without%20Forgetting%20%28LwF%29%2C%20LoRA%2C%20Mixture-of-Experts%2C%20and%20weight-space%20interpolation%20%28WiSE-FT%29%2C%20and%20find%20that%20our%20selective%20tuning%20recipes%20match%20or%20exceed%20their%20learning-stability%20balance%20while%20remaining%20simpler%2C%20requiring%20no%20replay%2C%20auxiliary%20parameters%2C%20or%20per-stage%20tuning.%20These%20results%20hold%20across%20LLaVA-OneVision%2C%20LLaVA-NeXT%2C%20and%20Qwen2.5-VL%2C%20confirming%20that%20the%20key%20to%20teaching%20LMMs%20new%20skills%20without%20forgetting%20lies%20in%20controlling%20output%20distribution%20shift%20by%20choosing%20which%20components%20to%20tune.%20Code%20will%20be%20made%20available.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222025%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2510.08564%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2510.08564%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-23T22%3A45%3A03Z%22%7D%7D%2C%7B%22key%22%3A%2267SDD539%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Tang%20et%20al.%22%2C%22parsedDate%22%3A%222026-09-11%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BTang%2C%20B.%2C%20Spalding%2C%20Z.%20%26amp%3B%20Cogan%2C%20G.%20B.%20Pretraining%20for%20Sample-Efficient%20Neural%20Interfaces.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2609.13507%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2609.13507%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Pretraining%20for%20Sample-Efficient%20Neural%20Interfaces%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ben%22%2C%22lastName%22%3A%22Tang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zachary%22%2C%22lastName%22%3A%22Spalding%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Gregory%20B.%22%2C%22lastName%22%3A%22Cogan%22%7D%5D%2C%22abstractNote%22%3A%22Brain-computer%20interfaces%20%28BCIs%29%20decode%20neural%20activity%20to%20restore%20lost%20function.%20Typically%2C%20training%20a%20high-performance%20neural%20decoder%20requires%20a%20large%20labeled%20dataset%20to%20be%20collected%20from%20every%20new%20subject.%20One%20way%20to%20reduce%20the%20labeled%20data%20cost%20is%20self-supervised%20pretraining%2C%20which%20learns%20general%20neural%20representations%20from%20unlabeled%20recordings%20that%20accumulate%20across%20subjects.%20However%2C%20for%20intracranial%20electroencephalography%20%28iEEG%29%20recordings%2C%20self-supervised%20learning%20has%20been%20challenging%20due%20to%20differences%20in%20contact%20placement%20and%20neuroanatomy%20between%20subjects.%20We%20propose%20MAPA%2C%20an%20otherwise%20vanilla%20masked%20autoencoder%20with%20two%20spatial%20encodings%2C%20an%20anatomical%20region%20embedding%20and%20a%20relative%20positional%20encoding%2C%20which%20together%20enable%20it%20to%20learn%20neural%20representations%20that%20transfer%20to%20unseen%20subjects%20and%20across%20various%20tasks.%20MAPA%20sets%20a%20new%20state%20of%20the%20art%20across%20all%20three%20regimes%20of%20the%20Neuroprobe%20benchmark%20without%20fine-tuning%3A%20within-session%2C%20cross-session%2C%20and%20cross-subject.%20In%20the%20cross-subject%20regime%2C%20a%20linear%20probe%20on%20MAPA%26%23039%3Bs%20features%20needs%20only%20%24%7B%5C%5Csim%7D164%24%20labeled%20trials%20to%20reach%20the%20accuracy%20that%20takes%203%2C500%20without%20pretraining.%20Our%20results%20show%20that%20self-supervised%20pretraining%20can%20scale%20across%20heterogeneous%20iEEG%20recordings%20and%20reduce%20the%20labeled%20data%20needed%20for%20accurate%20decoding%20in%20new%20subjects.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22arXiv%3A2609.13507%22%2C%22date%22%3A%222026-09-11%22%2C%22DOI%22%3A%2210.48550%5C%2FarXiv.2609.13507%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22http%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2609.13507%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-23T22%3A39%3A26Z%22%7D%7D%2C%7B%22key%22%3A%22WHLP4YBD%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Kwark%20et%20al.%22%2C%22parsedDate%22%3A%222026-09-09%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BKwark%2C%20D.%20H.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20Seamless%20Whole%20Slide%20Label-Free%20Virtual%20Staining.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2609.10914%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2609.10914%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Seamless%20Whole%20Slide%20Label-Free%20Virtual%20Staining%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Dou%20Hoon%22%2C%22lastName%22%3A%22Kwark%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Kianoush%22%2C%22lastName%22%3A%22Falahkheirkhah%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ji-hun%22%2C%22lastName%22%3A%22Oh%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shirui%22%2C%22lastName%22%3A%22Luo%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Volodymyr%22%2C%22lastName%22%3A%22Kindratenko%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Rohit%22%2C%22lastName%22%3A%22Bhargava%22%7D%5D%2C%22abstractNote%22%3A%22Label-free%20virtual%20staining%20offers%20a%20compelling%2C%20non-destructive%20alternative%20to%20standard%20histopathology%3B%20however%2C%20its%20clinical%20adoption%20is%20hindered%20by%20the%20computational%20bottlenecks%20inherent%20to%20processing%20gigapixel%20Whole%20Slide%20Images%20%28WSIs%29.%20Current%20deep%20learning%20approaches%20require%20patch-based%20inference%20to%20avoid%20memory%20constraints%2C%20which%20disrupts%20global%20tissue%20continuity%20and%20introduces%20tiling%20artifacts--displaying%20visible%20seams%20and%20color%20shifts.%20To%20address%20this%2C%20we%20introduce%20the%20Consistency%20Memory%20Bank%20%28COMB%29%2C%20a%20novel%20label-free%20virtual%20staining%20framework%20that%20enforces%20spatial%20and%20channel%20consistency%20across%20tiles%20without%20memory%20bottlenecks.%20COMB%20decouples%20context%20storage%20from%20computation%2C%20utilizing%20a%20dynamic%20retrieval%20mechanism%20to%20fetch%20feature%20representations%20from%20adjacent%20tiles.%20This%20enables%20a%20retrieval-based%20context%20integration%20strategy%20that%20adopts%20local%20padding%20to%20resolve%20spatial%20discontinuities%20and%20neighbor-aware%20channel%20attention%20to%20stabilize%20statistical%20drift.%20Further%20optimized%20with%20a%20sliding%20window%20schedule%20to%20ensure%20minimal%20memory%20overhead%2C%20our%20method%20demonstrates%20superior%20performance%20over%20state-of-the-art%20baselines%2C%20achieving%20significant%20improvements%20in%20both%20perceptual%20fidelity%20and%20tiling%20consistency%2C%20while%20suggesting%20its%20downstream%20utility%20in%20tumor%20segmentation.%20Code%20is%20available%20at%20https%3A%5C%2F%5C%2Fgithub.com%5C%2Fdou0000%5C%2FCOMB.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22arXiv%3A2609.10914%22%2C%22date%22%3A%222026-09-09%22%2C%22DOI%22%3A%2210.48550%5C%2FarXiv.2609.10914%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22http%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2609.10914%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-23T22%3A23%3A50Z%22%7D%7D%2C%7B%22key%22%3A%2269JMB5S6%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Koric%20et%20al.%22%2C%22parsedDate%22%3A%222026-09-07%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BKoric%2C%20A.%20A.%2C%20Liu%2C%20Q.%20%26amp%3B%20Koric%2C%20S.%20Disentangling%20Attention%20in%20Deep%20Operator%20Learning%3A%20A%20Controlled%20Study%20of%20Data-Driven%20and%20Physics-Informed%20Architectures.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2609.04407%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2609.04407%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Disentangling%20Attention%20in%20Deep%20Operator%20Learning%3A%20A%20Controlled%20Study%20of%20Data-Driven%20and%20Physics-Informed%20Architectures%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Amar%20Alem%22%2C%22lastName%22%3A%22Koric%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Qibang%22%2C%22lastName%22%3A%22Liu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Seid%22%2C%22lastName%22%3A%22Koric%22%7D%5D%2C%22abstractNote%22%3A%22Deep%20neural%20operators%20learn%20mappings%20between%20input%20functions%20and%20complete%20PDE%20solution%20fields%2C%20enabling%20forward%20evaluations%20of%20new%20problem%20instances%20orders%20of%20magnitude%20faster%20than%20conventional%20numerical%20solvers.%20Attention%20mechanisms%20have%20recently%20been%20introduced%20into%20neural%20operators%2C%20but%20most%20studies%20change%20several%20architectural%20components%20at%20once%2C%20making%20it%20difficult%20to%20identify%20what%20actually%20improves%20accuracy.%20This%20work%20presents%20a%20controlled%20and%20systematic%20study%20of%20five%20deep%20operator%20network%20%28DeepONet%29%20variants%20with%20distinct%20attention%20mechanisms%2C%20trained%20under%20both%20data-driven%20and%20physics-informed%20regimes%2C%20to%20isolate%20the%20effects%20of%20cross-attention%2C%20self-attention%2C%20tokenization%2C%20and%20attention%20depth.%20We%20evaluate%20them%20on%20a%20source-driven%20transient%20one-dimensional%20nonlinear%20diffusion-reaction%20equation%2C%20a%20transient%20one-dimensional%20viscous%20Burgers%20equation%20with%20variable%20initial%20conditions%2C%20and%20a%20two-dimensional%20Poisson%20heat-conduction%20problem%20with%20heterogeneous%20source%20fields.%20Per-sensor%20tokenization%20with%20cross-attention%20reduces%20the%20mean%20relative%20L_2%20error%20of%20the%20classical%20DeepONet%20in%20all%20benchmark-training%20combinations%20by%20factors%20of%202.4-28.0%2C%20while%20the%20best%20attention%20configurations%20reach%203.5-32.3.%20Branch%20self-attention%20paired%20only%20with%20dot-product%20fusion%20is%20inconsistent%2C%20degrading%20the%20one-dimensional%20problems%20while%20helping%20the%20more%20complex%20two-dimensional%20source%20field%3B%20added%20on%20top%20of%20cross-attention%20it%20improves%20all%20six%20cases%2C%20though%20by%20less%20than%20cross-attention%20fusion%20alone.%20Global%20pre-mixing%20provides%20no%20consistent%20benefit.%20Increasing%20cross-attention%20depth%20further%20improves%20accuracy%2C%20but%20with%20diminishing%20returns%20and%20a%20substantially%20higher%20cost%20under%20physics-informed%20training.%20Overall%2C%20query-dependent%20cross-attention%20is%20the%20most%20reliable%20mechanism%2C%20whereas%20branch%20self-attention%20is%20most%20useful%20for%20large%2C%20spatially%20complex%20functional%20inputs.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-09-07T00%3A06%3A03Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2609.04407%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2609.04407%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-23T22%3A09%3A08Z%22%7D%7D%2C%7B%22key%22%3A%22RNH6LX49%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Mahajan%20et%20al.%22%2C%22parsedDate%22%3A%222026-09-15%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BMahajan%2C%20J.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20ReaDiT%20Guidance%3A%20Control%20for%20Image%20and%20Video%20Generation%20using%20Diffusion%20Transformer%20Features.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2609.04649%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2609.04649%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22ReaDiT%20Guidance%3A%20Control%20for%20Image%20and%20Video%20Generation%20using%20Diffusion%20Transformer%20Features%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jay%22%2C%22lastName%22%3A%22Mahajan%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Chang%22%2C%22lastName%22%3A%22Liu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Rauf%22%2C%22lastName%22%3A%22Makharov%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Viraj%22%2C%22lastName%22%3A%22Shah%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Alexander%22%2C%22lastName%22%3A%22Schwing%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Svetlana%22%2C%22lastName%22%3A%22Lazebnik%22%7D%5D%2C%22abstractNote%22%3A%22We%20present%20DiT%20Readout%20%28ReaDiT%29%20Guidance%2C%20a%20lightweight%20framework%20for%20controlling%20generation%20with%20Diffusion%20Transformer%20%28DiT%29%20models%20via%20their%20internal%20feature%20representations.%20ReaDiT%20Guidance%20uses%20features%20from%20a%20single%20DiT%20block%20to%20steer%20the%20generative%20process%20according%20to%20spatial%20targets%20-%20like%20depth%2C%20pose%2C%20or%20edge%20maps%20-%20provided%20at%20test%20time.%20Furthermore%2C%20since%20modern%20text-to-video%20models%20are%20largely%20built%20on%20DiT%20backbones%2C%20ReaDiT%20Guidance%20naturally%20extends%20to%20video%20generation%2C%20enabling%20camera%20and%20motion%20control.%20Experimental%20results%20demonstrate%20that%20our%20approach%20achieves%20competitive%20or%20improved%20results%20compared%20to%20existing%20feature-based%20and%20off-the-shelf%20adapter-based%20approaches%20while%20requiring%20fewer%20parameters.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-09-15T00%3A53%3A48Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2609.04649%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2609.04649%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-23T22%3A06%3A46Z%22%7D%7D%2C%7B%22key%22%3A%22JAAQBUAG%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Ramasubramanian%20et%20al.%22%2C%22parsedDate%22%3A%222026-09-10%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BRamasubramanian%2C%20S.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20Tail-Likelihood%20Reinforcement%20Learning.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2609.02987%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2609.02987%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Tail-Likelihood%20Reinforcement%20Learning%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shrinivas%22%2C%22lastName%22%3A%22Ramasubramanian%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Daman%22%2C%22lastName%22%3A%22Arora%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Fahim%22%2C%22lastName%22%3A%22Tajwar%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Guanning%22%2C%22lastName%22%3A%22Zeng%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Qingyang%22%2C%22lastName%22%3A%22Wu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zhongzhu%22%2C%22lastName%22%3A%22Zhou%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Chenfeng%22%2C%22lastName%22%3A%22Xu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Haiwen%22%2C%22lastName%22%3A%22Feng%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yuda%22%2C%22lastName%22%3A%22Song%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Aarti%22%2C%22lastName%22%3A%22Singh%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ruslan%22%2C%22lastName%22%3A%22Salakhutdinov%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22J.%20Andrew%22%2C%22lastName%22%3A%22Bagnell%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jeff%22%2C%22lastName%22%3A%22Schneider%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Andrea%22%2C%22lastName%22%3A%22Zanette%22%7D%5D%2C%22abstractNote%22%3A%22Reinforcement%20learning%20typically%20optimizes%20average%20reward.%20For%20generative%20policies%2C%20the%20average%20can%20hide%20an%20important%20distinction%3A%20two%20policies%20can%20achieve%20the%20same%20mean%20reward%20while%20having%20very%20different%20chances%20of%20producing%20a%20rare%20but%20high-reward%20rollout.%20This%20matters%20as%20sampling%20increases%20during%20training%20and%20inference%2C%20since%20its%20benefit%20depends%20on%20retaining%20probability%20mass%20on%20high-reward%20outcomes.%20We%20propose%20to%20optimize%20this%20coverage%20directly.%20Rather%20than%20considering%20only%20expected%20reward%2C%20we%20consider%20all%20of%20its%20upper%20tails%3A%20for%20each%20reward%20threshold%2C%20how%20likely%20is%20the%20policy%20to%20exceed%20it%3F%20This%20turns%20a%20continuous%20reward%20into%20a%20family%20of%20binary%20success%20events.%20We%20introduce%20Tail-Likelihood%20Reinforcement%20Learning%20%28TailRL%29%2C%20which%20maximizes%20the%20log-probability%20of%20exceeding%20a%20randomly%20chosen%20reward%20threshold.%20Its%20gradient%20gives%20more%20weight%20to%20rare%2C%20high-reward%20rollouts%20and%20can%20be%20interpreted%20as%20a%20mixture%20of%20Best-of-k%20gradients.%20TailRL%20requires%20only%20a%20simple%20modification%20to%20the%20advantage%20function%2C%20making%20it%20compatible%20with%20existing%20reinforcement%20learning%20pipelines.%20Across%20object%20localization%2C%20maze%20navigation%2C%20GUI%20grounding%2C%20and%20code%20optimization%2C%20TailRL%20leverages%20rare%20high-reward%20training%20samples%20to%20avoid%20suboptimal%20solutions%20and%20yields%20models%20that%20benefit%20more%20from%20additional%20samples%20at%20inference%20time.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-09-10T01%3A00%3A36Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2609.02987%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2609.02987%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-23T22%3A02%3A46Z%22%7D%7D%2C%7B%22key%22%3A%22A2U2II5X%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Li%20et%20al.%22%2C%22parsedDate%22%3A%222026-08-27%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BLi%2C%20A.%2C%20Li%2C%20Y.%2C%20DeHaas%2C%20D.%20%26amp%3B%20Guidi%2C%20G.%20Cheaper%20by%20the%20Batch%3A%20Shared%20Traversal%20for%20Genotype%20Graph%20Editing.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2608.26488%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2608.26488%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Cheaper%20by%20the%20Batch%3A%20Shared%20Traversal%20for%20Genotype%20Graph%20Editing%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Aaron%22%2C%22lastName%22%3A%22Li%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yifan%22%2C%22lastName%22%3A%22Li%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Drew%22%2C%22lastName%22%3A%22DeHaas%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Giulia%22%2C%22lastName%22%3A%22Guidi%22%7D%5D%2C%22abstractNote%22%3A%22Updating%20a%20graph%20by%20inserting%20or%20replacing%20nodes%20while%20preserving%20semantics%20and%20reusing%20existing%20structure%20is%20a%20recurring%20computational%20problem.%20In%20population%20genetics%2C%20this%20problem%20arises%20in%20the%20genotype%20representation%20graph%20%28GRG%29%2C%20a%20directed%20acyclic%20graph%20that%20losslessly%20encodes%20phased%20genetic%20variation%20across%20hundreds%20of%20thousands%20of%20samples%20by%20sharing%20subgraph%20structure%20for%20individual%20mutations.%20In%20a%20GRG%2C%20each%20mutation%26%23039%3Bs%20carrier%20set%20is%20implicitly%20encoded%20as%20the%20set%20of%20leaf%20nodes%20reachable%20from%20the%20node%20it%20is%20assigned%20to.%20Updating%20a%20mutation%20is%20therefore%20a%20structural%20editing%20problem%2C%20and%20current%20approaches%20remap%20mutations%20individually.%20This%20paper%20introduces%20a%20batched%20mutation-remapping%20algorithm%20that%20replaces%20independent%20reuse-aware%20traversals%20with%20a%20single%20shared%20reverse-topological%20pass%2C%20identifying%20reuse%20candidates%20for%20an%20entire%20batch%20at%20once.%20The%20pass%20propagates%20compact%20bit-parallel%20per-mutation%20state%20and%20uses%20an%20adaptive%20sparse%5C%2Fdense%20carrier%20set%20representation%20spanning%20rare-to-common%20variant%20densities.%20Batching%20is%20the%20memory-scalable%20complement%20to%20split-based%20parallelism%2C%20which%20instead%20replicates%20graph%20and%20traversal%20state%20per%20worker.%20Our%20remapping%20is%20evaluated%20on%20a%20controlled%20update%20workload%20and%20on%20end-to-end%20allele%20polarization%2C%20a%20bulk%20carrier%20set%20update%20that%20is%20common%20in%20population%20genetic%20analysis.%20Our%20approach%20is%20up%20to%2010.5%24%5C%5Ctimes%24%20faster%20than%20independent%20remapping%20while%20preserving%20exact%20carrier-set%20semantics.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22arXiv%3A2608.26488%22%2C%22date%22%3A%222026-08-27%22%2C%22DOI%22%3A%2210.48550%5C%2FarXiv.2608.26488%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22http%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2608.26488%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-23T21%3A44%3A26Z%22%7D%7D%2C%7B%22key%22%3A%22A5CT83NE%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Yu%20and%20Ji%22%2C%22parsedDate%22%3A%222026-10%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BYu%2C%20C.%20%26amp%3B%20Ji%2C%20H.%20Sememe%3A%20Causal%20Modeling%20for%20Task-Aware%20Depth%20Pruning%20of%20Large%20Language%20Models.%20in%20%26lt%3Bi%26gt%3BCOLM%26%23xA0%3B%202026%26lt%3B%5C%2Fi%26gt%3B%20%28San%20Francisco%2C%20CA%2C%202026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22conferencePaper%22%2C%22title%22%3A%22Sememe%3A%20Causal%20Modeling%20for%20Task-Aware%20Depth%20Pruning%20of%20Large%20Language%20Models%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Charles%22%2C%22lastName%22%3A%22Yu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Henry%22%2C%22lastName%22%3A%22Ji%22%7D%5D%2C%22abstractNote%22%3A%22%22%2C%22proceedingsTitle%22%3A%22COLM%20%202026%22%2C%22conferenceName%22%3A%22Conference%20on%20Language%20Modeling%22%2C%22date%22%3A%222026-10%22%2C%22eventPlace%22%3A%22%22%2C%22DOI%22%3A%22%22%2C%22ISBN%22%3A%22%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Fopenreview.net%5C%2Fpdf%3Fid%3DUTMigobeqk%22%2C%22ISSN%22%3A%22%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-23T21%3A32%3A17Z%22%7D%7D%2C%7B%22key%22%3A%22W6ULBAPQ%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Howes%20et%20al.%22%2C%22parsedDate%22%3A%222026-08-26%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BHowes%2C%20W.%2C%20Ahmed%2C%20F.%20%26amp%3B%20Alam%2C%20S.%20B.%20Low-Latency%20Activation-Regularized%20Sparse%20Neural%20Operators%20with%20Distillation%20Assistance%20Towards%20Real-Time%20Edge-Deployable%20Virtual%20Sensing.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2608.23987%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2608.23987%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Low-Latency%20Activation-Regularized%20Sparse%20Neural%20Operators%20with%20Distillation%20Assistance%20Towards%20Real-Time%20Edge-Deployable%20Virtual%20Sensing%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22William%22%2C%22lastName%22%3A%22Howes%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Farid%22%2C%22lastName%22%3A%22Ahmed%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Syed%20Bahauddin%22%2C%22lastName%22%3A%22Alam%22%7D%5D%2C%22abstractNote%22%3A%22Virtual%20sensing%20enables%20digital%20twins%20and%20safety-critical%20systems%20to%20reconstruct%20and%20forecast%20spatial-temporal%20physics%20in%20real%20time.%20However%2C%20conventional%20computational%20and%20data-driven%20methods%20often%20face%20challenges%20in%20generalization%2C%20latency%2C%20and%20energy%20efficiency%20for%20edge%20deployment.%20Neural%20operators%20offer%20a%20promising%20alternative%20but%20remain%20reliant%20on%20power-intensive%20hardware.%20Spiking%20neurons%20and%20neuromorphic%20computing%20can%20improve%20efficiency%2C%20yet%20surrogate-gradient%20training%20and%20multi-step%20spiking%20introduce%20convergence%20and%20latency%20challenges.%20We%20propose%20the%20Sparse-Activation-ReLU%20%28SAR%29%20layer%2C%20a%20single-step%20alternative%20that%20promotes%20activation%20sparsity%20without%20surrogate-gradient%20training%20while%20remaining%20compatible%20with%20event-based%20computing.%20Within%20a%20trunk-based%20NOMAD%20architecture%2C%20SAR%20achieves%20over%20a%20fivefold%20improvement%20in%20the%20combined%20Latency-Error-Energy%20%28LEE%29%20metric%20compared%20with%20Variable%20Spiking%20Neuron%20%28VSN%29%20and%20Leaky%20Integrate-and-Fire%20%28LIF%29%20implementations.%20We%20further%20analyze%20spiking%20entropy%20and%20feature%20usage%20and%20introduce%20synthetic%20knowledge%20distillation%2C%20reducing%20the%20LEE%20score%20by%20more%20than%20twofold.%20Finally%2C%20we%20improve%20VSN%20through%20a%20ReLU-based%20spiking%20loss%20and%20graph-neighbor%20thresholding.%20On%20the%20Heat%20Exchanger%20dataset%2C%20these%20approaches%20reduce%20L2%20error%20by%20more%20than%20twofold%20and%20nearly%20sevenfold%2C%20respectively%2C%20while%20reducing%20spiking%20and%20spatial%20aggregation.%20Overall%2C%20the%20work%20presented%20is%20a%20step%20towards%20energy-efficient%20virtual%20sensing%20by%20providing%20an%20alternative%20framework%20that%20can%20be%20positioned%20towards%20neuromorphic%20or%20other%20edge%20device%20integration%20that%20can%20be%20a%20gold%20standard%20to%20compare%20latency%2C%20energy%2C%20and%20error%20performance%20for%20future%20efficient%20designs%20that%20are%20sparsity%20or%20brain-inspired%20spiking%20based.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-08-26T00%3A21%3A51Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2608.23987%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2608.23987%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-23T20%3A57%3A59Z%22%7D%7D%2C%7B%22key%22%3A%22LD8JBR8Q%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Someki%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BSomeki%2C%20M.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20PlanRAG-Audio%3A%20Planning%20and%20Retrieval%20Augmented%20Generation%20for%20Long-form%20Audio%20Understanding.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2605.20414%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2605.20414%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22PlanRAG-Audio%3A%20Planning%20and%20Retrieval%20Augmented%20Generation%20for%20Long-form%20Audio%20Understanding%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Masao%22%2C%22lastName%22%3A%22Someki%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Chien-yu%22%2C%22lastName%22%3A%22Huang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Siddhant%22%2C%22lastName%22%3A%22Arora%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Samuele%22%2C%22lastName%22%3A%22Cornell%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Markus%22%2C%22lastName%22%3A%22M%5Cu00fcller%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Nathan%22%2C%22lastName%22%3A%22Susanj%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Rupak%20V%22%2C%22lastName%22%3A%22Swaminathan%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Grant%20P%22%2C%22lastName%22%3A%22Strimel%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jing%22%2C%22lastName%22%3A%22Liu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shinji%22%2C%22lastName%22%3A%22Watanabe%22%7D%5D%2C%22abstractNote%22%3A%22Long-form%20audio%20understanding%20poses%20significant%20challenges%20for%20large%20audio%20language%20models%20%28LALMs%29%20due%20to%20the%20extreme%20length%20of%20audio%20sequences%20and%20the%20need%20to%20reason%20over%20heterogeneous%20acoustic%20cues%20distributed%20over%20time%2C%20such%20as%20speech%20content%2C%20speaker%20identity%2C%20emotion%2C%20and%20sound%20events.%20To%20address%20these%20challenges%2C%20we%20propose%20%5C%5Ctextbf%7BPlanRAG-Audio%7D%2C%20a%20planning-based%20retrieval-augmented%20generation%20framework%20for%20scalable%20long-form%20audio%20understanding.%20Rather%20than%20having%20audio%20LALMs%20process%20entire%20recordings%20directly%2C%20PlanRAG-Audio%20explicitly%20plans%20which%20modalities%20and%20temporal%20spans%20are%20required%20for%20a%20given%20query%2C%20and%20retrieves%20only%20query-relevant%20information%20from%20a%20structured%20text%20and%20audio%20database.%20This%20retrieval%20planning%20enables%20effective%20reasoning%20over%20complex%2C%20cross-domain%20audio%20queries%20while%20substantially%20reducing%20the%20input%20length%20passed%20to%20the%20large%20language%20models.%20Experiments%20across%20a%20wide%20range%20of%20speech%5C%2Faudio%20retrieval%20demonstrate%20that%20PlanRAG-Audio%20improves%20reasoning%20accuracy%20and%20stabilizes%20performance%20as%20audio%20duration%20increases%20by%20decoupling%20inference%20cost%20from%20raw%20audio%20length.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2605.20414%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2605.20414%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-09-21T17%3A50%3A28Z%22%7D%7D%2C%7B%22key%22%3A%22MMQ6I5I8%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Fan%20et%20al.%22%2C%22parsedDate%22%3A%222026-08-20%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BFan%2C%20X.%2C%20Liu%2C%20M.%2C%20Yang%2C%20R.%2C%20Ouyang%2C%20S.%20%26amp%3B%20Han%2C%20J.%20Can%20Agent%20Memory%20Systems%20Track%20Evolving%20State%3F%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2608.19652%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2608.19652%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Can%20Agent%20Memory%20Systems%20Track%20Evolving%20State%3F%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Xinyi%22%2C%22lastName%22%3A%22Fan%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Miri%22%2C%22lastName%22%3A%22Liu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ruozhen%22%2C%22lastName%22%3A%22Yang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Siru%22%2C%22lastName%22%3A%22Ouyang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jiawei%22%2C%22lastName%22%3A%22Han%22%7D%5D%2C%22abstractNote%22%3A%22As%20LLM-based%20agents%20are%20deployed%20for%20longer%20and%20higher-stakes%20tasks%2C%20their%20memory%20systems%20continue%20to%20have%20crucial%20gaps.%20While%20existing%20memory%20benchmarks%20focus%20largely%20on%20recall-shaped%20tasks%2C%20we%20argue%20an%20effective%20memory%20system%20must%20track%20the%20evolving%20state%20of%20the%20world%3B%20as%20facts%2C%20constraints%2C%20and%20decisions%20are%20revised%20over%20a%20long%20interaction%2C%20answers%20must%20reflect%20the%20current%20state%20and%20not%20a%20superseded%20one.%20We%20define%20this%20capability%20as%20state%20tracking%20and%20instantiate%20it%20in%20StateMemBench%2C%20a%20benchmark%20of%20234%20multi-session%20scenarios%20spanning%20two%20conversation-length%20regimes.%20Its%20closed-pool%20grading%20scores%20whether%20an%20answer%20reflects%20the%20current%20state%2C%20the%20superseded%20state%2C%20or%20fails%20otherwise%2C%20separating%20state-tracking%20failures%20from%20other%20errors%20by%20construction.%20Our%20analysis%20shows%20that%20this%20task%20is%20challenging%20for%20existing%20memory%20systems%2C%20retrieval-augmented%20baselines%2C%20and%20long-context%20baselines.%20We%20then%20present%20StateMem%2C%20a%20state-first%20memory%20method%20that%20explicitly%20tracks%20supersession%20and%20relational%20dependencies%2C%20and%20show%20it%20improves%20current-state%20accuracy%20over%20the%20strongest%20same-backbone%20baseline%20by%201.8x%20%280.205%20-%26gt%3B%200.363%29%20on%20DeepSeek-V4-Flash%20and%20over%20the%20strongest%20memory%20system%20by%201.6x%20%280.149%20-%26gt%3B%200.233%29%20on%20Qwen-3.5-9B%2C%20while%20remaining%20competitive%20with%20the%20long-context%20baselines.%20Finally%2C%20we%20show%20the%20same%20state%20approach%20can%20be%20applied%20as%20a%20lightweight%20single-call%20wrapper%20over%20existing%20memory%20systems%2C%20lifting%20current-state%20accuracy%20by%20%2B32%20to%20%2B67%20points%20on%20StateMemBench%20across%20six%20memory%20and%20retrieval%20backends.%20A%20length-%20and%20cost-matched%20control%20attributes%20%2B15%20to%20%2B32%20of%20those%20points%20to%20state%20structure%20rather%20than%20added%20context.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22arXiv%3A2608.19652%22%2C%22date%22%3A%222026-08-20%22%2C%22DOI%22%3A%2210.48550%5C%2FarXiv.2608.19652%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22http%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2608.19652%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A54%3A25Z%22%7D%7D%2C%7B%22key%22%3A%22IVHKYTH8%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Feng%20et%20al.%22%2C%22parsedDate%22%3A%222026-08-20%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BFeng%2C%20Y.%2C%20Bijoy%2C%20B.%20S.%2C%20Balasubramanian%2C%20N.%20%26amp%3B%20Zhou%2C%20J.%20Break%20It%20Down%2C%20Pass%20It%20On%3A%20Cross-Task%20Skill%20Transfer%20in%20LLM%20Agents.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2608.20274%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2608.20274%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Break%20It%20Down%2C%20Pass%20It%20On%3A%20Cross-Task%20Skill%20Transfer%20in%20LLM%20Agents%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yiyang%22%2C%22lastName%22%3A%22Feng%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Biddut%20Sarker%22%2C%22lastName%22%3A%22Bijoy%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Niranjan%22%2C%22lastName%22%3A%22Balasubramanian%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jiawei%22%2C%22lastName%22%3A%22Zhou%22%7D%5D%2C%22abstractNote%22%3A%22Large%20language%20model%20%28LLM%29%20agents%20can%20induce%20skills%20from%20completed%20tasks%20and%20reuse%20them%20later%20to%20grow%20more%20capable%20with%20experience.%20In%20practice%2C%20induced%20skills%20may%20transfer%20unreliably%20and%20can%20even%20harm%20the%20agent%20that%20retrieves%20them.%20When%20agent-induced%20skills%20transfer%20reliably%20across%20tasks%20remains%20an%20open%20question.%20We%20conduct%20a%20comprehensive%20and%20controlled%20study%20of%20how%20the%20way%20skills%20are%20induced%20shapes%20their%20transfer%20across%20tasks.%20Specifically%2C%20we%20compare%20task-level%20with%20subtask-level%20skill%20induction%20and%20text%20with%20code%20skill%20formats%2C%20the%20two%20axes%20along%20which%20existing%20methods%20differ.%20Task-level%20skills%20mostly%20reduce%20the%20agent%26%23039%3Bs%20performance%20below%20its%20no-memory%20baseline%20while%20subtask-level%20skills%20raise%20it%20above%20on%20average%2C%20and%20text%20skills%20transfer%20better%20than%20code%20skills.%20To%20further%20understand%20our%20findings%2C%20we%20examine%20two%20complementary%20properties%20of%20the%20induced%20skills%3A%20specificity%2C%20which%20measures%20how%20closely%20a%20skill%20matches%20real%20tasks%2C%20and%20abstractness%2C%20which%20measures%20how%20evenly%20its%20relevance%20spreads%20across%20tasks.%20Neither%20property%20alone%20predicts%20task%20success%2C%20but%20their%20combined%20effect%20does%2C%20which%20we%20propose%20as%20a%20skill%20utility%20score.%20The%20score%20correlates%20consistently%20with%20task%20success%20when%20skills%20are%20transferred%2C%20and%20subtask-level%20and%20text%20skills%20score%20higher.%20Computing%20skill%20utility%20only%20needs%20the%20skills%20and%20task%20descriptions%20but%20not%20any%20task%20execution%2C%20so%20our%20score%20serves%20as%20a%20practical%20diagnostic%20of%20a%20skill%20memory%20before%20any%20new%20task%20runs.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22arXiv%3A2608.20274%22%2C%22date%22%3A%222026-08-20%22%2C%22DOI%22%3A%2210.48550%5C%2FarXiv.2608.20274%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22http%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2608.20274%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A52%3A23Z%22%7D%7D%2C%7B%22key%22%3A%22PR4UN2VC%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Chen%20et%20al.%22%2C%22parsedDate%22%3A%222026-08-04%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BChen%2C%20N.%2C%20Yang%2C%20Z.%20%26amp%3B%20Hayou%2C%20S.%20Training-Free%20versus%20Training-Based%20Intent%20Classification%20in%20LLMs%3A%20Accuracy%2C%20Robustness%2C%20and%20Failure%20Modes.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2608.02415%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2608.02415%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Training-Free%20versus%20Training-Based%20Intent%20Classification%20in%20LLMs%3A%20Accuracy%2C%20Robustness%2C%20and%20Failure%20Modes%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Nan%22%2C%22lastName%22%3A%22Chen%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zhouhao%22%2C%22lastName%22%3A%22Yang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Soufiane%22%2C%22lastName%22%3A%22Hayou%22%7D%5D%2C%22abstractNote%22%3A%22Intent%20classification%20in%20Large%20Language%20Models%20%28LLMs%29%20involves%20categorizing%20user%20prompts%20into%20predefined%20classes.%20For%20instance%2C%20given%20a%20user%20prompt%2C%20the%20system%20must%20determine%20whether%20it%20primarily%20concerns%20mathematics%2C%20coding%2C%20or%20general%20text%20processing.%20Such%20classification%20enables%20routing%20prompts%20to%20specialized%20models%20optimized%20for%20specific%20domains%2C%20improving%20both%20accuracy%20and%20computational%20efficiency.%20In%20this%20work%2C%20we%20conduct%20a%20systematic%20study%20comparing%20training-free%20vs%20training-based%20approaches%20for%20intent%20classification.%20For%20this%20purpose%2C%20we%20consider%20two%20lightweight%2C%20training-free%20methods%20based%20on%20statistics%20of%20internal%20representations%20and%20compare%20them%20against%20MLP%20classifiers%20and%20linear%20probes.%20Our%20comprehensive%20empirical%20evaluation%20reveals%20that%201%29%20Both%20training-free%20and%20training-based%20methods%20saturate%20easy%20benchmarks%20%28mathematics%20vs.%20coding%20vs.%20natural%20language%29%2C%202%29%20Training-based%20classifiers%20have%20an%20advantage%20on%20harder%20classification%20tasks%20%28e.g.%20Java%20vs%20Python%29%2C%20and%203%29%20Training-free%20methods%20are%20generally%20more%20robust%20to%20mixed-intent%20and%20adversarial%20prompts.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-08-04T02%3A07%3A56Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2608.02415%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2608.02415%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A50%3A59Z%22%7D%7D%2C%7B%22key%22%3A%22BL94W77S%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Rodr%5Cu00edguez%22%2C%22parsedDate%22%3A%222026-07-31%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BRodr%26%23xED%3Bguez%2C%20C.%20C.%20Tails%20and%20Walls%3A%20Reality%26%23x2019%3Bs%20Two%20Honest%20Priors.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.21203%5C%2Frs.3.rs-10473971%5C%2Fv1%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.21203%5C%2Frs.3.rs-10473971%5C%2Fv1%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Tails%20and%20Walls%3A%20Reality%27s%20Two%20Honest%20Priors%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Carlos%20C.%22%2C%22lastName%22%3A%22Rodr%5Cu00edguez%22%7D%5D%2C%22abstractNote%22%3A%22Abstract%5Cn%20%20%20%20%20%20%20%20%20%20The%20maximally%20ignorant%20priors%20of%20a%20regular%20statistical%20model%20come%20in%20two%20dual%20branches%3A%20a%20heavy%20branch%20%5B1%20%2B%20%5Cu03b1%5Cu03bd%20I%5Cu03b4%5D%5Cu22121%5C%2F%5Cu03bd%20%282002%3B%20Student-t%20for%20scalar%20location%20models%2C%20up%20to%20Cauchy%3B%20multivariate%20for%20%5Cu03bd%20%26lt%3B%202%5C%2Fd%29%20and%20a%20compact-support%20droplet%20%5B1%20%5Cu2212%20%5Cu03b1%5Cu03bd%20I%5Cu03b4%5D1%5C%2F%5Cu03bd%20%2B%20%282026%29%2C%20with%20I%5Cu03b4%20an%20information%20deviation%2C%200%20%5Cu2264%20%5Cu03bd%20%5Cu2264%201%20the%20tail%20index%2C%20and%20%5Cu03b4%20%5Cu2208%20%5B0%2C%201%5D%20a%20robustness%20%28loss%29%20parameter.%20In%20infinite%20dimensions%20there%20is%20no%20canonical%20volume%20%28no%20locally%20finite%20translation-invariant%20measure%29%2C%20so%20densities%20exist%20only%20against%20a%20reference%3B%20but%20on%20a%20finite-dimensional%20model%20the%20Information%20Volume%20dV%20%3D%20%5Cu221a%20det%20I%20d%5Cu03b8%20is%20canonical.%20Writing%20a%20measure%20in%20its%20%5Cu03b4-coordinates%20%5Cu2113%5Cu03b4%28p%29%20%3D%20p%5Cu03b4%5C%2F%5Cu03b4%5Cu2014a%20canonical%20vector%20embedding%20into%20L1%5C%2F%5Cu03b4%2C%20self-dual%20Hilbert%20at%20%5Cu03b4%20%3D%201%202%5Cu2014the%20pairing%20of%20a%20coordinate%20against%20its%20dual%20returns%20the%20same%20metric%20I%20for%20every%20%5Cu03b4%2C%20so%20dV%20is%20deformation-independent%20and%20computable%20%28I%20%3D%20J%5Cu22a4J%3B%20Cauchy%5Cu2013Binet%20and%20Hutchinson%5C%2FLanczos%20probes%2C%20the%20operator%20the%20companion%20diffusion%20paper%20builds%29.%20Extremizing%20the%20same%20action%20over%20the%20truth-guess%20gives%20t%5Cu22c6%2C%20the%20models%5Cu2019%20barycenter%20in%20these%20coordinates%5Cu2014the%20Bayesian%20predictive%20at%20%5Cu03b4%20%3D%201%2C%20Robbins%5Cu2019s%20empirical%20Bayes%5Cu2014so%20the%20empirical%20marginal%20one%20measures%20is%20t%5Cu22c6%2C%20and%20its%20branch%20is%20the%20prior%5Cu2019s%3A%20a%20heavy%20prior%20stays%20heavy%2C%20while%20a%20droplet%5Cu2019s%20wall%20survives%20a%20bounded%20observation%20but%20washes%20under%20an%20unbounded%20one.%20The%20branches%20map%20onto%20the%20extreme-value%20trichotomy%20%28heavy%20%5Cu03be%20%3D%20%5Cu03bd%5C%2F%282%5Cu2212%5Cu03bd%29%20%5Cu2208%20%280%2C%201%5D%2C%20droplet%20%5Cu03be%20%3D%20%5Cu2212%5Cu03bd%5C%2F%281%20%2B%20%5Cu03bd%29%20%5Cu2208%20%5B%5Cu22121%202%20%2C%200%29%2C%20meeting%20at%20Gauss%29%2C%20so%20the%20tail%20shape%20%5Cu03be%20classifies%20an%20observable%20as%20droplet-%20or%20heavy-type%20within%20the%20footprint%20%5Cu03be%20%5Cu2208%20%5B%5Cu22121%202%20%2C%201%5D.%20In%20QM9%20the%20molecular%20ensemble%20hosts%20both%20branches%20%28C%5Cu2013O%20walled%20on%20both%20sides%2C%20the%20hydrogen-bonded%20O%5Cu2013H%20heavy%29%3B%20sound%20is%20heavy%2C%20speech%20relaxes%20to%20Gauss%2C%20and%208-bit%20images%20are%20droplets%20whose%20wall%20a%20pre-registered%20test%20attributes%20to%20the%20representation%2C%20not%20the%20radiance%20field.%20Every%20sign-stable%20marginal%20but%20one%20lands%20inside%20the%20footprint%5Cu2014the%20exception%2C%20HDR%20linear%20luminance%2C%20is%20a%20super-Cauchy%20stress%20test%20we%20leave%20open.%5CnMSC%20Classification%3A%2062B11%20%2C%2060G70%20%2C%2062F15%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-07-31%22%2C%22DOI%22%3A%2210.21203%5C%2Frs.3.rs-10473971%5C%2Fv1%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Fwww.researchsquare.com%5C%2Farticle%5C%2Frs-10473971%5C%2Fv1%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A49%3A38Z%22%7D%7D%2C%7B%22key%22%3A%22NEABKW62%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Berres%20et%20al.%22%2C%22parsedDate%22%3A%222026-08-04%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BBerres%2C%20A.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20DB-Bench%3A%20Benchmarking%20Deblenders%20for%20LSST%20DESC%20Using%20the%20Blending%20ToolKit.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.28475%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.28475%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22DB-Bench%3A%20Benchmarking%20Deblenders%20for%20LSST%20DESC%20Using%20the%20Blending%20ToolKit%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Aidan%22%2C%22lastName%22%3A%22Berres%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Grant%22%2C%22lastName%22%3A%22Merz%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Xin%22%2C%22lastName%22%3A%22Liu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22%5Cu00c9ric%22%2C%22lastName%22%3A%22Aubourg%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22C%5Cu00e9cile%22%2C%22lastName%22%3A%22Roucelle%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22the%20LSST%20Dark%20Energy%20Science%22%2C%22lastName%22%3A%22Collaboration%22%7D%5D%2C%22abstractNote%22%3A%22Blending%20will%20be%20a%20major%20source%20of%20systematic%20uncertainty%20in%20downstream%20science%20analyses%20of%20LSST%20data.%20We%20benchmark%20the%20performance%20of%20several%20deblenders%2C%20leveraging%20the%20Blending%20ToolKit%20%28BTK%29%20to%20perform%20rigorous%2C%20end-to-end%20testing.%20This%20benchmark%20incorporates%20key%20deblending%20algorithms%2C%20including%20SourceExtractor%2C%20SCARLET%2C%20and%20DeepDISC%2C%20with%20the%20goal%20of%20comparing%20their%20effectiveness%20in%20handling%20blended%20galaxy%20images%20from%20LSST%5C%2FRubin%20simulations.%20A%20key%20focus%20is%20characterizing%20algorithm%20performance%20in%20the%20regime%20of%20unrecognized%20blends%2C%20where%20multiple%20galaxies%20are%20misidentified%20as%20a%20single%20object%2C%20as%20these%20cases%20introduce%20systematic%20biases%20that%20propagate%20into%20downstream%20cosmological%20analyses%20for%20galaxy%20surveys.%20By%20utilizing%20BTK%26%23039%3Bs%20ability%20to%20create%20customized%2C%20reproducible%20blends%2C%20we%20systematically%20test%20these%20deblenders%20against%20different%20blending%20conditions%2C%20such%20as%20source%20separation%20and%20brightness.%20The%20toolkit%26%23039%3Bs%20standardized%20evaluation%20metrics%2C%20including%20detection%20precision%2C%20segmentation%20accuracy%2C%20and%20source%20reconstruction%2C%20are%20comprehensive%20assessments%20of%20each%20algorithm%26%23039%3Bs%20strengths%20and%20limitations.%20Each%20deblender%20has%20performance%20caveats%20that%20may%20impact%20their%20true%20performance%20in%20real%20survey%20conditions.%20We%20find%20that%20SCARLET%20has%20high%20segmentation%20and%20reconstruction%20performance%2C%20whereas%20DeepDISC%20has%20strong%20detection%20recall%20for%20faint%20and%20low-SNR%20sources%2C%20and%20SourceExtractor%20has%20accurate%20peak%20finding%20abilities%20but%20low%20segmentation%20and%20reconstruction%20performance.%20This%20benchmark%20provides%20valuable%20insights%20into%20the%20performance%20of%20existing%20deblenders%20and%20highlights%20areas%20for%20future%20development.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-08-04T00%3A07%3A46Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2607.28475%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2607.28475%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A47%3A16Z%22%7D%7D%2C%7B%22key%22%3A%226E22IEJS%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Radhakrishnan%20and%20Bryngelson%22%2C%22parsedDate%22%3A%222026-07-31%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BRadhakrishnan%2C%20A.%20%26amp%3B%20Bryngelson%2C%20S.%20H.%20A%20sub-grid-scale%20model%20for%20polydisperse%20bubbly%20flows%20with%20heat%20and%20mass%20transfer.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.27426%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.27426%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22A%20sub-grid-scale%20model%20for%20polydisperse%20bubbly%20flows%20with%20heat%20and%20mass%20transfer%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Anand%22%2C%22lastName%22%3A%22Radhakrishnan%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Spencer%20H.%22%2C%22lastName%22%3A%22Bryngelson%22%7D%5D%2C%22abstractNote%22%3A%22Ensemble-averaged%20models%20of%20polydisperse%20bubbly%20flows%20require%20statistics%20of%20the%20evolving%20bubble%20population.%20Prior%20quadrature-based%20moment%20formulations%20close%20bubble%20pressure%20with%20a%20polytropic%20relation%20that%20omits%20heat%20and%20mass%20transfer%20at%20the%20bubble%20wall.%20We%20formulate%20constant-transfer%20equations%20for%20bubble%20pressure%20and%20vapor%20mass%20within%20a%20conditional%20hyperbolic%20quadrature%20method.%20Second-order%20conditional%20inversion%20produces%20four%20joint%20radius--radial-velocity%20nodes%20per%20equilibrium-radius%20bin.%20Bubble%20pressure%20and%20vapor%20mass%20are%20advanced%20at%20each%20node.%20The%20node%20values%20close%20the%20ensemble-averaged%20flow%20equations%20without%20adding%20mixed%20pressure%20or%20vapor-mass%20moments%20to%20the%20transported%20moment%20set.%20The%20model%20is%20implemented%20in%20MFC.%20Monte%20Carlo%20calculations%20verify%20the%20evolution%20of%20quadrature%20nodes%20and%20the%20mean%20bubble%20variables%20for%20a%20harmonically%20forced%20population.%20Bubble-screen%20calculations%20quantify%20closure%20error%20as%20the%20equilibrium%20radius%20is%20discretized%20and%20the%20initial%20distributions%20vary.%20The%20constant-transfer%20calculation%20does%20not%20exhibit%20the%20high-frequency%20pressure%20oscillations%20observed%20with%20the%20polytropic%20closure%20under%20the%20conditions%20considered.%203D%20bubble-screen%20calculations%20give%20a%201.5%25%20relative%20root-mean-square%20error%20between%20the%20Euler--Euler%20center%20pressure%20and%20the%20mean%20of%2040%20volume-averaged%20Euler--Lagrange%20realizations.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-07-31T00%3A08%3A01Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2607.27426%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2607.27426%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A44%3A58Z%22%7D%7D%2C%7B%22key%22%3A%224JVBZ6QE%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Zhou%20et%20al.%22%2C%22parsedDate%22%3A%222026-07-30%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BZhou%2C%20S.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20Filesystem-Based%20Memory%20for%20LLM%20Agents%3A%20Organization%2C%20Evolution%2C%20and%20Sustainability.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.26637%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.26637%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Filesystem-Based%20Memory%20for%20LLM%20Agents%3A%20Organization%2C%20Evolution%2C%20and%20Sustainability%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Sizhe%22%2C%22lastName%22%3A%22Zhou%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Sheldon%22%2C%22lastName%22%3A%22Yu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Hui%22%2C%22lastName%22%3A%22Wei%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Junda%22%2C%22lastName%22%3A%22Wu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Siru%22%2C%22lastName%22%3A%22Ouyang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yizhu%22%2C%22lastName%22%3A%22Jiao%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shijia%22%2C%22lastName%22%3A%22Pan%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Julian%22%2C%22lastName%22%3A%22McAuley%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yu%22%2C%22lastName%22%3A%22Zhang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Tong%22%2C%22lastName%22%3A%22Yu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jiawei%22%2C%22lastName%22%3A%22Han%22%7D%5D%2C%22abstractNote%22%3A%22Deployed%20LLM%20agents%20increasingly%20keep%20their%20long-term%20memory%20as%20a%20filesystem%3A%20a%20directory%20tree%20of%20markdown%20files%20that%20the%20agent%20itself%20reads%2C%20writes%2C%20and%20reorganizes%20through%20generic%20file%20tools.%20Yet%20research%20has%20largely%20passed%20over%20this%20medium%3A%20prior%20systems%20design%20bespoke%20memory%20representations%20and%20study%20retrieval%20over%20them%2C%20leaving%20the%20default%26%23039%3Bs%20two%20working%20assumptions%20untested%3A%20that%20an%20agent%20can%20keep%20a%20growing%20store%20organized%20as%20memories%20accumulate%2C%20conflict%2C%20and%20go%20stale%2C%20and%20that%20this%20organization%20pays.%20We%20present%20the%20first%20systematic%20exploration%20of%20filesystem-based%20memory%20for%20LLM%20agents.%20We%20formalize%20the%20setting%20as%20three%20roles%20around%20one%20memory%20filesystem%3A%20a%20management%20agent%20integrates%20and%20organizes%20incoming%20content%2C%20a%20search%20agent%20answers%20queries%20with%20cited%20sources%2C%20and%20an%20execution%20agent%20supplies%20task%20trajectories%20that%20are%20distilled%20into%20skills%2C%20unifying%20declarative%20memory%20and%20skills%20in%20a%20single%20store.%20Across%20long-conversation%20benchmarks%20and%20embodied%20tasks%2C%20we%20vary%20memory%20shape%20%28agent-organized%20hierarchy%2C%20verbatim%20dump%2C%20chunk%20retrieval%29%2C%20stream%20scale%2C%20tool%20harness%20%28sandboxed%20shell%2C%20memory-tool-style%20functions%2C%20varied%20search%20tooling%29%2C%20and%20the%20strengths%20of%20the%20management%20and%20search%20agents%2C%20tracking%20answer%20quality%2C%20cost%2C%20and%20store%20health%20as%20memory%20grows.%20What%20organization%20reliably%20buys%20is%20search%20economy%3A%20organized%20stores%20roughly%20halve%20retrieval%20cost%20where%20material%20is%20large.%20Today%26%23039%3Bs%20agents%2C%20however%2C%20fall%20short%20of%20the%20default%26%23039%3Bs%20promise%3A%20in%20our%20growth%20study%2C%20organization%20erodes%20for%20all%20but%20the%20strongest%20management%20agent%2C%20and%20no%20agent%20we%20measure%20converts%20organization%20itself%20into%20better%20answers.%20And%20the%20model%20is%20not%20the%20only%20lever%20over%20a%20store%26%23039%3Bs%20shape%3A%20changing%20the%20tool%20set%20alone%20reshapes%20the%20store%20as%20strongly%20as%20swapping%20the%20model.%20The%20study%20turns%20the%20filesystem%20default%20from%20an%20assumption%20into%20a%20design%20space%20for%20agent%20memory.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-07-30T00%3A34%3A02Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2607.26637%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2607.26637%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A42%3A44Z%22%7D%7D%2C%7B%22key%22%3A%229NV6N3NW%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Hidalgo%20and%20Draper%22%2C%22parsedDate%22%3A%222026-07-30%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BHidalgo%2C%20L.%20%26amp%3B%20Draper%2C%20P.%20Lattice%20Quantum%20Chromodynamics%20for%20Quantum%20Simulations.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.26445%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.26445%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Lattice%20Quantum%20Chromodynamics%20for%20Quantum%20Simulations%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Luis%22%2C%22lastName%22%3A%22Hidalgo%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Patrick%22%2C%22lastName%22%3A%22Draper%22%7D%5D%2C%22abstractNote%22%3A%22We%20develop%20a%20framework%20for%20quantum%20simulations%20of%20lattice%20SU%28%24N_c%24%29%20gauge%20theory%20with%20quarks.%20Staggered%20and%20Wilson%20lattice%20fermions%20are%20considered%20in%20two%20and%20three%20spatial%20dimensions%20and%20a%20theta%20angle%20is%20included%20in%20three%20dimensions.%20The%20physical%2C%20gauge-invariant%20Hilbert%20space%20is%20formulated%20in%20a%20representation%20basis%2C%20where%20gauge%20and%20fermionic%20degrees%20of%20freedom%20are%20encoded%20by%20irreducible%20representations%20of%20SU%28%24N_c%24%29%20that%20tensor%20at%20each%20lattice%20site%20to%20contain%20a%20singlet.%20We%20discuss%20algorithms%20for%20simulating%20time%20evolution%20on%20quantum%20computers%20and%20carry%20out%20noiseless%20simulations%20of%20lattice%20quantum%20chromodynamics%20%28%24N_c%3D3%24%29%20on%20small%20lattices%20with%20up%20to%2032%20qubits%2C%20showcasing%20theta%20angle%20effects%2C%20hadronic%20states%2C%20string%20dynamics%2C%20and%20a%20baryon%20chemical%20potential%20for%20the%20first%20time%20in%20three%20spatial%20dimensions.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-07-30T00%3A21%3A26Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2607.26445%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2607.26445%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A38%3A37Z%22%7D%7D%2C%7B%22key%22%3A%22TAFPAA9B%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Gladstone%20et%20al.%22%2C%22parsedDate%22%3A%222026-07-31%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BGladstone%2C%20A.%2C%20Ji%2C%20H.%20%26amp%3B%20Du%2C%20Y.%20Explorative%20Modeling%3A%20Unlocking%20a%20Third%20Pretraining%20Axis%20and%20End-to-End%20Generation.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.27372%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.27372%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Explorative%20Modeling%3A%20Unlocking%20a%20Third%20Pretraining%20Axis%20and%20End-to-End%20Generation%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Alexi%22%2C%22lastName%22%3A%22Gladstone%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Heng%22%2C%22lastName%22%3A%22Ji%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yilun%22%2C%22lastName%22%3A%22Du%22%7D%5D%2C%22abstractNote%22%3A%22The%20deep%20learning%20revolution%2C%20kicked%20off%20by%20AlexNet%2C%20taught%20us%20that%20end-to-end%20training%20beats%20decomposing%20a%20problem%20into%20hand-designed%20stages.%20Generative%20modeling%2C%20however%2C%20has%20remained%20the%20exception-despite%20generative%20models%20being%20remarkably%20capable%2C%20they%20are%20still%20not%20trained%20end-to-end.%20This%20is%20because%2C%20at%20its%20core%2C%20generative%20modeling%20is%20about%20handling%20distributions%20with%20many%20modes%2C%20and%20existing%20scalable%20approaches%20handle%20this%20the%20same%20way%2C%20by%20factoring%20the%20generation%20procedure%2C%20which%20prevents%20end-to-end%20generation.%20In%20this%20work%2C%20we%20introduce%20Explorative%20Modeling%2C%20a%20new%20paradigm%20that%20instead%20factors%20the%20training%20loop%2C%20exploring%20K%20candidate%20matches%20between%20model%20generations%20and%20data%2C%20and%20training%20on%20the%20best%2C%20so%20predictions%20commit%20to%20modes%20rather%20than%20blurring%20them.%20We%20find%20Explorative%20Models%20%28XMs%29%20useful%20in%20two%20settings.%20First%2C%20increasing%20exploration%20adds%20a%20third%20pretraining%20axis%20beyond%20parameters%20and%20data%20for%20existing%20generative%20models-where%20scaling%20exploration%20monotonically%20improves%20performance%20across%20both%20continuous%20and%20discrete%20domains%20%28images%2C%20video%2C%20and%20language%29.%20Notably%2C%20gains%20from%20exploration%20increase%20with%20scale%2C%20climbing%20from%207%25%20to%2036%25%20as%20data%20scales%20and%20from%2013%25%20to%2023%25%20as%20models%20grow%2C%20with%20efficiency%20gains%20more%20than%20doubling%20at%203x%20the%20compute.%20Concretely%2C%20exploration%20improves%20FLOP%20efficiency%20by%204.1x%2C%20sample%20efficiency%20by%206.2x%2C%20parameter%20efficiency%20by%2047%25%2C%20lifts%20the%20strongest%20of%20image-generation%20recipes%20to%20a%20near-state-of-the-art%201.43%20FID%20on%20ImageNet%20without%20guidance%2C%20enables%20scaling%20how%20end-to-end%20existing%20models%20are%2C%20and%20unlocks%20scaling%20generalization.%20Second%2C%20XMs%20enable%20end-to-end%20reconstructive%20generative%20modeling%2C%20matching%20diffusion%20on%20control%20tasks%20with%2016-256x%20fewer%20inference%20steps.%20Together%2C%20these%20results%20establish%20XMs%20as%20both%20a%20new%20pretraining%20axis%20for%20existing%20generative%20models%20and%20a%20standalone%20end-to-end%20generative%20modeling%20paradigm.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-07-31T00%3A04%3A45Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2607.27372%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2607.27372%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A36%3A43Z%22%7D%7D%2C%7B%22key%22%3A%22L2J5WXF3%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Wu%20et%20al.%22%2C%22parsedDate%22%3A%222026-07-27%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BWu%2C%20S.%2C%20Qian%2C%20C.%2C%20Chen%2C%20X.%20%26amp%3B%20Ji%2C%20H.%20Teaching%20LLMs%20to%20Self-Evolve%3A%20Cultivating%20Core%20Meta-Skills%20with%20Reinforcement%20Learning.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.21971%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.21971%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Teaching%20LLMs%20to%20Self-Evolve%3A%20Cultivating%20Core%20Meta-Skills%20with%20Reinforcement%20Learning%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shujin%22%2C%22lastName%22%3A%22Wu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Cheng%22%2C%22lastName%22%3A%22Qian%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Xiusi%22%2C%22lastName%22%3A%22Chen%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Heng%22%2C%22lastName%22%3A%22Ji%22%7D%5D%2C%22abstractNote%22%3A%22Test-time%20scaling%20through%20iterative%20self-evolution%20with%20environment%20feedback%2C%20as%20demonstrated%20by%20AlphaEvolve%2C%20shows%20remarkable%20performance%20gains.%20We%20hypothesize%20that%20the%20success%20of%20such%20evolution%20frameworks%20hinges%20on%20meta-skills%2C%20such%20as%20self-reflection%20with%20environment%20feedback%2C%20that%20enable%20effective%20multi-round%20refinement%2C%20yet%20are%20largely%20neglected%20by%20traditional%20post-training.%20To%20bridge%20this%20gap%2C%20we%20present%20MetaEvolve%2C%20a%20framework%20designed%20to%20develop%20these%20meta-skills%20via%20a%20data%20synthesis%20pipeline%2C%20evolution-aware%20reinforcement%20learning%20%28RL%29%2C%20and%20inference-time%20evolutionary%20search.%20Concretely%2C%20we%20ground%20MetaEvolve%20in%20coding%2C%20where%20program%20execution%20provides%20natural%2C%20continuous%20reward%20signals%20beyond%20binary%20correctness.%20Building%20on%20these%20signals%2C%20we%20synthesize%20evolution%20trajectories%20as%20training%20data%2C%20each%20containing%20a%20current%20program%2C%20its%20fitness%20score%20%28combining%20correctness%20and%20efficiency%29%2C%20and%20a%20history%20of%20prior%20attempts%2C%20and%20train%20the%20model%20via%20RL%20with%20verifiable%20rewards%20derived%20from%20test%20case%20execution.%20By%20training%20on%20large-scale%20code%20data%2C%20we%20aim%20to%20inspire%20generalizable%20domain-agnostic%20meta-skills%20that%20can%20transfer%20broadly%20to%20open-ended%20problems%20where%20such%20rich%20training%20signals%20are%20scarce.%20Across%20seven%20coding%20benchmarks%2C%20MetaEvolve%20outperforms%20the%20strongest%20baseline%20by%2010.01%25%20absolute%20on%20in-distribution%20tasks%20and%2024.12%25%20on%20out-of-distribution%20tasks.%20On%20open-ended%20algorithm%20optimization%20problems%20entirely%20outside%20the%20training%20domain%2C%20it%20further%20achieves%20a%2046.9%25%20relative%20improvement.%20These%20results%20demonstrate%20that%20explicitly%20cultivating%20self-evolution%20meta-skills%20offers%20a%20principled%20path%20toward%20more%20capable%20and%20autonomously%20self-evolving%20AI.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-07-27T00%3A18%3A31Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2607.21971%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2607.21971%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A34%3A30Z%22%7D%7D%2C%7B%22key%22%3A%22N4YEVY84%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Lin%20et%20al.%22%2C%22parsedDate%22%3A%222026-08-04%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BLin%2C%20S.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20D%24_4%24CNN%24%5C%5Ctimes%24AnaCal%3A%20Physics-Informed%20Machine%20Learning%20for%20Accurate%20and%20Precise%20Weak%20Lensing%20Shear%20Estimation.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2603.19046%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2603.19046%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22D%24_4%24CNN%24%5C%5Ctimes%24AnaCal%3A%20Physics-Informed%20Machine%20Learning%20for%20Accurate%20and%20Precise%20Weak%20Lensing%20Shear%20Estimation%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shurui%22%2C%22lastName%22%3A%22Lin%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Xiangchong%22%2C%22lastName%22%3A%22Li%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ji%22%2C%22lastName%22%3A%22Li%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shengcao%22%2C%22lastName%22%3A%22Cao%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Xin%22%2C%22lastName%22%3A%22Liu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yu-Xiong%22%2C%22lastName%22%3A%22Wang%22%7D%5D%2C%22abstractNote%22%3A%22Traditional%20weak%20gravitational%20lensing%20shear%20estimators%20are%20carefully%20calibrated%20but%20struggle%20to%20fully%20capture%20realistic%20galaxy%20morphologies%2C%20point-spread-function%20%28PSF%29%20effects%2C%20blending%2C%20and%20noise%20in%20deep%20surveys%2C%20while%20blindly%20trained%20machine%20learning%20%28ML%29%20models%20can%20introduce%20significant%20calibration%20biases.%20Here%20we%20construct%20a%20fully%20D%24_4%24-equivariant%20deep%20neural%20network%20for%20galaxy%20shape%20measurement%20whose%20architecture%20enforces%20symmetry%20under%2090%24%5E%7B%5C%5Ccirc%7D%24%20rotations%20and%20mirror%20transformations%2C%20and%20adopt%20the%20Analytical%20Calibration%20framework%20%28AnaCal%29%20to%20calibrate%20the%20model%20using%20its%20backpropagated%20gradients.%20For%20isolated%20galaxies%20in%20LSST-like%20single-band%20simulations%2C%20we%20demonstrate%20that%20our%20approach%20achieves%20%24%5C%5Csim%2410%25%20lower%20shape%20noise%20than%20the%20traditional%20moment-based%20Fourier%20Power%20Function%20Shapelets%20estimator%20in%20the%20high-noise%20regime%2C%20equivalent%20to%20a%20%24%5C%5Csim%2420%25%20gain%20in%20effective%20galaxy%20number%20density%2C%20while%20simultaneously%20achieving%20multiplicative%20biases%20consistent%20with%20zero%20across%20a%20wide%20range%20of%20noise%20levels%2C%20PSF%20sizes%20and%20ellipticities%2C%20and%20magnitude%20selection%20cuts%2C%20with%20all%20measurements%20satisfying%20%24%7Cm%7C%20%7B%26lt%3B%7D%202%20%5C%5Ctimes%2010%5E%7B-3%7D%24%20%28i.e.%2C%20within%20the%200.2%25%20LSST%20requirement%29%20and%20most%20at%20the%20%24%7B%5C%5Csim%7D10%5E%7B-4%7D%24%20level.%20We%20demonstrate%20this%20framework%20on%20isolated%20single-band%20galaxy%20images%20with%20Gaussian%20noise%20and%20known%20PSF%2C%20establishing%20a%20rigorous%2C%20physics-informed%20foundation%20for%20future%20extensions%20of%20ML-based%20shear%20estimation%20to%20blended%20sources%20and%20multi-band%20observations%20in%20Stage-IV%20surveys.%20All%20codes%20and%20data%20products%20will%20be%20made%20publicly%20available%20upon%20acceptance.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-08-04T00%3A57%3A11Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2603.19046%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2603.19046%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A33%3A08Z%22%7D%7D%2C%7B%22key%22%3A%22RYUDW2VS%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Yang%20et%20al.%22%2C%22parsedDate%22%3A%222026-07-20%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BYang%2C%20Y.%2C%20Zhao%2C%20Y.%2C%20Dasgupta%2C%20A.%20%26amp%3B%20Misailovic%2C%20S.%20PagedWeight%3A%20Efficient%20MoE%20LLM%20Serving%20with%20Dynamic%20Quality-Aware%20Weight%20Quantization.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.16184%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.16184%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22PagedWeight%3A%20Efficient%20MoE%20LLM%20Serving%20with%20Dynamic%20Quality-Aware%20Weight%20Quantization%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yuchen%22%2C%22lastName%22%3A%22Yang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yifan%22%2C%22lastName%22%3A%22Zhao%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Anisha%22%2C%22lastName%22%3A%22Dasgupta%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Sasa%22%2C%22lastName%22%3A%22Misailovic%22%7D%5D%2C%22abstractNote%22%3A%22Mixture-of-Experts%20%28MoE%29%20is%20a%20popular%20class%20of%20large%20language%20models%20%28LLMs%29%2C%20offering%20high%20efficiency%20and%20accuracy.%20However%2C%20in%20KV-cache-intensive%20serving%20scenarios%2C%20MoEs%20often%20exhibit%20a%20tension%20between%20the%20GPU%20memory%20requirements%20of%20the%20model%20weights%20and%20the%20growing%20KV%20cache.%20We%20propose%20PagedWeight%2C%20a%20novel%20management%20method%20for%20MoE%20LLM%20serving%20that%20dynamically%20quantizes%20MoE%20model%26%23039%3Bs%20weights%20at%20runtime%20and%20balances%20expert-weight%20precision%20with%20the%20KV%20cache%20sizes.%20PagedWeight%20exposes%20and%20effectively%20navigates%20the%20complex%20tradeoff%20between%20the%20model%26%23039%3Bs%20task%20accuracy%2C%20memory%20consumption%2C%20and%20throughput%5C%2Flatency.%20Across%20several%20memory-sensitive%20MoE%20serving%20scenarios%2C%20PagedWeight%20improves%20the%20quality-memory%20tradeoff%20over%20several%20existing%20quantization%20baselines.%20PagedWeight%20achieves%20FP16-equivalent%20accuracy%20with%20up%20to%2072.0%25%20GPU%20memory%20savings%20and%201.94%24%5C%5Ctimes%24%20throughput%20improvement%2C%20and%20improves%20quality%20over%20quantization%20methods%20by%20up%20to%2039.3%25%20at%20a%20similar%20memory%20budget%20with%20at%20most%204.1%25%20throughput%20loss.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-07-20T00%3A54%3A47Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2607.16184%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2607.16184%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T18%3A27%3A48Z%22%7D%7D%2C%7B%22key%22%3A%22W68E72QY%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Park%20et%20al.%22%2C%22parsedDate%22%3A%222026-07-03%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BPark%2C%20J.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20Network%20architecture%20follows%20coupling%20in%20multiphysics%20systems%3A%20single%20vs.%20multiple%20branches%20in%20DeepONet%20and%20S-DeepONet.%20%26lt%3Bi%26gt%3BCommun%20Eng%26lt%3B%5C%2Fi%26gt%3B%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.1038%5C%2Fs44172-026-00714-4%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.1038%5C%2Fs44172-026-00714-4%26lt%3B%5C%2Fa%26gt%3B%20%282026%29%20doi%3A10.1038%5C%2Fs44172-026-00714-4.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22journalArticle%22%2C%22title%22%3A%22Network%20architecture%20follows%20coupling%20in%20multiphysics%20systems%3A%20single%20vs.%20multiple%20branches%20in%20DeepONet%20and%20S-DeepONet%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jaewan%22%2C%22lastName%22%3A%22Park%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Kazuma%22%2C%22lastName%22%3A%22Kobayashi%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Qibang%22%2C%22lastName%22%3A%22Liu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Diab%22%2C%22lastName%22%3A%22Abueidda%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Seid%22%2C%22lastName%22%3A%22Koric%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Syed%20Bahauddin%22%2C%22lastName%22%3A%22Alam%22%7D%5D%2C%22abstractNote%22%3A%22%22%2C%22date%22%3A%222026-07-03%22%2C%22section%22%3A%22%22%2C%22partNumber%22%3A%22%22%2C%22partTitle%22%3A%22%22%2C%22DOI%22%3A%2210.1038%5C%2Fs44172-026-00714-4%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Fwww.nature.com%5C%2Farticles%5C%2Fs44172-026-00714-4%22%2C%22PMID%22%3A%22%22%2C%22PMCID%22%3A%22%22%2C%22ISSN%22%3A%222731-3395%22%2C%22language%22%3A%22en%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T13%3A55%3A19Z%22%7D%7D%2C%7B%22key%22%3A%229QDPSFSL%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Wang%20et%20al.%22%2C%22parsedDate%22%3A%222026-07-03%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BWang%2C%20H.%2C%20Tian%2C%20J.%2C%20Arora%2C%20S.%20%26amp%3B%20Watanabe%2C%20S.%20An%20Efficient%20vLLM-Based%20Inference%20Pipeline%20for%20Unified%20Audio%20Understanding%20and%20Generation.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.02119%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2607.02119%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22An%20Efficient%20vLLM-Based%20Inference%20Pipeline%20for%20Unified%20Audio%20Understanding%20and%20Generation%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Haoran%22%2C%22lastName%22%3A%22Wang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jinchuan%22%2C%22lastName%22%3A%22Tian%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Siddhant%22%2C%22lastName%22%3A%22Arora%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shinji%22%2C%22lastName%22%3A%22Watanabe%22%7D%5D%2C%22abstractNote%22%3A%22While%20Large%20Multimodal%20Models%20excel%20in%20comprehension%2C%20high-throughput%20inference%20engines%20lack%20native%20support%20for%20multimodal%20generation.%20This%20is%20severe%20in%20Speech%20Language%20Models%2C%20where%20generating%20multi-layered%20audio%20tokens%20via%20decoupled%20AR%2BNAR%20or%20synchronous%20Multi-Token%20Prediction%20%28MTP%29%20with%20delay-pattern%20interleaving%20conflicts%20with%20standard%20single-stream%20loops.%20We%20present%20a%20vLLM-based%20inference%20pipeline%20for%20unified%20speech%20understanding%20and%20generation.%20We%20extend%20autoregressive%20decoding%20to%20natively%20execute%20delay-pattern%20de-interleaving%20and%20coordinated%20multi-stream%20sampling%2C%20integrating%20an%20on-GPU%20acoustic%20decoder%20for%20end-to-end%20waveform%20synthesis.%20Crucially%2C%20we%20overcome%20the%20shared%20intuition%20that%20Classifier-Free%20Guidance%20%28CFG%29%20halves%20throughput.%20By%20co-scheduling%20paired%20conditional%20and%20unconditional%20requests%20within%20a%20continuous%20batch%2C%20our%20CFG%20implementation%20sustains%2080%25%20of%20non-CFG%20throughput%2C%20absorbing%20dual-request%20and%20logit%20merging%20overheads.%20We%20open-source%20our%20framework.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-07-03T00%3A49%3A42Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2607.02119%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2607.02119%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T13%3A49%3A25Z%22%7D%7D%2C%7B%22key%22%3A%22GGLEXEWM%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Remy%20et%20al.%22%2C%22parsedDate%22%3A%222026-07-01%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BRemy%2C%20B.%2C%20Chang%2C%20C.%20%26amp%3B%20Willett%2C%20R.%20Joint%20inference%20of%20weak%20lensing%20convergence%20map%20and%20cosmology%20with%20diffusion%20models.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.31988%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.31988%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Joint%20inference%20of%20weak%20lensing%20convergence%20map%20and%20cosmology%20with%20diffusion%20models%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Benjamin%22%2C%22lastName%22%3A%22Remy%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Chihway%22%2C%22lastName%22%3A%22Chang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Rebecca%22%2C%22lastName%22%3A%22Willett%22%7D%5D%2C%22abstractNote%22%3A%22We%20present%20a%20method%20for%20joint%20inference%20of%20cosmological%20parameters%20and%20convergence%20maps%20from%20weak%20lensing%20observations%2C%20targeting%20the%20full%20posterior%20conditioned%20on%20the%20observed%20shear%20field.%20Our%20approach%20uses%20implicit%20inference%20with%20diffusion%20models%2C%20learning%20the%20joint%20distribution%20from%20simulations%2C%20without%20the%20need%20to%20have%20an%20explicit%20and%20differentiable%20forward%20model%20for%20gradient-based%20MCMC%20sampling.%20We%20introduce%20a%20transformer-based%20architecture%20that%20operates%20in%20pixel%20space%20and%20treats%20cosmological%20parameters%20as%20additional%20tokens%20in%20a%20unified%20sequence%2C%20enabling%20efficient%20multimodal%20processing%20within%20a%20single%20network.%20At%20inference%20time%2C%20the%20trained%20model%20generates%20posterior%20samples%20of%20joint%20convergence%20maps%20and%20cosmological%20parameters%20conditioned%20on%20observed%20noisy%20shear%20fields.%20We%20demonstrate%20the%20method%20on%20simulated%20weak%20lensing%20data%20generated%20from%20log-normal%20fields%20in%20a%20wcdm%20cosmology.%20The%20model%20accurately%20reconstructs%20convergence%20maps%20and%20recovers%20cosmological%20posteriors%20that%20agree%20with%20traditional%20MCMC%2C%20while%20remaining%20well%20calibrated%20across%20the%20prior%2C%20with%20a%20MIRA%20calibration%20score%20of%20%240.635%20%5C%5Cpm%200.017%24%20on%20the%20joint%20posterior%20%28where%20%240.667%24%20is%20optimal%29.%20The%20inferred%20fields%20reproduce%20the%20correct%20two-point%20statistics%20as%20well%20as%20non-Gaussian%20statistics%20such%20as%20the%20one-point%20distribution.%20This%20work%20establishes%20diffusion-based%20implicit%20inference%20as%20a%20viable%20route%20toward%20full%20field-level%20cosmological%20analyses%2C%20paving%20the%20way%20for%20applications%20to%20more%20realistic%2C%20non-differentiable%20simulators.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-07-01T01%3A16%3A39Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2606.31988%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.31988%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T13%3A48%3A12Z%22%7D%7D%2C%7B%22key%22%3A%22HAFH5VJQ%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Kim%20and%20Kumar%22%2C%22parsedDate%22%3A%222026-07-01%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BKim%2C%20T.%20%26amp%3B%20Kumar%2C%20A.%20Addressing%20Over-Refusal%20in%20LLMs%20with%20Competing%20Rewards.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.31748%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.31748%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Addressing%20Over-Refusal%20in%20LLMs%20with%20Competing%20Rewards%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Taeyoun%22%2C%22lastName%22%3A%22Kim%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Aviral%22%2C%22lastName%22%3A%22Kumar%22%7D%5D%2C%22abstractNote%22%3A%22Safety%20training%20on%20language%20models%20often%20induces%20over-refusal%3A%20improved%20safety%20on%20harmful%20prompts%20at%20the%20cost%20of%20increased%20refusal%20on%20harmless%20ones.%20Though%20this%20trade-off%20can%20be%20mitigated%20by%20training%20models%20with%20reinforcement%20learning%20%28RL%29%20to%20reason%20before%20answering%2C%20it%20does%20not%20remove%20the%20underlying%20problem%20that%20reasoning%20can%20often%20be%20a%20%26quot%3Brubber%20stamp%26quot%3B%20for%20a%20predetermined%20response.%20In%20this%20paper%2C%20we%20address%20the%20safety-refusal%20trade-off%20by%20rethinking%20how%20models%20are%20trained%20to%20reason%20about%20safety.%20Our%20key%20insight%20is%20that%20unsafe%20reasoning%20can%20itself%20serve%20as%20a%20useful%20exploratory%20signal.%20Rather%20than%20preemptively%20blocking%20harmful%20thoughts%2C%20we%20encourage%20the%20model%20to%20sufficiently%20explore%20unsafe%20reasoning%20but%20produce%20a%20safe%20response.%20The%20harmful%20exploration%20improves%20the%20model%26%23039%3Bs%20ability%20to%20distinguish%20harmful%20from%20harmless%20prompts%20by%20resolving%20ambiguity%2C%20allowing%20it%20to%20remain%20safe%20while%20complying%20only%20when%20appropriate.%20We%20cast%20this%20as%20an%20adversarial%20optimization%20problem%20in%20which%20a%20reasoning%20player%20explores%20strategies%20for%20producing%20an%20unsafe%20response%20and%20an%20answer%20player%20ensures%20that%20the%20final%20output%20is%20safe.%20We%20train%20a%20single%20model%20with%20dense%20rewards%20to%20play%20both%20roles%20within%20one%20chain-of-thought%2C%20across%20different%20segments.%20To%20achieve%20this%2C%20we%20find%20that%20process%20rewards%20are%20crucial%20for%20stable%20optimization%20of%20competing%20objectives.%20Our%20resulting%20model%20SEAR%20deliberately%20engages%20in%20harmful%20reasoning%20as%20exploration%20while%20reliably%20flipping%20back%20to%20a%20safe%20answer.%20We%20demonstrate%20that%20this%20behavior%20helps%20mitigate%20over-refusal%20and%20defend%20against%20attacks%20that%20directly%20manipulate%20the%20reasoning%20to%20be%20harmful.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-07-01T01%3A04%3A39Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2606.31748%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.31748%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T13%3A43%3A26Z%22%7D%7D%2C%7B%22key%22%3A%22MIM53W2C%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Chen%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BChen%2C%20Z.%2C%20Mayne%20i%20Comas%2C%20O.%2C%20Jin%2C%20Z.%2C%20Luo%2C%20D.%20%26amp%3B%20Soljacic%2C%20M.%20L-CUBE%3A%20Isolating%20Long-Context%20Capacity%20from%20Knowledge%20with%20Controllable%20Mutual%20Information%20Scaling.%20in%20%26lt%3Bi%26gt%3BProceedings%20of%20the%2043rd%20International%20Conference%20on%20Machine%20Learning%20%28ICML%29%26lt%3B%5C%2Fi%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22conferencePaper%22%2C%22title%22%3A%22L-CUBE%3A%20Isolating%20Long-Context%20Capacity%20from%20Knowledge%20with%20Controllable%20Mutual%20Information%20Scaling%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zhuo%22%2C%22lastName%22%3A%22Chen%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Oriol%22%2C%22lastName%22%3A%22Mayne%20i%20Comas%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zhuotao%22%2C%22lastName%22%3A%22Jin%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Di%22%2C%22lastName%22%3A%22Luo%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Marin%22%2C%22lastName%22%3A%22Soljacic%22%7D%5D%2C%22abstractNote%22%3A%22%22%2C%22proceedingsTitle%22%3A%22Proceedings%20of%20the%2043rd%20International%20Conference%20on%20Machine%20Learning%20%28ICML%29%22%2C%22conferenceName%22%3A%22ICML%22%2C%22date%22%3A%222026%22%2C%22eventPlace%22%3A%22%22%2C%22DOI%22%3A%22%22%2C%22ISBN%22%3A%22%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Fopenreview.net%5C%2Fpdf%3Fid%3DU3QqOOZK6P%22%2C%22ISSN%22%3A%22%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T13%3A35%3A26Z%22%7D%7D%2C%7B%22key%22%3A%22UHJBBVUJ%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Zhao%20et%20al.%22%2C%22parsedDate%22%3A%222026-07-01%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BZhao%2C%20D.%2C%20Cao%2C%20S.%2C%20Wang%2C%20Y.-X.%20%26amp%3B%20Gui%2C%20L.-Y.%20Think%20in%20Latent%2C%20Explain%20in%20Language%3A%20Self-Explainable%20Latent%20Reasoning.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2608.13570%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2608.13570%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Think%20in%20Latent%2C%20Explain%20in%20Language%3A%20Self-Explainable%20Latent%20Reasoning%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Dayuan%22%2C%22lastName%22%3A%22Zhao%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shengcao%22%2C%22lastName%22%3A%22Cao%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yu-Xiong%22%2C%22lastName%22%3A%22Wang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Liang-Yan%22%2C%22lastName%22%3A%22Gui%22%7D%5D%2C%22abstractNote%22%3A%22Latent%20reasoning%20has%20emerged%20as%20a%20powerful%20alternative%20to%20text-based%20Chain-of-Thought%20%28CoT%29%2C%20offering%20significant%20gains%20in%20computational%20efficiency%20by%20compressing%20verbose%20reasoning%20into%20compact%20embeddings.%20However%2C%20compressing%20reasoning%20into%20the%20latent%20space%20renders%20the%20thinking%20opaque%2C%20hindering%20its%20interpretability.%20Current%20methods%20present%20a%20stark%20trade-off%3A%20they%20either%20function%20as%20unexplainable%20%26%23039%3B%26%23039%3Bblack%20boxes%26%23039%3B%26%23039%3B%20%28e.g.%2C%20Coconut%29%2C%20where%20the%20latent%20reasoning%20is%20not%20human-readable%2C%20or%20rely%20on%20separate%20post-hoc%20decoders%20for%20explainability%20%28e.g.%2C%20Heima%29%2C%20introducing%20architectural%20overhead%20and%20decoupling%20the%20explanation%20from%20the%20actual%20reasoning%20process.%20In%20this%20work%2C%20we%20present%20a%20unified%20framework%20for%20Self-Explainable%20Latent%20Reasoning%20%28SELR%29%20that%20trains%20a%20single%20model%20to%20perform%20efficient%20and%20inherently%20explainable%20latent%20reasoning.%20Our%20core%20contribution%20is%20a%20novel%20multi-task%20training%20objective%20that%20optimizes%20for%20two%20goals%20simultaneously%3A%20%281%29%20an%20Answer%20Loss%20that%20optimizes%20the%20latent%20reasoning%20trajectory%20to%20produce%20accurate%20final%20answers%2C%20and%20%282%29%20a%20CoT%20Loss%20that%20explicitly%20trains%20the%20same%20model%20to%20decode%20its%20own%20latent%20representations%20back%20into%20human-understandable%20reasoning%20steps.%20This%20design%20ensures%20that%20generated%20latent%20representations%20are%20both%20task-effective%20and%20semantically%20interpretable%2C%20eliminating%20the%20need%20for%20external%20decoders.%20We%20validate%20the%20effectiveness%20of%20SELR%20on%20both%20Large%20Language%20Models%20%28LLMs%29%20and%20Vision-Language%20Models%20%28VLMs%29%2C%20demonstrating%20that%20SELR%20achieves%20superior%20token%20efficiency%20and%20accuracy%20compared%20to%20baselines%2C%20while%20uniquely%20providing%20self-contained%20explainability%20without%20auxiliary%20models.%20Project%20page%20is%20available%20at%20https%3A%5C%2F%5C%2Fjasondayuan.github.io%5C%2FSELR%5C%2F.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22arXiv%3A2608.13570%22%2C%22date%22%3A%222026-07-01%22%2C%22DOI%22%3A%2210.48550%5C%2FarXiv.2608.13570%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22http%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2608.13570%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T13%3A28%3A21Z%22%7D%7D%2C%7B%22key%22%3A%22RIJJZD7L%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Lian%20et%20al.%22%2C%22parsedDate%22%3A%222026-07-01%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BLian%2C%20X.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20Quantization%20Inflates%20Reasoning%3A%20Token%20Inflation%20as%20a%20Hidden%20Cost%20of%20Low-Bit%20Reasoning%20Models.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.25519%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.25519%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Quantization%20Inflates%20Reasoning%3A%20Token%20Inflation%20as%20a%20Hidden%20Cost%20of%20Low-Bit%20Reasoning%20Models%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Xinyu%22%2C%22lastName%22%3A%22Lian%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Walid%22%2C%22lastName%22%3A%22Krichene%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Beichen%22%2C%22lastName%22%3A%22Huang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Masahiro%22%2C%22lastName%22%3A%22Tanaka%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Olatunji%22%2C%22lastName%22%3A%22Ruwase%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Li%22%2C%22lastName%22%3A%22Zhang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Minjia%22%2C%22lastName%22%3A%22Zhang%22%7D%5D%2C%22abstractNote%22%3A%22Quantization%20is%20widely%20used%20to%20reduce%20the%20inference%20cost%20of%20large%20language%20models%2C%20but%20its%20effect%20on%20reasoning%20models%20is%20not%20fully%20captured%20by%20final-answer%20accuracy%20or%20per-token%20latency.%20We%20show%20that%20low-bit%20post-training%20quantization%20can%20introduce%20a%20hidden%20test-time%20compute%20cost%3A%20quantized%20reasoning%20models%20often%20generate%20longer%20chains%20of%20thought%20even%20when%20they%20still%20answer%20correctly.%20Across%20mathematical%20reasoning%2C%20code%20generation%2C%20scientific%20question%20answering%2C%20and%20agentic%20tool-use%20benchmarks%2C%20we%20find%20that%20INT4%5C%2FINT3%20quantization%20can%20preserve%20accuracy%20but%20increase%20reasoning-token%20usage%2C%20offsetting%20the%20expected%20per-token%20speedup.%20To%20measure%20this%20effect%2C%20we%20introduce%20the%20CoT%20Token%20Inflation%20Ratio%2C%20which%20compares%20reasoning%20length%20between%20quantized%20and%20full-precision%20models%20averaged%20across%20all%20evaluation%20benchmarks.%20We%20further%20show%20that%20token%20inflation%20is%20accompanied%20by%20behavioral%20changes%20in%20the%20reasoning%20trace%2C%20including%20more%20intermediate%20steps%20and%20greater%20semantic%20repetition.%20These%20changes%20translate%20into%20measurable%20end-to-end%20real-world%20serving%20penalties.%20Finally%2C%20we%20evaluate%20mitigation%20strategies%20and%20find%20that%20prompting%20and%20decoding-time%20sampling%20offer%20inconsistent%20accuracy-length%20trade-offs%2C%20while%20quantization-aware%20training%20shows%20more%20promise%20in%20reducing%20both%20accuracy%20degradation%20and%20token%20inflation.%20Our%20results%20suggest%20that%20reasoning-token%20usage%20should%20be%20reported%20alongside%20accuracy%20when%20evaluating%20quantized%20reasoning%20models.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-07-01T00%3A09%3A06Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2606.25519%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.25519%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T13%3A24%3A18Z%22%7D%7D%2C%7B%22key%22%3A%22CT8STTL6%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Tian%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BTian%2C%20J.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20Bagpiper-TTS%3A%20Natural%20Language%20Guided%20Universal%20Speech%20Synthesis.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.22811%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.22811%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Bagpiper-TTS%3A%20Natural%20Language%20Guided%20Universal%20Speech%20Synthesis%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jinchuan%22%2C%22lastName%22%3A%22Tian%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Haoran%22%2C%22lastName%22%3A%22Wang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Siddhant%22%2C%22lastName%22%3A%22Arora%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Takashi%22%2C%22lastName%22%3A%22Maekaku%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Keita%22%2C%22lastName%22%3A%22Goto%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jin%22%2C%22lastName%22%3A%22Sakuma%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yusuke%22%2C%22lastName%22%3A%22Shinohara%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Chao-Han%20Huck%22%2C%22lastName%22%3A%22Yang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shinji%22%2C%22lastName%22%3A%22Watanabe%22%7D%5D%2C%22abstractNote%22%3A%22Classical%20TTS%20systems%20typically%20rely%20on%20rigid%20input%20formats%20and%20predefined%20metadata%20slots%2C%20limiting%20their%20ability%20to%20fulfill%20flexible%20user%20requirements.%20This%20paper%20introduces%20Bagpiper-TTS%2C%20a%20universal%20speech%20synthesis%20system%20that%20deals%20with%20diverse%20natural%20language%20user%20requests.%20Given%20a%20natural%20language%20prompt%2C%20Bagpiper-TTS%20first%20reasons%20over%20the%20users%26%23039%3B%20intent%20to%20derive%20a%20rich%20caption%2C%20i.e.%2C%20a%20comprehensive%20textual%20blueprint%20encompassing%20both%20transcription%20and%20nuanced%20metadata.%20Subsequently%2C%20this%20caption%20guides%20the%20synthesis%20of%20the%20target%20speech.%20Our%20model%20inherently%20supports%20a%20broad%20spectrum%20of%20tasks%20besides%20classical%20TTS%20applications%2C%20including%20multi-talker%2C%20intent-to-speech%2C%20role-play%20synthesis%2C%20singing%20voice%20synthesis%2C%20and%20more.%20Experimental%20results%20demonstrate%20that%20Bagpiper-TTS%20achieves%20an%201.7%25%20Word%20Error%20Rate%20%28WER%29%20on%20the%20Seed-TTS-Eval%20benchmark%20and%20match%20the%20performance%20of%20dedicated%20models%20in%20both%20LLM-as-a-judge%20and%20human%20subjective%20evaluations%20across%20multiple%20applications.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2606.22811%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.22811%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T13%3A22%3A45Z%22%7D%7D%2C%7B%22key%22%3A%22EQXN53RS%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Tran%20et%20al.%22%2C%22parsedDate%22%3A%222026-05%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BTran%2C%20B.%2C%20Maithert%2C%20M.%2C%20Shin%2C%20W.%2C%20Sinclair%2C%20M.%20D.%20%26amp%3B%20Venkataraman%2C%20S.%20ET%3A%20Bridging%20the%20Gap%20on%20Energy%20Telemetry%20for%20Multi-GPU%20Communication%20Collectives.%20in%20%26lt%3Bi%26gt%3BProceedings%20of%20the%2016th%20International%20Workshop%20on%20Accelerators%20and%20Hybrid%20Emerging%20Systems%20%28AsHES%29%26lt%3B%5C%2Fi%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22conferencePaper%22%2C%22title%22%3A%22ET%3A%20Bridging%20the%20Gap%20on%20Energy%20Telemetry%20for%20Multi-GPU%20Communication%20Collectives%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Brandon%22%2C%22lastName%22%3A%22Tran%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Matthias%22%2C%22lastName%22%3A%22Maithert%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Woong%22%2C%22lastName%22%3A%22Shin%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Matthew%20D.%22%2C%22lastName%22%3A%22Sinclair%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shivaram%22%2C%22lastName%22%3A%22Venkataraman%22%7D%5D%2C%22abstractNote%22%3A%22%22%2C%22proceedingsTitle%22%3A%22Proceedings%20of%20the%2016th%20International%20Workshop%20on%20Accelerators%20and%20Hybrid%20Emerging%20Systems%20%28AsHES%29%22%2C%22conferenceName%22%3A%22AsHES%22%2C%22date%22%3A%222026-05%22%2C%22eventPlace%22%3A%22%22%2C%22DOI%22%3A%22%22%2C%22ISBN%22%3A%22%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Fpages.cs.wisc.edu%5C%2F~sinclair%5C%2Fpapers%5C%2Fbtran-ashes26-et.pdf%22%2C%22ISSN%22%3A%22%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T13%3A18%3A41Z%22%7D%7D%2C%7B%22key%22%3A%22YYWHCRPU%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Liang%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BLiang%2C%20W.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20MedQPA-Gen%3A%20Medical%20Question%20Proposing%20and%20Answering%20for%20Report%20Generation.%20in%20%26lt%3Bi%26gt%3BFindings%20of%20the%20Association%20for%20Computational%20Linguistics%3A%20ACL%202026%26lt%3B%5C%2Fi%26gt%3B%2043120%26%23x2013%3B43132%20%28Association%20for%20Computational%20Linguistics%2C%20San%20Diego%2C%20California%2C%20United%20States%2C%202026%29.%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttp%3A%5C%2F%5C%2Fdoi.org%5C%2F10.18653%5C%2Fv1%5C%2F2026.findings-acl.2139%26%23039%3B%26gt%3Bhttp%3A%5C%2F%5C%2Fdoi.org%5C%2F10.18653%5C%2Fv1%5C%2F2026.findings-acl.2139%26lt%3B%5C%2Fa%26gt%3B.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22conferencePaper%22%2C%22title%22%3A%22MedQPA-Gen%3A%20Medical%20Question%20Proposing%20and%20Answering%20for%20Report%20Generation%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Weijie%22%2C%22lastName%22%3A%22Liang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Xiyue%22%2C%22lastName%22%3A%22Zhu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ruike%22%2C%22lastName%22%3A%22Zhu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Chenhao%22%2C%22lastName%22%3A%22Li%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Cheng%22%2C%22lastName%22%3A%22Tang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zhiyu%22%2C%22lastName%22%3A%22Liu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zhihua%22%2C%22lastName%22%3A%22Gong%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shirui%22%2C%22lastName%22%3A%22Luo%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yudu%22%2C%22lastName%22%3A%22Li%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Volodymyr%22%2C%22lastName%22%3A%22Kindratenko%22%7D%5D%2C%22abstractNote%22%3A%22%22%2C%22proceedingsTitle%22%3A%22Findings%20of%20the%20Association%20for%20Computational%20Linguistics%3A%20ACL%202026%22%2C%22conferenceName%22%3A%22Findings%20of%20the%20Association%20for%20Computational%20Linguistics%3A%20ACL%202026%22%2C%22date%22%3A%222026%22%2C%22eventPlace%22%3A%22%22%2C%22DOI%22%3A%2210.18653%5C%2Fv1%5C%2F2026.findings-acl.2139%22%2C%22ISBN%22%3A%22%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Faclanthology.org%5C%2F2026.findings-acl.2139%22%2C%22ISSN%22%3A%22%22%2C%22language%22%3A%22en%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T13%3A06%3A01Z%22%7D%7D%2C%7B%22key%22%3A%22R4VY6UP9%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Wang%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BWang%2C%20Z.%2C%20Luo%2C%20X.%2C%20Zeng%2C%20S.%20%26amp%3B%20Yan%2C%20X.%20Dual%20Dimensionality%20for%20Local%20and%20Global%20Attention.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.18587%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.18587%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Dual%20Dimensionality%20for%20Local%20and%20Global%20Attention%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zhiyuan%22%2C%22lastName%22%3A%22Wang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Xuan%22%2C%22lastName%22%3A%22Luo%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Sirui%22%2C%22lastName%22%3A%22Zeng%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Xifeng%22%2C%22lastName%22%3A%22Yan%22%7D%5D%2C%22abstractNote%22%3A%22Decoder-only%20Transformers%20compute%20attention%20over%20the%20KV%20cache%20of%20preceding%20tokens.%20Keys%20%28and%20Values%29%20are%20typically%20represented%20with%20the%20same%20dimensionality%2C%20regardless%20of%20its%20distance%20from%20the%20prediction%20target.%20In%20natural%20language%2C%20however%2C%20the%20next%20word%20is%20most%20strongly%20influenced%20by%20the%20immediately%20preceding%20tokens.%20We%20hypothesize%20that%20local%20and%20distant%20tokens%20impose%20asymmetric%20demands%20on%20representational%20capacity%3A%20local%20tokens%20are%20more%20critical%20for%20predicting%20immediate%20outputs%20and%20thus%20require%20richer%20representations%2C%20whereas%20distant%20tokens%20primarily%20serve%20as%20long-range%20memory%2C%20for%20which%20lower-dimensional%20representations%20may%20suffice.%20We%20formalize%20this%20idea%20as%20Distance-Adaptive%20Representation%20%28DAR%29%2C%20implemented%20in%20a%20controlled%20setting%20that%20preserves%20full-dimensional%20representations%20within%20a%20local%20context%20window%20while%20assigning%20reduced-dimensional%20representations%20%28e.g.%201%5C%2F4%20of%20the%20original%20dimensionality%29%20to%20tokens%20beyond%20that%20window.%20Across%20multiple%20pretraining%20scales%20%2870M%20to%20410M%20parameters%29%2C%20as%20well%20as%20continued%20supervised%20fine-tuning%20on%20a%201B-scale%20model%2C%20this%20approach%20closely%20matches%20the%20performance%20of%20full-dimensional%20baselines.%20In%20contrast%2C%20uniformly%20reducing%20dimensionality%20across%20all%20token%20positions%20leads%20to%20worse%20performance.%20These%20results%20challenge%20the%20common%20assumption%20that%20key%20and%20value%20dimensionality%20should%20be%20uniform%20across%20token%20positions.%20Our%20findings%20suggest%20a%20new%20direction%20for%20designing%20attention%20architectures%20that%20adaptively%20allocate%20representational%20capacity%20across%20sequences%2C%20enabling%20further%20reductions%20in%20KV%20cache%20during%20inference.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2606.18587%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.18587%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T13%3A00%3A04Z%22%7D%7D%2C%7B%22key%22%3A%226M2DMKQM%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Daithankar%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BDaithankar%2C%20N.%2C%20Gladstone%2C%20A.%2C%20LeCun%2C%20Y.%20%26amp%3B%20Ji%2C%20H.%20You%20Don%26%23x2019%3Bt%20Need%20Strong%20Assumptions%3A%20Visual%20Representation%20Learning%20via%20Temporal%20Differences.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.15956%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.15956%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22You%20Don%27t%20Need%20Strong%20Assumptions%3A%20Visual%20Representation%20Learning%20via%20Temporal%20Differences%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ninad%22%2C%22lastName%22%3A%22Daithankar%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Alexi%22%2C%22lastName%22%3A%22Gladstone%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yann%22%2C%22lastName%22%3A%22LeCun%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Heng%22%2C%22lastName%22%3A%22Ji%22%7D%5D%2C%22abstractNote%22%3A%22Progress%20in%20AI%20has%20largely%20been%20driven%20by%20methods%20that%20assume%20less.%20As%20compute%20and%20data%20increase%2C%20approaches%20with%20weaker%20inductive%20biases%20generally%20outperform%20those%20with%20stronger%20assumptions.%20This%20is%20particularly%20characteristic%20of%20the%20field%20of%20Visual%20Representation%20Learning%2C%20where%20approaches%20have%20gone%20from%20being%20dominated%20by%20Supervised%20Learning%2C%20to%20Weakly%20Supervised%20Learning%2C%20to%20the%20now%20widespread%20success%20of%20Self-Supervised%20Learning%20without%20human%20labels.%20Yet%2C%20even%20modern%20Self-Supervised%20Learning%20approaches%20still%20depend%20on%20strong%20inductive%20biases%20such%20as%20augmentations%2C%20masking%2C%20or%20cropping.%20If%20this%20trend%20holds%2C%20even%20these%20remaining%20biases%20should%20become%20bottlenecks%20at%20scale%20--%20and%20our%20experiments%20confirm%20this%3A%20the%20optimal%20strength%20of%20inductive%20biases%20decreases%20as%20data%20grows.%20This%20motivates%20the%20search%20for%20approaches%20that%20rely%20on%20fewer%20assumptions.%20To%20this%20end%2C%20we%20introduce%20Temporal%20Difference%20in%20Vision%20%28TDV%29%2C%20a%20new%20paradigm%20for%20self-supervised%20learning%20from%20video%20that%20avoids%20existing%20inductive%20biases%2C%20relying%20instead%20on%20a%20causal%20assumption%20that%20the%20past%20causes%20the%20future.%20TDV%20functions%20by%20jointly%20training%20an%20image%20encoder%20and%20a%20motion%20encoder%20so%20that%20the%20current%20frame%26%23039%3Bs%20representation%20plus%20the%20encoded%20motion%20equals%20the%20next%20frame%26%23039%3Bs%20representation.%20Despite%20not%20leveraging%20any%20strong%20inductive%20biases%2C%20TDV%20matches%20state-of-the-art%20recipes%20on%20dense%20spatial%20tasks%2C%20laying%20the%20foundation%20for%20representation%20learning%20without%20strong%20assumptions.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2606.15956%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.15956%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A57%3A10Z%22%7D%7D%2C%7B%22key%22%3A%22KKJDBIZN%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Yoo%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BYoo%2C%20J.%20P.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20Towards%20Data-Efficient%20Cross-Device%20Generalization%20of%20Grad-Shafranov%20Equilibria%20via%20Transfer%20Learning%20Neural%20Operator.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.15512%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.15512%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Towards%20Data-Efficient%20Cross-Device%20Generalization%20of%20Grad-Shafranov%20Equilibria%20via%20Transfer%20Learning%20Neural%20Operator%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jay%20Phil%22%2C%22lastName%22%3A%22Yoo%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22William%22%2C%22lastName%22%3A%22Howes%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yashika%22%2C%22lastName%22%3A%22Ghai%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Kazuma%22%2C%22lastName%22%3A%22Kobayashi%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Souvik%22%2C%22lastName%22%3A%22Chakraborty%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Syed%20Bahauddin%22%2C%22lastName%22%3A%22Alam%22%7D%5D%2C%22abstractNote%22%3A%22Real-time%20reconstruction%20of%20magnetohydrodynamic%20equilibria%20is%20essential%20for%20plasma%20shaping%2C%20stability%20assessment%20and%20feedback%20control%20in%20magnetic%20confinement%20fusion.%20However%2C%20Grad-Shafranov%20equilibrium%20calculations%20remain%20largely%20device-specific%20and%20iterative%2C%20limiting%20their%20use%20in%20latency-constrained%20control%20settings.%20Existing%20neural%20approaches%20can%20accelerate%20individual%20equilibrium%20predictions%2C%20but%20they%20do%20not%20generally%20provide%20reusable%20models%20across%20changing%20plasma%20boundaries%20or%20tokamak%20geometries.%20Here%20we%20show%20that%20equilibrium%20reconstruction%20can%20be%20recast%20as%20a%20cross-device%20operator%20learning%20problem.%20We%20develop%20a%20domain-specific%20neural%20operator%20framework%20that%20maps%20geometry%20and%20profile%20parameters%20directly%20to%20the%20poloidal%20flux%20field%2C%20replacing%20repeated%20solve-on-demand%20computation%20with%20amortized%20operator%20inference.%20Using%20the%20analytically%20tractable%20Solov%26%23039%3Bev%20family%20as%20a%20controlled%20Grad-Shafranov%20testbed%2C%20we%20generate%20equilibria%20across%20eight%20geometrically%20distinct%20tokamak-like%20configurations%20and%20benchmark%20five%20neural%20operator%20architectures%20under%20four%20transfer-learning%20strategies.%20Single-geometry%20pretraining%20gives%20poor%20transfer%20to%20unseen%20devices%2C%20whereas%20multi-geometry%20pretraining%20enables%20data-efficient%20adaptation.%20The%20Wavelet%20Neural%20Operator%20gives%20the%20strongest%20cross-geometry%20performance%2C%20reaching%20mean%20relative%20L2%20errors%20below%204%25%20with%20100%20labelled%20target%20equilibria%20and%20below%202%25%20with%20full%20fine-tuning.%20The%20predicted%20magnetic%20fields%20satisfy%20the%20divergence-free%20constraint%20to%20numerical%20precision%2C%20and%20four%20architectures%20achieve%20millisecond%20or%20sub-millisecond%20inference.%20These%20results%20identify%20neural%20operator%20pretraining%20as%20a%20route%20towards%20reusable%2C%20real-time%20equilibrium%20inference%20across%20fusion%20device%20configurations.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2606.15512%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.15512%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A55%3A13Z%22%7D%7D%2C%7B%22key%22%3A%22LE9K2NPW%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Tam%22%2C%22parsedDate%22%3A%222026-06-07%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BTam%2C%20W.%20K.%20The%20Amplifying%20Mirror%3A%20Locating%20and%20Steering%20the%20Partisan%20Direction%20inside%20a%20Large%20Language%20Model.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2606.08792%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FarXiv.2606.08792%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22The%20Amplifying%20Mirror%3A%20Locating%20and%20Steering%20the%20Partisan%20Direction%20inside%20a%20Large%20Language%20Model%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Wendy%20K.%22%2C%22lastName%22%3A%22Tam%22%7D%5D%2C%22abstractNote%22%3A%22Large%20language%20models%20are%20rapicly%20replacing%20search%20engines%20as%20the%20primary%20interface%20between%20people%20and%20information.%20Unlike%20search%20engines%2C%20which%20retrieve%20existing%20content%2C%20LLMs%20generate%20novel%20text%20shaped%20by%20internal%20representations%20learned%20during%20training.%20Here%20we%20show%20that%20partisan%20political%20identity%20is%20encoded%20in%20the%20model%26%23039%3Bs%20activation%20space%2C%20and%20that%20this%20direction%20directly%20shapes%20generation.%20Using%20190%2C491%20tweets%20from%20sitting%20members%20of%20the%20U.S.%20Congress%20as%20labeled%20training%20data%2C%20we%20train%20linear%20probes%20on%20the%20hidden%20states%20of%20the%20Llama%203.1%208B%20Instruct%20model.%20We%20identify%20a%20single%20geometric%20axis%20at%20layer%2018%20that%20separates%20Republican%20from%20Democratic%20text%20with%20an%20AUC%20of%200.945%20and%20a%20Cohen%26%23039%3Bs%20d%20of%201.94%2C%20and%20use%20sparse%20autoencoders%20to%20decompose%20that%20axis%20into%20interpretable%20partisan%20features.%20Causally%20intervening%20along%20this%20axis%2C%20ablating%20or%20amplifying%20the%20partisan%20component%20mid-generation%2C%20produces%20systematic%20shifts%20in%20the%20model%26%23039%3Bs%20output.%20We%20witness%20stance%20reversals%2C%20register%20shifting%2C%20and%20structured%20fabrications%20of%20authority.%20Our%20results%20demonstrate%20that%20partisan%20bias%20in%20language%20models%20is%20not%20a%20vague%20emergent%20property%20but%20a%20learned%20geometric%20feature%20that%20can%20be%20precisely%20located%20and%20steered.%20Partisan%20bias%20is%20not%20a%20bug%20to%20be%20patched%2C%20but%20a%20structural%20property%20of%20how%20these%20models%20encode%20information%20about%20their%20users.%20As%20LLMs%20displace%20search%20engines%20as%20the%20interface%20to%20knowledge%2C%20understanding%20that%20product%20design%20%28and%20its%20consequences%29%20will%20be%20essential%20for%20navigating%20the%20legal%2C%20social%2C%20and%20political%20transitions%20from%20an%20information%20ecosystem%20that%20is%20curated%20to%20one%20that%20is%20generated.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22arXiv%3A2606.08792%22%2C%22date%22%3A%222026-06-07%22%2C%22DOI%22%3A%2210.48550%5C%2FarXiv.2606.08792%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22http%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.08792%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A52%3A53Z%22%7D%7D%2C%7B%22key%22%3A%22P2VA59V9%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Jiang%20et%20al.%22%2C%22parsedDate%22%3A%222026-08-05%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BJiang%2C%20E.%2C%20Gj%26%23xF8%3Blbye%2C%20A.%2C%20Zhang%2C%20Y.%20J.%20%26amp%3B%20Koyejo%2C%20S.%20When%20Behavioral%20Safety%20Evaluation%20Fails%3A%20A%20Representation-Level%20Perspective.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.08044%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.08044%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22When%20Behavioral%20Safety%20Evaluation%20Fails%3A%20A%20Representation-Level%20Perspective%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Enyi%22%2C%22lastName%22%3A%22Jiang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Anders%22%2C%22lastName%22%3A%22Gj%5Cu00f8lbye%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yibo%20Jacky%22%2C%22lastName%22%3A%22Zhang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Sanmi%22%2C%22lastName%22%3A%22Koyejo%22%7D%5D%2C%22abstractNote%22%3A%22Safety%20evaluation%20of%20large%20language%20models%20%28LLMs%29%20is%20largely%20behavioral%3A%20a%20model%20is%20certified%20safe%20when%20it%20refuses%20harmful%20requests%20and%20answers%20benign%20ones.%20But%20refusing%20on%20the%20prompts%20an%20auditor%20happens%20to%20try%20does%20not%20show%20that%20the%20model%20is%20far%20from%20harmful%20behavior.%20Behavioral%20tests%20observe%20outputs%3B%20they%20do%20not%20measure%20how%20easily%20an%20intervention%20on%20the%20model%20turns%20a%20refusal%20into%20compliance.%20We%20call%20the%20gap%20between%20what%20static%20audits%20certify%20and%20what%20an%20intervention%20can%20reach%20the%20audit%20gap%2C%20and%20we%20show%20it%20is%20realizable%3A%20one%20can%20build%20a%20model%20that%20matches%20its%20safety-aligned%20base%20on%20every%20static%20audit%20yet%20gives%20way%20to%20a%20small%2C%20known%20perturbation%20of%20its%20internal%20state.%20We%20construct%20such%20dissociated%20models%20from%20three%20safety-aligned%20bases%20%28Gemma%202%202B%2C%20Llama%203.2%203B%2C%20Qwen%202.5%203B%29%20and%20audit%20the%20base%2C%20dissociated%2C%20and%20openly%20harmful%20models%20with%20the%20same%20soft%20interventions%20in%20parameter%20and%20latent%20space%3B%20the%20latent%20attacks%20are%20summarized%20by%20the%20Latent%20Vulnerability%20Score%20%28LVS%29%2C%20the%20safety%20degradation%20produced%20per%20unit%20of%20bounded%20latent%20perturbation.%20Every%20static%20audit%20we%20run%20gives%20the%20dissociated%20model%20the%20same%20verdict%20as%20its%20base%2C%20since%20its%20refusals%20match%20the%20base%2C%20jailbreaks%20show%20no%20consistent%20signature%2C%20and%20a%20strong%20fixed%20probe%20on%20clean%20activations%20cannot%20tell%20it%20from%20the%20base.%20The%20same%20interventions%20an%20auditor%20could%20run%20reverse%20the%20verdict.%20At%20the%20targeted%20mid%20layer%20the%20dissociated%20models%20score%202.5%20to%203.1%20times%20higher%20LVS%20than%20their%20bases.%20A%20bounded%20latent%20attack%20elicits%20harmful%20compliance%20on%2054%20to%2086%25%20of%20prompts%2C%20against%203%20to%2048%25%20for%20the%20bases%2C%20while%20matched%20random%20perturbations%20stay%20at%20or%20below%2012%25.%20Harmful%20fine-tuning%20reaches%20high%20compliance%20within%20five%20gradient%20steps%2C%20where%20the%20bases%20need%2010%20to%2025.%20Behavioral%20testing%2C%20even%20with%20static%20latent%20probing%2C%20cannot%20certify%20representation-level%20robustness%3A%20a%20safety%20audit%20must%20intervene%20on%20the%20model%2C%20not%20only%20observe%20it.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-08-05T00%3A51%3A58Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2606.08044%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.08044%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A51%3A01Z%22%7D%7D%2C%7B%22key%22%3A%227BL5NCFD%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Wang%20et%20al.%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BWang%2C%20Y.%2C%20Geng%2C%20R.%2C%20Chen%2C%20Y.%20%26amp%3B%20Jia%2C%20J.%20AttnTrace%3A%20Contextual%20Attribution%20of%20Prompt%20Injection%20and%20Knowledge%20Corruption.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2508.03793%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2508.03793%26lt%3B%5C%2Fa%26gt%3B%20%282025%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22AttnTrace%3A%20Contextual%20Attribution%20of%20Prompt%20Injection%20and%20Knowledge%20Corruption%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yanting%22%2C%22lastName%22%3A%22Wang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Runpeng%22%2C%22lastName%22%3A%22Geng%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ying%22%2C%22lastName%22%3A%22Chen%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jinyuan%22%2C%22lastName%22%3A%22Jia%22%7D%5D%2C%22abstractNote%22%3A%22Long-context%20large%20language%20models%20%28LLMs%29%2C%20such%20as%20Gemini-2.5-Pro%20and%20Claude-Sonnet-4%2C%20are%20increasingly%20used%20to%20empower%20advanced%20AI%20systems%2C%20including%20retrieval-augmented%20generation%20%28RAG%29%20pipelines%20and%20autonomous%20agents.%20In%20these%20systems%2C%20an%20LLM%20receives%20an%20instruction%20along%20with%20a%20context--often%20consisting%20of%20texts%20retrieved%20from%20a%20knowledge%20database%20or%20memory--and%20generates%20a%20response%20that%20is%20contextually%20grounded%20by%20following%20the%20instruction.%20Recent%20studies%20have%20designed%20solutions%20to%20trace%20back%20to%20a%20subset%20of%20texts%20in%20the%20context%20that%20contributes%20most%20to%20the%20response%20generated%20by%20the%20LLM.%20These%20solutions%20have%20numerous%20real-world%20applications%2C%20including%20performing%20post-attack%20forensic%20analysis%20and%20improving%20the%20interpretability%20and%20trustworthiness%20of%20LLM%20outputs.%20While%20significant%20efforts%20have%20been%20made%2C%20state-of-the-art%20solutions%20such%20as%20TracLLM%20often%20lead%20to%20a%20high%20computation%20cost%2C%20e.g.%2C%20it%20takes%20TracLLM%20hundreds%20of%20seconds%20to%20perform%20traceback%20for%20a%20single%20response-context%20pair.%20In%20this%20work%2C%20we%20propose%20AttnTrace%2C%20a%20new%20context%20traceback%20method%20based%20on%20the%20attention%20weights%20produced%20by%20an%20LLM%20for%20a%20prompt.%20To%20effectively%20utilize%20attention%20weights%2C%20we%20introduce%20two%20techniques%20designed%20to%20enhance%20the%20effectiveness%20of%20AttnTrace%2C%20and%20we%20provide%20theoretical%20insights%20for%20our%20design%20choice.%20We%20also%20perform%20a%20systematic%20evaluation%20for%20AttnTrace.%20The%20results%20demonstrate%20that%20AttnTrace%20is%20more%20accurate%20and%20efficient%20than%20existing%20state-of-the-art%20context%20traceback%20methods.%20We%20also%20show%20that%20AttnTrace%20can%20improve%20state-of-the-art%20methods%20in%20detecting%20prompt%20injection%20under%20long%20contexts%20through%20the%20attribution-before-detection%20paradigm.%20As%20a%20real-world%20application%2C%20we%20demonstrate%20that%20AttnTrace%20can%20effectively%20pinpoint%20injected%20instructions%20in%20a%20paper%20designed%20to%20manipulate%20LLM-generated%20reviews.%20The%20code%20is%20at%20https%3A%5C%2F%5C%2Fgithub.com%5C%2FWang-Yanting%5C%2FAttnTrace.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222025%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2508.03793%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2508.03793%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A49%3A11Z%22%7D%7D%2C%7B%22key%22%3A%22AT93I2L3%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Caplan%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BCaplan%2C%20M.%20E.%2C%20Smith%2C%20N.%20T.%2C%20Bransgrove%2C%20A.%20J.%20%26amp%3B%20Horowitz%2C%20C.%20J.%20Plasticity%20of%20Neutron%20Star%20Crusts.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.06706%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.06706%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Plasticity%20of%20Neutron%20Star%20Crusts%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Matthew%20E%22%2C%22lastName%22%3A%22Caplan%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Nevin%20T%22%2C%22lastName%22%3A%22Smith%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ashley%20J%22%2C%22lastName%22%3A%22Bransgrove%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Charles%20J%22%2C%22lastName%22%3A%22Horowitz%22%7D%5D%2C%22abstractNote%22%3A%22We%20use%20first-principles%20molecular%20dynamics%20simulations%20to%20study%20the%20deformation%20and%20breaking%20of%20neutron%20star%20crusts.%20When%20simulating%20with%20strain%20rates%20several%20orders%20of%20magnitude%20slower%20than%20prior%20work%2C%20we%20find%20a%20new%20regime%20of%20steady%20plastic%20flow%20beyond%20the%20breaking%20point%20that%20is%20independent%20of%20the%20initial%20crystal%20structure.%20Polycrystals%20exhibit%20a%20robust%20transition%20from%20linear%20elasticity%20to%20perfect%20plastic%20flow%20at%20shear%20strains%20of%20%24%5Cu03b5%3D%200.05%24%2C%20while%20monocrystals%20break%20at%20%24%5Cu03b5%3D%200.11%24%20and%20then%20flow%20plastically.%20The%20universal%20post-break%20plasticity%20may%20arise%20because%20the%20crystal%20self-consistently%20assumes%20a%20defect%20density%20to%20accommodate%20the%20imposed%20strain%20rate.%20If%20broken%20crusts%20can%20re-anneal%20to%20large%20crystal%20sizes%2C%20crust%20breaking%20may%20repeat%20with%20implications%20for%20magnetar%20bursts%20and%20flares.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2606.06706%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.06706%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A46%3A40Z%22%7D%7D%2C%7B%22key%22%3A%22X3XD3UKS%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Bai%20et%20al.%22%2C%22parsedDate%22%3A%222026-08-10%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BBai%2C%20H.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20AsyncWebRL%3A%20Efficient%20Asynchronous%20Reinforcement%20Learning%20for%20Multi-Step%20Visual%20Web%20Agents.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.05597%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.05597%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22AsyncWebRL%3A%20Efficient%20Asynchronous%20Reinforcement%20Learning%20for%20Multi-Step%20Visual%20Web%20Agents%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Hao%22%2C%22lastName%22%3A%22Bai%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Rui%22%2C%22lastName%22%3A%22Yang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Chenlu%22%2C%22lastName%22%3A%22Ye%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Spencer%22%2C%22lastName%22%3A%22Whitehead%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Aviral%22%2C%22lastName%22%3A%22Kumar%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Tong%22%2C%22lastName%22%3A%22Zhang%22%7D%5D%2C%22abstractNote%22%3A%22Training%20vision-language%20web%20agents%20with%20multi-step%20RL%20is%20compute-intensive%2C%20with%20two%20dominant%20forms%20of%20inefficiency%3A%20idle%20GPUs%20in%20synchronous%20RL%2C%20and%20trajectories%20that%20use%20more%20steps%20and%20tokens%20than%20necessary.%20We%20present%20AsyncWebRL%2C%20which%20addresses%20both.%20On%20the%20system%20side%2C%20an%20asynchronous%20design%20overlaps%20rollout%2C%20gradient%20update%2C%20and%20policy%20refresh%20across%20iterations%2C%20paired%20with%20two%20web-agent-specific%20adaptations%2C%20namely%20an%20everlasting%20rollout%20pool%20and%20lightweight%20screenshot%20handling%2C%20that%20together%20deliver%20up%20to%20a%20%242.9%5C%5Ctimes%24%20end-to-end%20training-throughput%20speedup%20over%20the%20previously%20fastest%20open%20synchronous%20pipeline%20%28WebGym%29.%20On%20the%20algorithmic%20side%2C%20we%20identify%20the%20per-trajectory%20normalizer%20%241%5C%2F%7C%5Cu03c4_i%7C%24%20in%20multi-step%20GRPO%20as%20the%20root%20cause%20of%20trajectory-level%20and%20token-level%20inefficiency%3A%20because%20failures%20are%20systematically%20longer%20than%20successes%2C%20it%20down-weights%20the%20negative%20gradient%20on%20failed%20tokens%2C%20so%20the%20policy%20keeps%20producing%20verbose%20memory%20schemas.%20Replacing%20%241%5C%2F%7C%5Cu03c4_i%7C%24%20with%20a%20constant%20%241%5C%2Fk%24%20breaks%20this%20coupling%2C%20contracting%20trajectories%20while%20preserving%20aggregate%20success.%20Together%2C%20these%20contributions%20set%20a%20new%20open-source%20state%20of%20the%20art%20on%20the%20WebGym%20out-of-distribution%20test%20split%20%28%2B5.8%25%20relative%20over%20the%2042.9%25%20prior%20best%29%2C%20with%20the%20largest%20gains%20on%20the%20harder%20slices%20%28%2B42%25%20relative%20on%20Medium%2C%20%2B48%25%20relative%20on%20Hard%29.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026-08-10T00%3A58%3A01Z%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2606.05597%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.05597%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A35%3A48Z%22%7D%7D%2C%7B%22key%22%3A%223S68ERTZ%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22La%20et%20al.%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BLa%2C%20H.%2C%20Gupta%2C%20A.%2C%20Morehead%2C%20A.%2C%20Cheng%2C%20J.%20%26amp%3B%20Zhang%2C%20M.%20MegaFold%3A%20Efficient%20Training%20of%20Next-Generation%203D%20Attention%20Protein%20Models%20on%20Cross-Platform%20GPUs.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2506.20686%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2506.20686%26lt%3B%5C%2Fa%26gt%3B%20%282025%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22MegaFold%3A%20Efficient%20Training%20of%20Next-Generation%203D%20Attention%20Protein%20Models%20on%20Cross-Platform%20GPUs%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Hoa%22%2C%22lastName%22%3A%22La%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ahan%22%2C%22lastName%22%3A%22Gupta%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Alex%22%2C%22lastName%22%3A%22Morehead%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jianlin%22%2C%22lastName%22%3A%22Cheng%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Minjia%22%2C%22lastName%22%3A%22Zhang%22%7D%5D%2C%22abstractNote%22%3A%22Recent%20advances%20in%20biomolecular%20modeling%20have%20been%20catalyzed%20by%20models%20such%20as%20AlphaFold3%20%28AF3%29%2C%20which%20introduce%20science-informed%20changes%20to%20the%20transformer%20architecture.%20Unlike%20transformers%2C%20a%20defining%20characteristic%20of%20AF3-style%20models%20is%20their%203D%20attention%20over%202D%20pairwise%20representations%20which%20produces%20tensors%20whose%20computation%20and%20memory%20costs%20scale%20cubically%20with%20sequence%20length.%20As%20a%20result%2C%20despite%20moderate%20parameter%20counts%2C%20AF3-style%20models%20are%20far%20more%20expensive%20to%20train%20than%20size-equivalent%20transformers%2C%20and%20are%20severely%20constrained%20by%20GPU%20memory%20capacity.%20Our%20characterization%20shows%203D%20attention%20fundamentally%20changes%20the%20training%20workload%2C%20causing%20massive%203D%20attention%20maps%2C%20complex%20inter-operator%20dependencies%2C%20kernel%20fragmentation%2C%20and%20heavy%20host-side%20data%20pipelines%20which%20differ%20substantially%20from%20LLM%20training%2C%20leading%20to%20poor%20utilization%20on%20modern%20GPU%20systems.%20Moreover%2C%20existing%20GPU%20optimizations%20do%20not%20adequately%20address%20these%20challenges%20due%20to%20complex%20cross-layer%20inter-operator%20dependencies%20introduced%20by%203D%20attention.%20Motivated%20by%20these%20challenges%2C%20we%20introduce%20MegaFold%2C%20a%20novel%20cross-platform%20system%20for%20efficient%20training%20of%20next-generation%203D-attention%20protein%20models.%20MegaFold%20combines%20a%20memory-efficient%203D-attention%20kernel%2C%20a%20communication-efficient%20sharding%20strategy%20for%20quadratic%20representations%2C%20fused%20operator%20implementations%20for%20critical%20execution%20paths%2C%20and%20a%20determinism-aware%20host-device%20pipeline%20that%20eliminates%20preprocessing%20stalls.%20Evaluation%20on%20both%20NVIDIA%20H200%20and%20AMD%20MI250%20GPUs%20shows%20that%20MegaFold%20enables%20training%20with%20up%20to%203.36%24%5C%5Ctimes%24%20longer%20sequence%20lengths%20on%2032%20GPUs%20while%20reducing%20end-to-end%20execution%20time%20by%20up%20to%201.73%24%5C%5Ctimes%24%20%28NVIDIA%29%20and%201.62%24%5C%5Ctimes%24%20%28AMD%29.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222025%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2506.20686%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2506.20686%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A28%3A32Z%22%7D%7D%2C%7B%22key%22%3A%22J5GZ6KLM%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Lancaster%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BLancaster%2C%20L.%2C%20Fielding%2C%20D.%20B.%2C%20Mohapatra%2C%20R.%20%26amp%3B%20Bryan%2C%20G.%20L.%20The%20Origin%20of%20Da%20Scaling%3A%20Suppressed%20Cooling%20in%20Fast-Cooling%20Mixing%20Layers.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.04093%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2606.04093%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22The%20Origin%20of%20Da%20Scaling%3A%20Suppressed%20Cooling%20in%20Fast-Cooling%20Mixing%20Layers%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Lachlan%22%2C%22lastName%22%3A%22Lancaster%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Drummond%20B.%22%2C%22lastName%22%3A%22Fielding%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Rajsekhar%22%2C%22lastName%22%3A%22Mohapatra%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Greg%20L.%22%2C%22lastName%22%3A%22Bryan%22%7D%5D%2C%22abstractNote%22%3A%22In%20numerical%20experiments%20simulating%20Turbulent%20Radiative%20Mixing%20Layers%20%28TRMLs%29%20it%20is%20observed%20that%20as%20the%20cooling%20time%20in%20the%20mixed%20gas%2C%20%24t_%7B%5C%5Crm%20cool%7D%24%2C%20becomes%20very%20short%20compared%20to%20the%20dynamical%20time%20of%20the%20turbulence%2C%20%24t_%7B%5C%5Crm%20eddy%7D%5C%2Ft_%7B%5C%5Crm%20cool%7D%20%5C%5Cgg%201%24%2C%20there%20is%20a%20change%20in%20the%20scaling%20behavior%20of%20the%20total%20energy%20radiated%20in%20the%20TRML%20as%20a%20function%20of%20this%20ratio%2C%20also%20known%20as%20the%20Damk%5Cu00f6hler%20number%2C%20%24%7B%5C%5Crm%20Da%7D%20%5C%5Cequiv%20t_%7B%5C%5Crm%20eddy%7D%5C%2Ft_%7B%5C%5Crm%20cool%7D%24%2C%20from%20%24%5C%5Cdot%7BE%7D_%7B%5C%5Crm%20cool%7D%20%5C%5Cpropto%20%7B%5C%5Crm%20Da%7D%5E%7B1%5C%2F2%7D%24%20to%20%24%5C%5Cdot%7BE%7D_%7B%5C%5Crm%20cool%7D%20%5C%5Cpropto%20%7B%5C%5Crm%20Da%7D%5E%7B1%5C%2F4%7D%24.%20The%20latter%2C%20so-called%20%26quot%3Bfast-cooling%2C%26quot%3B%20regime%20is%20of%20particular%20interest%20as%20many%20astrophysical%20mixing%20layers%20lie%20in%20this%20regime.%20We%20demonstrate%20that%20the%20origin%20of%20this%20change%20is%20the%20suppression%20of%20turbulent%20folding%20of%20the%20surface%20by%20the%20ram-pressure%20of%20the%20inflowing%20gas%2C%20which%20becomes%20much%20greater%20than%20the%20turbulent%20pressure%20in%20this%20regime.%20We%20present%20an%20argument%20that%20reproduces%20the%20%24%5C%5Cdot%7BE%7D_%7B%5C%5Crm%20cool%7D%20%5C%5Cpropto%20%7B%5C%5Crm%20Da%7D%5E%7B1%5C%2F4%7D%24%20behavior%20by%20appealing%20to%20the%20suppression%20of%20the%20fractal%20structure%20of%20the%20interface%20by%20the%20ram-pressure%20of%20the%20inflowing%20gas.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2606.04093%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2606.04093%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A22%3A54Z%22%7D%7D%2C%7B%22key%22%3A%22UR3WKBSI%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Anand%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BAnand%2C%20A.%2C%20Singh%2C%20J.%2C%20Sun%2C%20Z.%2C%20Hakkani-T%26%23xFC%3Br%2C%20D.%20%26amp%3B%20Feamster%2C%20N.%20Measuring%2C%20Localizing%2C%20and%20Ablating%20Alignment%20Signatures%20in%20LLMs.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2605.30526%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2605.30526%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Measuring%2C%20Localizing%2C%20and%20Ablating%20Alignment%20Signatures%20in%20LLMs%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Aniket%22%2C%22lastName%22%3A%22Anand%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Janvijay%22%2C%22lastName%22%3A%22Singh%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zhewei%22%2C%22lastName%22%3A%22Sun%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Dilek%22%2C%22lastName%22%3A%22Hakkani-T%5Cu00fcr%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Nick%22%2C%22lastName%22%3A%22Feamster%22%7D%5D%2C%22abstractNote%22%3A%22Aligned%20language%20models%20often%20exhibit%20a%20recognizable%20AI-like%20style%2C%20yet%20its%20connection%20to%20post-training%20and%20internal%20representations%20remains%20poorly%20understood.%20In%20this%20work%2C%20we%20study%20whether%20post-training%20introduces%20or%20amplifies%20AI-like%20stylistic%20regularities%20and%20whether%20these%20regularities%20have%20a%20localized%20internal%20signature.%20To%20this%20end%2C%20we%20compare%20human%20text%2C%20base-model%20generations%2C%20and%20aligned-model%20generations%20under%20matched%20human-source%20prefixes.%20Aligned%20generations%20show%20lower%20human-corpus%20affinity%20and%20higher%20AI-detection%20rates%20than%20base%20generations%2C%20suggesting%20that%20post-training%20shifts%20generated%20text%20away%20from%20human-corpus%20style%20and%20toward%20detector-visible%20AI-like%20text.%20We%20then%20introduce%20PASTA%20%28Post-training%20Alignment%20Signature%20Targeted%20Ablation%29%2C%20a%20training-free%20method%20that%20estimates%20a%20post-training%20alignment%20signature%20from%20aligned-base%20residual%20contrasts%20and%20ablates%20the%20corresponding%20direction%20during%20decoding.%20Across%2011%20aligned%20models%20and%206%20AI%20detectors%2C%20PASTA%20lowers%20the%20detection%20rate%20for%20most%20aligned%20models%3B%20this%20effect%20transfers%20well%20across%20detectors%20and%20is%20not%20reproduced%20by%20random%20directions.%20Qualitative%20analysis%20suggests%20that%20PASTA%20generations%20remain%20relevant%20and%20coherent%20while%20exhibiting%20greater%20stylistic%20variation.%20Together%2C%20these%20results%20show%20that%20AI-like%20stylistic%20effects%20of%20post-training%20can%20be%20measured%2C%20localized%2C%20and%20causally%20tested%20through%20activation%20ablation.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2605.30526%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2605.30526%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A21%3A06Z%22%7D%7D%2C%7B%22key%22%3A%22HX9HDQFS%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Bai%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BBai%2C%20H.%2C%20Taymanov%2C%20A.%2C%20Zhang%2C%20T.%2C%20Kumar%2C%20A.%20%26amp%3B%20Whitehead%2C%20S.%20WebGym%3A%20Scaling%20Training%20Environments%20for%20Visual%20Web%20Agents%20with%20Realistic%20Tasks.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2601.02439%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2601.02439%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22WebGym%3A%20Scaling%20Training%20Environments%20for%20Visual%20Web%20Agents%20with%20Realistic%20Tasks%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Hao%22%2C%22lastName%22%3A%22Bai%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Alexey%22%2C%22lastName%22%3A%22Taymanov%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Tong%22%2C%22lastName%22%3A%22Zhang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Aviral%22%2C%22lastName%22%3A%22Kumar%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Spencer%22%2C%22lastName%22%3A%22Whitehead%22%7D%5D%2C%22abstractNote%22%3A%22We%20present%20WebGym%2C%20the%20largest-to-date%20open-source%20environment%20for%20training%20realistic%20visual%20web%20agents.%20Real%20websites%20are%20non-stationary%20and%20diverse%2C%20making%20artificial%20or%20small-scale%20task%20sets%20insufficient%20for%20robust%20policy%20learning.%20WebGym%20contains%20nearly%20300%2C000%20tasks%20with%20rubric-based%20evaluations%20across%20diverse%2C%20real-world%20websites%20and%20difficulty%20levels.%20We%20train%20agents%20with%20a%20simple%20reinforcement%20learning%20%28RL%29%20recipe%2C%20which%20trains%20on%20the%20agent%26%23039%3Bs%20own%20interaction%20traces%20%28rollouts%29%2C%20using%20task%20rewards%20as%20feedback%20to%20guide%20learning.%20To%20enable%20scaling%20RL%2C%20we%20speed%20up%20sampling%20of%20trajectories%20in%20WebGym%20by%20developing%20a%20high-throughput%20asynchronous%20rollout%20system%2C%20designed%20specifically%20for%20web%20agents.%20Our%20system%20achieves%20a%204-5x%20rollout%20speedup%20compared%20to%20naive%20implementations.%20Second%2C%20we%20scale%20the%20task%20set%20breadth%2C%20depth%2C%20and%20size%2C%20which%20results%20in%20continued%20performance%20improvement.%20Fine-tuning%20a%20strong%20base%20vision-language%20model%2C%20Qwen-3-VL-8B-Instruct%2C%20on%20WebGym%20results%20in%20an%20improvement%20in%20success%20rate%20on%20an%20out-of-distribution%20test%20set%20from%2026.2%25%20to%2042.9%25%2C%20significantly%20outperforming%20agents%20based%20on%20proprietary%20models%20such%20as%20GPT-4o%20and%20GPT-5-Thinking%20that%20achieve%2027.1%25%20and%2029.8%25%2C%20respectively.%20This%20improvement%20is%20substantial%20because%20our%20test%20set%20consists%20only%20of%20tasks%20on%20websites%20never%20seen%20during%20training%2C%20unlike%20many%20other%20prior%20works%20on%20training%20visual%20web%20agents.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2601.02439%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2601.02439%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A18%3A14Z%22%7D%7D%2C%7B%22key%22%3A%22JFGQVU8R%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Chen%20et%20al.%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BChen%2C%20Z.%20J.%20%26lt%3Bi%26gt%3Bet%20al.%26lt%3B%5C%2Fi%26gt%3B%20Everything%20at%20Every%20Scale%3A%20Scale-Invariant%20Diffusion%20with%20Continuous%20Super-Resolution.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2605.26032%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2605.26032%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Everything%20at%20Every%20Scale%3A%20Scale-Invariant%20Diffusion%20with%20Continuous%20Super-Resolution%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zixin%20Jessie%22%2C%22lastName%22%3A%22Chen%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zhuo%22%2C%22lastName%22%3A%22Chen%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Archer%22%2C%22lastName%22%3A%22Wang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jeff%22%2C%22lastName%22%3A%22Gore%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22William%20T.%22%2C%22lastName%22%3A%22Freeman%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Congyue%22%2C%22lastName%22%3A%22Deng%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Marin%22%2C%22lastName%22%3A%22Solja%5Cu010di%5Cu0107%22%7D%5D%2C%22abstractNote%22%3A%22Creating%20images%20from%20noise%20is%20image%20generation%3B%20reconstructing%20fine%20details%20from%20coarse%20inputs%20is%20super-resolution.%20Despite%20their%20practical%20differences%2C%20both%20can%20be%20understood%20as%20reversing%20information%20loss%20across%20scales.%20We%20introduce%20%24%5C%5Ctextbf%7BSKILD%7D%24%2C%20a%20%24%5C%5Ctextbf%7BS%7D%24cale-invariant%20%24%5C%5Ctextbf%7BK%7D%24-Space%20%24%5C%5Ctextbf%7BI%7D%24mage%20%24%5C%5Ctextbf%7BL%7D%24earning%20%24%5C%5Ctextbf%7BD%7D%24iffusion%20model%20that%20unifies%20generation%20and%20continuous%20super-resolution%20within%20a%20single%20unconditional%20framework.%20Both%20natural%20images%20and%20critical%20physical%20systems%20exhibit%20scale%20invariance%2C%20and%20we%20leverage%20it%20to%20design%20a%20forward%20process%20that%20attenuates%20image%20content%20from%20fine%20to%20coarse%20scales%20while%20injecting%20spectrum-matched%20Gaussian%20noise%2C%20making%20scale%20an%20explicit%20coordinate%20of%20the%20diffusion%20dynamics.%20The%20same%20trained%20reverse%20process%20performs%20generation%20and%20continuous%20super-resolution%20by%20varying%20only%20the%20starting%20timestep%3A%20%24%5C%5Ctextit%7Bno%20task-specific%20architecture%2C%20no%20conditioning%20branch%2C%20no%20classifier-free%20guidance%2C%20no%20retraining%20per%20scale%20factor%7D%24.%20Empirically%2C%20SKILD%20reaches%20FID%20%242.65%24%20and%20Inception%20Score%20%249.63%24%20on%20unconditional%20CIFAR-10%2C%20performs%20%242%5C%5Ctimes%24--%248%5C%5Ctimes%24%20super-resolution%20on%20ImageNet%20from%20a%20single%20unconditional%20checkpoint%20while%20outperforming%20conditional%20models%20across%20perceptual%20metrics%2C%20and%20reconstructs%20critical%20Ising%20models%20whose%20connected%20four-point%20correlations%20closely%20track%20the%20ground%20truth.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2605.26032%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2605.26032%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A15%3A23Z%22%7D%7D%2C%7B%22key%22%3A%2275RCPCX2%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Marreddy%20and%20Liu%22%2C%22parsedDate%22%3A%222026%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%26lt%3Bdiv%20class%3D%26quot%3Bcsl-bib-body%26quot%3B%20style%3D%26quot%3Bline-height%3A%202%3B%20%26quot%3B%26gt%3B%5Cn%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-entry%26quot%3B%20style%3D%26quot%3Bclear%3A%20left%3B%20%26quot%3B%26gt%3B%5Cn%20%20%20%20%26lt%3Bdiv%20class%3D%26quot%3Bcsl-left-margin%26quot%3B%20style%3D%26quot%3Bfloat%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%26quot%3B%26gt%3B1.%26lt%3B%5C%2Fdiv%26gt%3B%26lt%3Bdiv%20class%3D%26quot%3Bcsl-right-inline%26quot%3B%20style%3D%26quot%3Bmargin%3A%200%20.4em%200%201.5em%3B%26quot%3B%26gt%3BMarreddy%2C%20R.%20%26amp%3B%20Liu%2C%20C.%20Label-NTK%20Alignments%20and%20A%20Tighter%20Convergence%20Bound%20in%20the%20NTK%20Regime.%20Preprint%20at%20%26lt%3Ba%20class%3D%26%23039%3Bzp-DOIURL%26%23039%3B%20href%3D%26%23039%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2605.25275%26%23039%3B%26gt%3Bhttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2605.25275%26lt%3B%5C%2Fa%26gt%3B%20%282026%29.%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%20%20%26lt%3B%5C%2Fdiv%26gt%3B%5Cn%26lt%3B%5C%2Fdiv%26gt%3B%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Label-NTK%20Alignments%20and%20A%20Tighter%20Convergence%20Bound%20in%20the%20NTK%20Regime%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ruchirinkil%22%2C%22lastName%22%3A%22Marreddy%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Chaoyue%22%2C%22lastName%22%3A%22Liu%22%7D%5D%2C%22abstractNote%22%3A%22The%20Neural%20Tangent%20Kernel%20%28NTK%29%20framework%20explains%20optimization%20in%20over-parameterized%20neural%20networks%20via%20approximately%20linearized%20dynamics%2C%20yielding%20exponential%20convergence%20guarantees.%20However%2C%20existing%20results%20are%20often%20overly%20pessimistic%20and%20do%20not%20match%20the%20fast%20training%20in%20practice%2C%20as%20they%20depend%20on%20the%20smallest%20NTK%20eigenvalue%2C%20which%20is%20typically%20extremely%20small%20in%20practice.%20In%20this%20work%2C%20we%20develop%20sharper%20convergence%20guarantees%20by%20characterizing%20the%20interaction%20between%20data%20labels%20and%20the%20NTK%20eigen-spectrum.%20We%20identify%20two%20key%20phenomena%2C%20Label-NTK%20alignment%20and%20Residual-NTK%20alignment%2C%20showing%20that%20projections%20of%20labels%20and%20residuals%20onto%20NTK%20eigenvectors%20scale%20with%20the%20corresponding%20eigenvalues.%20We%20provide%20empirical%20evidence%20and%20theoretical%20justification%20under%20mild%20data%20assumptions.%20Exploiting%20these%20alignment%20properties%2C%20we%20derive%20a%20refined%20convergence%20bound%20that%20depends%20on%20the%20full%20spectrum%20and%20closely%20matches%20practical%20training%20dynamics%2C%20significantly%20improving%20over%20classical%20worst-case%20results.%20We%20further%20obtain%20improved%20generalization%20bounds.%20Experiments%20on%20MLPs%20and%20CNNs%20across%20multiple%20datasets%20validate%20our%20theory.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222026%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2605.25275%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2605.25275%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222026-08-23T12%3A09%3A25Z%22%7D%7D%5D%7D
1.
Mohapatra, R., Lancaster, L., Fielding, D., Quataert, E. & Bryan, G. L. How Magnetic Fields Regulate Cooling and Mixing in Turbulent Radiative Mixing Layers. Preprint at https://doi.org/10.48550/ARXIV.2609.29782 (2026).
1.
Gong, J. et al. SAM-V: Geometry-Aware Segment Anything for Multi-View Instance Segmentation. Preprint at https://doi.org/10.48550/arXiv.2609.25490 (2026).
1.
Hategan-Marandiuc, M. et al. Toward Autonomous Radio Follow-up of Multi-messenger Transients with RADAR: From Alert Parsing to Inference and Observation Scheduling. Preprint at https://doi.org/10.48550/arXiv.2609.18233 (2026).
1.
Trachu, T., Cornell, S., Chen, W. & Watanabe, S. LACE: Layer-Wise Compression for Dynamic Frame Rate Codecs. Preprint at https://doi.org/10.48550/ARXIV.2609.17509 (2026).
1.
Zhu, Z., Gong, Y., Xiao, Y., Liu, Y. & Hoiem, D. How to Teach Large Multimodal Models New Skills. Preprint at https://doi.org/10.48550/ARXIV.2510.08564 (2025).
1.
Tang, B., Spalding, Z. & Cogan, G. B. Pretraining for Sample-Efficient Neural Interfaces. Preprint at https://doi.org/10.48550/arXiv.2609.13507 (2026).
1.
Kwark, D. H. et al. Seamless Whole Slide Label-Free Virtual Staining. Preprint at https://doi.org/10.48550/arXiv.2609.10914 (2026).
1.
Koric, A. A., Liu, Q. & Koric, S. Disentangling Attention in Deep Operator Learning: A Controlled Study of Data-Driven and Physics-Informed Architectures. Preprint at https://doi.org/10.48550/ARXIV.2609.04407 (2026).
1.
Mahajan, J. et al. ReaDiT Guidance: Control for Image and Video Generation using Diffusion Transformer Features. Preprint at https://doi.org/10.48550/ARXIV.2609.04649 (2026).
1.
Ramasubramanian, S. et al. Tail-Likelihood Reinforcement Learning. Preprint at https://doi.org/10.48550/ARXIV.2609.02987 (2026).
1.
Li, A., Li, Y., DeHaas, D. & Guidi, G. Cheaper by the Batch: Shared Traversal for Genotype Graph Editing. Preprint at https://doi.org/10.48550/arXiv.2608.26488 (2026).
1.
Yu, C. & Ji, H. Sememe: Causal Modeling for Task-Aware Depth Pruning of Large Language Models. in COLM 2026 (San Francisco, CA, 2026).
1.
Howes, W., Ahmed, F. & Alam, S. B. Low-Latency Activation-Regularized Sparse Neural Operators with Distillation Assistance Towards Real-Time Edge-Deployable Virtual Sensing. Preprint at https://doi.org/10.48550/ARXIV.2608.23987 (2026).
1.
Someki, M. et al. PlanRAG-Audio: Planning and Retrieval Augmented Generation for Long-form Audio Understanding. Preprint at https://doi.org/10.48550/ARXIV.2605.20414 (2026).
1.
Fan, X., Liu, M., Yang, R., Ouyang, S. & Han, J. Can Agent Memory Systems Track Evolving State? Preprint at https://doi.org/10.48550/arXiv.2608.19652 (2026).
1.
Feng, Y., Bijoy, B. S., Balasubramanian, N. & Zhou, J. Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents. Preprint at https://doi.org/10.48550/arXiv.2608.20274 (2026).
1.
Chen, N., Yang, Z. & Hayou, S. Training-Free versus Training-Based Intent Classification in LLMs: Accuracy, Robustness, and Failure Modes. Preprint at https://doi.org/10.48550/ARXIV.2608.02415 (2026).
1.
Rodríguez, C. C. Tails and Walls: Reality’s Two Honest Priors. Preprint at https://doi.org/10.21203/rs.3.rs-10473971/v1 (2026).
1.
Berres, A. et al. DB-Bench: Benchmarking Deblenders for LSST DESC Using the Blending ToolKit. Preprint at https://doi.org/10.48550/ARXIV.2607.28475 (2026).
1.
Radhakrishnan, A. & Bryngelson, S. H. A sub-grid-scale model for polydisperse bubbly flows with heat and mass transfer. Preprint at https://doi.org/10.48550/ARXIV.2607.27426 (2026).
1.
Zhou, S. et al. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability. Preprint at https://doi.org/10.48550/ARXIV.2607.26637 (2026).
1.
Hidalgo, L. & Draper, P. Lattice Quantum Chromodynamics for Quantum Simulations. Preprint at https://doi.org/10.48550/ARXIV.2607.26445 (2026).
1.
Gladstone, A., Ji, H. & Du, Y. Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation. Preprint at https://doi.org/10.48550/ARXIV.2607.27372 (2026).
1.
Wu, S., Qian, C., Chen, X. & Ji, H. Teaching LLMs to Self-Evolve: Cultivating Core Meta-Skills with Reinforcement Learning. Preprint at https://doi.org/10.48550/ARXIV.2607.21971 (2026).
1.
Lin, S. et al. D$_4$CNN$\times$AnaCal: Physics-Informed Machine Learning for Accurate and Precise Weak Lensing Shear Estimation. Preprint at https://doi.org/10.48550/ARXIV.2603.19046 (2026).
1.
Yang, Y., Zhao, Y., Dasgupta, A. & Misailovic, S. PagedWeight: Efficient MoE LLM Serving with Dynamic Quality-Aware Weight Quantization. Preprint at https://doi.org/10.48550/ARXIV.2607.16184 (2026).
1.
Park, J. et al. Network architecture follows coupling in multiphysics systems: single vs. multiple branches in DeepONet and S-DeepONet. Commun Eng https://doi.org/10.1038/s44172-026-00714-4 (2026) doi:10.1038/s44172-026-00714-4.
1.
Wang, H., Tian, J., Arora, S. & Watanabe, S. An Efficient vLLM-Based Inference Pipeline for Unified Audio Understanding and Generation. Preprint at https://doi.org/10.48550/ARXIV.2607.02119 (2026).
1.
Remy, B., Chang, C. & Willett, R. Joint inference of weak lensing convergence map and cosmology with diffusion models. Preprint at https://doi.org/10.48550/ARXIV.2606.31988 (2026).
1.
Kim, T. & Kumar, A. Addressing Over-Refusal in LLMs with Competing Rewards. Preprint at https://doi.org/10.48550/ARXIV.2606.31748 (2026).
1.
Chen, Z., Mayne i Comas, O., Jin, Z., Luo, D. & Soljacic, M. L-CUBE: Isolating Long-Context Capacity from Knowledge with Controllable Mutual Information Scaling. in Proceedings of the 43rd International Conference on Machine Learning (ICML) (2026).
1.
Zhao, D., Cao, S., Wang, Y.-X. & Gui, L.-Y. Think in Latent, Explain in Language: Self-Explainable Latent Reasoning. Preprint at https://doi.org/10.48550/arXiv.2608.13570 (2026).
1.
Lian, X. et al. Quantization Inflates Reasoning: Token Inflation as a Hidden Cost of Low-Bit Reasoning Models. Preprint at https://doi.org/10.48550/ARXIV.2606.25519 (2026).
1.
Tian, J. et al. Bagpiper-TTS: Natural Language Guided Universal Speech Synthesis. Preprint at https://doi.org/10.48550/ARXIV.2606.22811 (2026).
1.
Tran, B., Maithert, M., Shin, W., Sinclair, M. D. & Venkataraman, S. ET: Bridging the Gap on Energy Telemetry for Multi-GPU Communication Collectives. in Proceedings of the 16th International Workshop on Accelerators and Hybrid Emerging Systems (AsHES) (2026).
1.
Liang, W. et al. MedQPA-Gen: Medical Question Proposing and Answering for Report Generation. in Findings of the Association for Computational Linguistics: ACL 2026 43120–43132 (Association for Computational Linguistics, San Diego, California, United States, 2026). http://doi.org/10.18653/v1/2026.findings-acl.2139.
1.
Wang, Z., Luo, X., Zeng, S. & Yan, X. Dual Dimensionality for Local and Global Attention. Preprint at https://doi.org/10.48550/ARXIV.2606.18587 (2026).
1.
Daithankar, N., Gladstone, A., LeCun, Y. & Ji, H. You Don’t Need Strong Assumptions: Visual Representation Learning via Temporal Differences. Preprint at https://doi.org/10.48550/ARXIV.2606.15956 (2026).
1.
Yoo, J. P. et al. Towards Data-Efficient Cross-Device Generalization of Grad-Shafranov Equilibria via Transfer Learning Neural Operator. Preprint at https://doi.org/10.48550/ARXIV.2606.15512 (2026).
1.
Tam, W. K. The Amplifying Mirror: Locating and Steering the Partisan Direction inside a Large Language Model. Preprint at https://doi.org/10.48550/arXiv.2606.08792 (2026).
1.
Jiang, E., Gjølbye, A., Zhang, Y. J. & Koyejo, S. When Behavioral Safety Evaluation Fails: A Representation-Level Perspective. Preprint at https://doi.org/10.48550/ARXIV.2606.08044 (2026).
1.
Wang, Y., Geng, R., Chen, Y. & Jia, J. AttnTrace: Contextual Attribution of Prompt Injection and Knowledge Corruption. Preprint at https://doi.org/10.48550/ARXIV.2508.03793 (2025).
1.
Caplan, M. E., Smith, N. T., Bransgrove, A. J. & Horowitz, C. J. Plasticity of Neutron Star Crusts. Preprint at https://doi.org/10.48550/ARXIV.2606.06706 (2026).
1.
Bai, H. et al. AsyncWebRL: Efficient Asynchronous Reinforcement Learning for Multi-Step Visual Web Agents. Preprint at https://doi.org/10.48550/ARXIV.2606.05597 (2026).
1.
La, H., Gupta, A., Morehead, A., Cheng, J. & Zhang, M. MegaFold: Efficient Training of Next-Generation 3D Attention Protein Models on Cross-Platform GPUs. Preprint at https://doi.org/10.48550/ARXIV.2506.20686 (2025).
1.
Lancaster, L., Fielding, D. B., Mohapatra, R. & Bryan, G. L. The Origin of Da Scaling: Suppressed Cooling in Fast-Cooling Mixing Layers. Preprint at https://doi.org/10.48550/ARXIV.2606.04093 (2026).
1.
Anand, A., Singh, J., Sun, Z., Hakkani-Tür, D. & Feamster, N. Measuring, Localizing, and Ablating Alignment Signatures in LLMs. Preprint at https://doi.org/10.48550/ARXIV.2605.30526 (2026).
1.
Bai, H., Taymanov, A., Zhang, T., Kumar, A. & Whitehead, S. WebGym: Scaling Training Environments for Visual Web Agents with Realistic Tasks. Preprint at https://doi.org/10.48550/ARXIV.2601.02439 (2026).
1.
Chen, Z. J. et al. Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution. Preprint at https://doi.org/10.48550/ARXIV.2605.26032 (2026).
1.
Marreddy, R. & Liu, C. Label-NTK Alignments and A Tighter Convergence Bound in the NTK Regime. Preprint at https://doi.org/10.48550/ARXIV.2605.25275 (2026).