Delta AI Publications
Publications resulting from research conducted using DeltaAI appear here. Check back to see how the list of exciting discoveries made using DeltaAI 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
1
3707
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%3A0%2C%22request_next%22%3A0%2C%22used_cache%22%3Atrue%7D%2C%22data%22%3A%5B%7B%22key%22%3A%22NCVA3VIY%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Butler%20et%20al.%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3EButler%2C%20L.%20%3Ci%3Eet%20al.%3C%5C%2Fi%3E%20ProxySPEX%3A%20Inference-Efficient%20Interpretability%20via%20Sparse%20Feature%20Interactions%20in%20LLMs.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2505.17495%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2505.17495%3C%5C%2Fa%3E%20%282025%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22ProxySPEX%3A%20Inference-Efficient%20Interpretability%20via%20Sparse%20Feature%20Interactions%20in%20LLMs%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Landon%22%2C%22lastName%22%3A%22Butler%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Abhineet%22%2C%22lastName%22%3A%22Agarwal%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Justin%20Singh%22%2C%22lastName%22%3A%22Kang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yigit%20Efe%22%2C%22lastName%22%3A%22Erginbas%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Bin%22%2C%22lastName%22%3A%22Yu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Kannan%22%2C%22lastName%22%3A%22Ramchandran%22%7D%5D%2C%22abstractNote%22%3A%22Large%20Language%20Models%20%28LLMs%29%20have%20achieved%20remarkable%20performance%20by%20capturing%20complex%20interactions%20between%20input%20features.%20To%20identify%20these%20interactions%2C%20most%20existing%20approaches%20require%20enumerating%20all%20possible%20combinations%20of%20features%20up%20to%20a%20given%20order%2C%20causing%20them%20to%20scale%20poorly%20with%20the%20number%20of%20inputs%20%24n%24.%20Recently%2C%20Kang%20et%20al.%20%282025%29%20proposed%20SPEX%2C%20an%20information-theoretic%20approach%20that%20uses%20interaction%20sparsity%20to%20scale%20to%20%24n%20%5C%5Capprox%2010%5E3%24%20features.%20SPEX%20greatly%20improves%20upon%20prior%20methods%20but%20requires%20tens%20of%20thousands%20of%20model%20inferences%2C%20which%20can%20be%20prohibitive%20for%20large%20models.%20In%20this%20paper%2C%20we%20observe%20that%20LLM%20feature%20interactions%20are%20often%20hierarchical%20--%20higher-order%20interactions%20are%20accompanied%20by%20their%20lower-order%20subsets%20--%20which%20enables%20more%20efficient%20discovery.%20To%20exploit%20this%20hierarchy%2C%20we%20propose%20ProxySPEX%2C%20an%20interaction%20attribution%20algorithm%20that%20first%20fits%20gradient%20boosted%20trees%20to%20masked%20LLM%20outputs%20and%20then%20extracts%20the%20important%20interactions.%20Experiments%20across%20four%20challenging%20high-dimensional%20datasets%20show%20that%20ProxySPEX%20more%20faithfully%20reconstructs%20LLM%20outputs%20by%2020%25%20over%20marginal%20attribution%20approaches%20while%20using%20%2410%5C%5Ctimes%24%20fewer%20inferences%20than%20SPEX.%20By%20accounting%20for%20interactions%2C%20ProxySPEX%20identifies%20features%20that%20influence%20model%20output%20over%2020%25%20more%20than%20those%20selected%20by%20marginal%20approaches.%20Further%2C%20we%20apply%20ProxySPEX%20to%20two%20interpretability%20tasks.%20Data%20attribution%2C%20where%20we%20identify%20interactions%20among%20CIFAR-10%20training%20samples%20that%20influence%20test%20predictions%2C%20and%20mechanistic%20interpretability%2C%20where%20we%20uncover%20interactions%20between%20attention%20heads%2C%20both%20within%20and%20across%20layers%2C%20on%20a%20question-answering%20task.%20ProxySPEX%20identifies%20interactions%20that%20enable%20more%20aggressive%20pruning%20of%20heads%20than%20marginal%20approaches.%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.2505.17495%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2505.17495%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-05-29T13%3A56%3A00Z%22%7D%7D%2C%7B%22key%22%3A%22KYYX6UI2%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Agarwal%20et%20al.%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3EAgarwal%2C%20S.%2C%20Zhang%2C%20Z.%2C%20Yuan%2C%20L.%2C%20Han%2C%20J.%20%26amp%3B%20Peng%2C%20H.%20The%20Unreasonable%20Effectiveness%20of%20Entropy%20Minimization%20in%20LLM%20Reasoning.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2505.15134%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2505.15134%3C%5C%2Fa%3E%20%282025%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22The%20Unreasonable%20Effectiveness%20of%20Entropy%20Minimization%20in%20LLM%20Reasoning%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shivam%22%2C%22lastName%22%3A%22Agarwal%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zimin%22%2C%22lastName%22%3A%22Zhang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Lifan%22%2C%22lastName%22%3A%22Yuan%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jiawei%22%2C%22lastName%22%3A%22Han%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Hao%22%2C%22lastName%22%3A%22Peng%22%7D%5D%2C%22abstractNote%22%3A%22Entropy%20minimization%20%28EM%29%20trains%20the%20model%20to%20concentrate%20even%20more%20probability%20mass%20on%20its%20most%20confident%20outputs.%20We%20show%20that%20this%20simple%20objective%20alone%2C%20without%20any%20labeled%20data%2C%20can%20substantially%20improve%20large%20language%20models%27%20%28LLMs%29%20performance%20on%20challenging%20math%2C%20physics%2C%20and%20coding%20tasks.%20We%20explore%20three%20approaches%3A%20%281%29%20EM-FT%20minimizes%20token-level%20entropy%20similarly%20to%20instruction%20finetuning%2C%20but%20on%20unlabeled%20outputs%20drawn%20from%20the%20model%3B%20%282%29%20EM-RL%3A%20reinforcement%20learning%20with%20negative%20entropy%20as%20the%20only%20reward%20to%20maximize%3B%20%283%29%20EM-INF%3A%20inference-time%20logit%20adjustment%20to%20reduce%20entropy%20without%20any%20training%20data%20or%20parameter%20updates.%20On%20Qwen-7B%2C%20EM-RL%2C%20without%20any%20labeled%20data%2C%20achieves%20comparable%20or%20better%20performance%20than%20strong%20RL%20baselines%20such%20as%20GRPO%20and%20RLOO%20that%20are%20trained%20on%2060K%20labeled%20examples.%20Furthermore%2C%20EM-INF%20enables%20Qwen-32B%20to%20match%20or%20exceed%20the%20performance%20of%20proprietary%20models%20like%20GPT-4o%2C%20Claude%203%20Opus%2C%20and%20Gemini%201.5%20Pro%20on%20the%20challenging%20SciCode%20benchmark%2C%20while%20being%203x%20more%20efficient%20than%20self-consistency%20and%20sequential%20refinement.%20Our%20findings%20reveal%20that%20many%20pretrained%20LLMs%20possess%20previously%20underappreciated%20reasoning%20capabilities%20that%20can%20be%20effectively%20elicited%20through%20entropy%20minimization%20alone%2C%20without%20any%20labeled%20data%20or%20even%20any%20parameter%20updates.%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.2505.15134%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2505.15134%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-05-27T19%3A10%3A03Z%22%7D%7D%2C%7B%22key%22%3A%22WPI9P3BW%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Liu%20et%20al.%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3ELiu%2C%20Q.%20%3Ci%3Eet%20al.%3C%5C%2Fi%3E%20Geometry-Informed%20Neural%20Operator%20Transformer.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2504.19452%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2504.19452%3C%5C%2Fa%3E%20%282025%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Geometry-Informed%20Neural%20Operator%20Transformer%22%2C%22creators%22%3A%5B%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%22Vincient%22%2C%22lastName%22%3A%22Zhong%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Hadi%22%2C%22lastName%22%3A%22Meidani%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%22Philippe%22%2C%22lastName%22%3A%22Geubelle%22%7D%5D%2C%22abstractNote%22%3A%22Machine-learning-based%20surrogate%20models%20offer%20significant%20computational%20efficiency%20and%20faster%20simulations%20compared%20to%20traditional%20numerical%20methods%2C%20especially%20for%20problems%20requiring%20repeated%20evaluations%20of%20partial%20differential%20equations.%20This%20work%20introduces%20the%20Geometry-Informed%20Neural%20Operator%20Transformer%20%28GINOT%29%2C%20which%20integrates%20the%20transformer%20architecture%20with%20the%20neural%20operator%20framework%20to%20enable%20forward%20predictions%20for%20arbitrary%20geometries.%20GINOT%20encodes%20the%20surface%20points%20cloud%20of%20a%20geometry%20using%20a%20sampling%20and%20grouping%20mechanism%20combined%20with%20an%20attention%20mechanism%2C%20ensuring%20invariance%20to%20point%20order%20and%20padding%20while%20maintaining%20robustness%20to%20variations%20in%20point%20density.%20The%20geometry%20information%20is%20seamlessly%20integrated%20with%20query%20points%20in%20the%20solution%20decoder%20through%20the%20attention%20mechanism.%20The%20performance%20of%20GINOT%20is%20validated%20on%20multiple%20challenging%20datasets%2C%20showcasing%20its%20high%20accuracy%20and%20strong%20generalization%20capabilities%20for%20complex%20and%20arbitrary%202D%20and%203D%20geometries.%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.2504.19452%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2504.19452%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-05-27T19%3A05%3A47Z%22%7D%7D%2C%7B%22key%22%3A%22AFKUP5GX%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Li%20and%20Shao%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3ELi%2C%20S.%20%26amp%3B%20Shao%2C%20C.%20Multi-Modal%20Fusion%20of%20In-Situ%20Video%20Data%20and%20Process%20Parameters%20for%20Online%20Forecasting%20of%20Cookie%20Drying%20Readiness.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2504.15599%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2504.15599%3C%5C%2Fa%3E%20%282025%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Multi-Modal%20Fusion%20of%20In-Situ%20Video%20Data%20and%20Process%20Parameters%20for%20Online%20Forecasting%20of%20Cookie%20Drying%20Readiness%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shichen%22%2C%22lastName%22%3A%22Li%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Chenhui%22%2C%22lastName%22%3A%22Shao%22%7D%5D%2C%22abstractNote%22%3A%22Food%20drying%20is%20essential%20for%20food%20production%2C%20extending%20shelf%20life%2C%20and%20reducing%20transportation%20costs.%20Accurate%20real-time%20forecasting%20of%20drying%20readiness%20is%20crucial%20for%20minimizing%20energy%20consumption%2C%20improving%20productivity%2C%20and%20ensuring%20product%20quality.%20However%2C%20this%20remains%20challenging%20due%20to%20the%20dynamic%20nature%20of%20drying%2C%20limited%20data%20availability%2C%20and%20the%20lack%20of%20effective%20predictive%20analytical%20methods.%20To%20address%20this%20gap%2C%20we%20propose%20an%20end-to-end%20multi-modal%20data%20fusion%20framework%20that%20integrates%20in-situ%20video%20data%20with%20process%20parameters%20for%20real-time%20food%20drying%20readiness%20forecasting.%20Our%20approach%20leverages%20a%20new%20encoder-decoder%20architecture%20with%20modality-specific%20encoders%20and%20a%20transformer-based%20decoder%20to%20effectively%20extract%20features%20while%20preserving%20the%20unique%20structure%20of%20each%20modality.%20We%20apply%20our%20approach%20to%20sugar%20cookie%20drying%2C%20where%20time-to-ready%20is%20predicted%20at%20each%20timestamp.%20Experimental%20results%20demonstrate%20that%20our%20model%20achieves%20an%20average%20prediction%20error%20of%20only%2015%20seconds%2C%20outperforming%20state-of-the-art%20data%20fusion%20methods%20by%2065.69%25%20and%20a%20video-only%20model%20by%2011.30%25.%20Additionally%2C%20our%20model%20balances%20prediction%20accuracy%2C%20model%20size%2C%20and%20computational%20efficiency%2C%20making%20it%20well-suited%20for%20heterogenous%20industrial%20datasets.%20The%20proposed%20model%20is%20extensible%20to%20various%20other%20industrial%20modality%20fusion%20tasks%20for%20online%20decision-making.%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.2504.15599%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2504.15599%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-05-27T18%3A58%3A55Z%22%7D%7D%2C%7B%22key%22%3A%222E96TFA7%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Gauba%20et%20al.%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3EGauba%2C%20A.%20%3Ci%3Eet%20al.%3C%5C%2Fi%3E%20AgMMU%3A%20A%20Comprehensive%20Agricultural%20Multimodal%20Understanding%20and%20Reasoning%20Benchmark.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2504.10568%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2504.10568%3C%5C%2Fa%3E%20%282025%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22AgMMU%3A%20A%20Comprehensive%20Agricultural%20Multimodal%20Understanding%20and%20Reasoning%20Benchmark%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Aruna%22%2C%22lastName%22%3A%22Gauba%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Irene%22%2C%22lastName%22%3A%22Pi%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yunze%22%2C%22lastName%22%3A%22Man%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ziqi%22%2C%22lastName%22%3A%22Pang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Vikram%20S.%22%2C%22lastName%22%3A%22Adve%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%22We%20curate%20a%20dataset%20AgMMU%20for%20evaluating%20and%20developing%20vision-language%20models%20%28VLMs%29%20to%20produce%20factually%20accurate%20answers%20for%20knowledge-intensive%20expert%20domains.%20Our%20AgMMU%20concentrates%20on%20one%20of%20the%20most%20socially%20beneficial%20domains%2C%20agriculture%2C%20which%20requires%20connecting%20detailed%20visual%20observation%20with%20precise%20knowledge%20to%20diagnose%2C%20e.g.%2C%20pest%20identification%2C%20management%20instructions%2C%20etc.%20As%20a%20core%20uniqueness%20of%20our%20dataset%2C%20all%20facts%2C%20questions%2C%20and%20answers%20are%20extracted%20from%20116%2C231%20conversations%20between%20real-world%20users%20and%20authorized%20agricultural%20experts.%20After%20a%20three-step%20dataset%20curation%20pipeline%20with%20GPT-4o%2C%20LLaMA%20models%2C%20and%20human%20verification%2C%20AgMMU%20features%20an%20evaluation%20set%20of%205%2C460%20multiple-choice%20questions%20%28MCQs%29%20and%20open-ended%20questions%20%28OEQs%29.%20We%20also%20provide%20a%20development%20set%20that%20contains%20205%2C399%20pieces%20of%20agricultural%20knowledge%20information%2C%20including%20disease%20identification%2C%20symptoms%20descriptions%2C%20management%20instructions%2C%20insect%20and%20pest%20identification%2C%20and%20species%20identification.%20As%20a%20multimodal%20factual%20dataset%2C%20it%20reveals%20that%20existing%20VLMs%20face%20significant%20challenges%20with%20questions%20requiring%20both%20detailed%20perception%20and%20factual%20knowledge.%20Moreover%2C%20open-source%20VLMs%20still%20demonstrate%20a%20substantial%20performance%20gap%20compared%20to%20proprietary%20ones.%20To%20advance%20knowledge-intensive%20VLMs%2C%20we%20conduct%20fine-tuning%20experiments%20using%20our%20development%20set%2C%20which%20improves%20LLaVA-1.5%20evaluation%20accuracy%20by%20up%20to%203.1%25.%20We%20hope%20that%20AgMMU%20can%20serve%20both%20as%20an%20evaluation%20benchmark%20dedicated%20to%20agriculture%20and%20a%20development%20suite%20for%20incorporating%20knowledge-intensive%20expertise%20into%20general-purpose%20VLMs.%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.2504.10568%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2504.10568%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-05-27T18%3A56%3A02Z%22%7D%7D%2C%7B%22key%22%3A%229YL9LIST%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%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3ELin%2C%20L.%2C%20Yu%2C%20X.%2C%20Pang%2C%20Z.%20%26amp%3B%20Wang%2C%20Y.-X.%20GLUS%3A%20Global-Local%20Reasoning%20Unified%20into%20A%20Single%20Large%20Language%20Model%20for%20Video%20Segmentation.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2504.07962%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2504.07962%3C%5C%2Fa%3E%20%282025%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22GLUS%3A%20Global-Local%20Reasoning%20Unified%20into%20A%20Single%20Large%20Language%20Model%20for%20Video%20Segmentation%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Lang%22%2C%22lastName%22%3A%22Lin%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Xueyang%22%2C%22lastName%22%3A%22Yu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ziqi%22%2C%22lastName%22%3A%22Pang%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%22This%20paper%20proposes%20a%20novel%20framework%20utilizing%20multi-modal%20large%20language%20models%20%28MLLMs%29%20for%20referring%20video%20object%20segmentation%20%28RefVOS%29.%20Previous%20MLLM-based%20methods%20commonly%20struggle%20with%20the%20dilemma%20between%20%5C%22Ref%5C%22%20and%20%5C%22VOS%5C%22%3A%20they%20either%20specialize%20in%20understanding%20a%20few%20key%20frames%20%28global%20reasoning%29%20or%20tracking%20objects%20on%20continuous%20frames%20%28local%20reasoning%29%2C%20and%20rely%20on%20external%20VOS%20or%20frame%20selectors%20to%20mitigate%20the%20other%20end%20of%20the%20challenge.%20However%2C%20our%20framework%20GLUS%20shows%20that%20global%20and%20local%20consistency%20can%20be%20unified%20into%20a%20single%20video%20segmentation%20MLLM%3A%20a%20set%20of%20sparse%20%5C%22context%20frames%5C%22%20provides%20global%20information%2C%20while%20a%20stream%20of%20continuous%20%5C%22query%20frames%5C%22%20conducts%20local%20object%20tracking.%20This%20is%20further%20supported%20by%20jointly%20training%20the%20MLLM%20with%20a%20pre-trained%20VOS%20memory%20bank%20to%20simultaneously%20digest%20short-range%20and%20long-range%20temporal%20information.%20To%20improve%20the%20information%20efficiency%20within%20the%20limited%20context%20window%20of%20MLLMs%2C%20we%20introduce%20object%20contrastive%20learning%20to%20distinguish%20hard%20false-positive%20objects%20and%20a%20self-refined%20framework%20to%20identify%20crucial%20frames%20and%20perform%20propagation.%20By%20collectively%20integrating%20these%20insights%2C%20our%20GLUS%20delivers%20a%20simple%20yet%20effective%20baseline%2C%20achieving%20new%20state-of-the-art%20for%20MLLMs%20on%20the%20MeViS%20and%20Ref-Youtube-VOS%20benchmark.%20Our%20project%20page%20is%20at%20https%3A%5C%2F%5C%2Fglus-video.github.io%5C%2F.%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.2504.07962%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2504.07962%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-05-27T18%3A50%3A51Z%22%7D%7D%2C%7B%22key%22%3A%22SZ8WZLSC%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Liu%20et%20al.%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3ELiu%2C%20Q.%2C%20Koric%2C%20S.%2C%20Abueidda%2C%20D.%2C%20Meidani%2C%20H.%20%26amp%3B%20Geubelle%2C%20P.%20Towards%20Signed%20Distance%20Function%20based%20Metamaterial%20Design%3A%20Neural%20Operator%20Transformer%20for%20Forward%20Prediction%20and%20Diffusion%20Model%20for%20Inverse%20Design.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2504.01195%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2504.01195%3C%5C%2Fa%3E%20%282025%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Towards%20Signed%20Distance%20Function%20based%20Metamaterial%20Design%3A%20Neural%20Operator%20Transformer%20for%20Forward%20Prediction%20and%20Diffusion%20Model%20for%20Inverse%20Design%22%2C%22creators%22%3A%5B%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%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%22Hadi%22%2C%22lastName%22%3A%22Meidani%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Philippe%22%2C%22lastName%22%3A%22Geubelle%22%7D%5D%2C%22abstractNote%22%3A%22The%20inverse%20design%20of%20metamaterial%20architectures%20presents%20a%20significant%20challenge%2C%20particularly%20for%20nonlinear%20mechanical%20properties%20involving%20large%20deformations%2C%20buckling%2C%20contact%2C%20and%20plasticity.%20Traditional%20methods%2C%20such%20as%20gradient-based%20optimization%2C%20and%20recent%20generative%20deep-learning%20approaches%20often%20rely%20on%20binary%20pixel-based%20representations%2C%20which%20introduce%20jagged%20edges%20that%20hinder%20finite%20element%20%28FE%29%20simulations%20and%203D%20printing.%20To%20overcome%20these%20challenges%2C%20we%20propose%20an%20inverse%20design%20framework%20that%20utilizes%20a%20signed%20distance%20function%20%28SDF%29%20representation%20combined%20with%20a%20conditional%20diffusion%20model.%20The%20SDF%20provides%20a%20smooth%20boundary%20representation%2C%20eliminating%20the%20need%20for%20post-processing%20and%20ensuring%20compatibility%20with%20FE%20simulations%20and%20manufacturing%20methods.%20A%20classifier-free%20guided%20diffusion%20model%20is%20trained%20to%20generate%20SDFs%20conditioned%20on%20target%20macroscopic%20stress-strain%20curves%2C%20enabling%20efficient%20one-shot%20design%20synthesis.%20To%20assess%20the%20mechanical%20response%20of%20the%20generated%20designs%2C%20we%20introduce%20a%20forward%20prediction%20model%20based%20on%20Neural%20Operator%20Transformers%20%28NOT%29%2C%20which%20accurately%20predicts%20homogenized%20stress-strain%20curves%20and%20local%20solution%20fields%20for%20arbitrary%20geometries%20with%20irregular%20query%20meshes.%20This%20approach%20enables%20a%20closed-loop%20process%20for%20general%20metamaterial%20design%2C%20offering%20a%20pathway%20for%20the%20development%20of%20advanced%20functional%20materials.%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.2504.01195%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2504.01195%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-05-27T18%3A38%3A38Z%22%7D%7D%2C%7B%22key%22%3A%22CVH7XHDW%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Cui%20et%20al.%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3ECui%2C%20S.%20%3Ci%3Eet%20al.%3C%5C%2Fi%3E%20Characterizing%20GPU%20Resilience%20and%20Impact%20on%20AI%5C%2FHPC%20Systems.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2503.11901%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2503.11901%3C%5C%2Fa%3E%20%282025%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Characterizing%20GPU%20Resilience%20and%20Impact%20on%20AI%5C%2FHPC%20Systems%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shengkun%22%2C%22lastName%22%3A%22Cui%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Archit%22%2C%22lastName%22%3A%22Patke%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ziheng%22%2C%22lastName%22%3A%22Chen%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Aditya%22%2C%22lastName%22%3A%22Ranjan%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Hung%22%2C%22lastName%22%3A%22Nguyen%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Phuong%22%2C%22lastName%22%3A%22Cao%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Saurabh%22%2C%22lastName%22%3A%22Jha%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Brett%22%2C%22lastName%22%3A%22Bode%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Gregory%22%2C%22lastName%22%3A%22Bauer%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Chandra%22%2C%22lastName%22%3A%22Narayanaswami%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Daby%22%2C%22lastName%22%3A%22Sow%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Catello%22%2C%22lastName%22%3A%22Di%20Martino%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Zbigniew%20T.%22%2C%22lastName%22%3A%22Kalbarczyk%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ravishankar%20K.%22%2C%22lastName%22%3A%22Iyer%22%7D%5D%2C%22abstractNote%22%3A%22In%20this%20study%2C%20we%20characterize%20GPU%20failures%20in%20Delta%2C%20the%20current%20large-scale%20AI%20system%20with%20over%20600%20petaflops%20of%20peak%20compute%20throughput.%20The%20system%20comprises%20GPU%20and%20non-GPU%20nodes%20with%20modern%20AI%20accelerators%2C%20such%20as%20NVIDIA%20A40%2C%20A100%2C%20and%20H100%20GPUs.%20The%20study%20uses%20two%20and%20a%20half%20years%20of%20data%20on%20GPU%20errors.%20We%20evaluate%20the%20resilience%20of%20GPU%20hardware%20components%20to%20determine%20the%20vulnerability%20of%20different%20GPU%20components%20to%20failure%20and%20their%20impact%20on%20the%20GPU%20and%20node%20availability.%20We%20measure%20the%20key%20propagation%20paths%20in%20GPU%20hardware%2C%20GPU%20interconnect%20%28NVLink%29%2C%20and%20GPU%20memory.%20Finally%2C%20we%20evaluate%20the%20impact%20of%20the%20observed%20GPU%20errors%20on%20user%20jobs.%20Our%20key%20findings%20are%3A%20%28i%29%20Contrary%20to%20common%20beliefs%2C%20GPU%20memory%20is%20over%2030x%20more%20reliable%20than%20GPU%20hardware%20in%20terms%20of%20MTBE%20%28mean%20time%20between%20errors%29.%20%28ii%29%20The%20newly%20introduced%20GSP%20%28GPU%20System%20Processor%29%20is%20the%20most%20vulnerable%20GPU%20hardware%20component.%20%28iii%29%20NVLink%20errors%20did%20not%20always%20lead%20to%20user%20job%20failure%2C%20and%20we%20attribute%20it%20to%20the%20underlying%20error%20detection%20and%20retry%20mechanisms%20employed.%20%28iv%29%20We%20show%20multiple%20examples%20of%20hardware%20errors%20originating%20from%20one%20of%20the%20key%20GPU%20hardware%20components%2C%20leading%20to%20application%20failure.%20%28v%29%20We%20project%20the%20impact%20of%20GPU%20node%20availability%20on%20larger%20scales%20with%20emulation%20and%20find%20that%20significant%20overprovisioning%20between%205-20%25%20would%20be%20necessary%20to%20handle%20GPU%20failures.%20If%20GPU%20availability%20were%20improved%20to%2099.9%25%2C%20the%20overprovisioning%20would%20be%20reduced%20by%204x.%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.2503.11901%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2503.11901%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-03-20T14%3A45%3A55Z%22%7D%7D%2C%7B%22key%22%3A%225IABTFLG%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Cheng%20and%20Schwing%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3ECheng%2C%20H.%20K.%20%26amp%3B%20Schwing%2C%20A.%20The%20Curse%20of%20Conditions%3A%20Analyzing%20and%20Improving%20Optimal%20Transport%20for%20Conditional%20Flow-Based%20Generation.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2503.10636%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2503.10636%3C%5C%2Fa%3E%20%282025%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22The%20Curse%20of%20Conditions%3A%20Analyzing%20and%20Improving%20Optimal%20Transport%20for%20Conditional%20Flow-Based%20Generation%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Ho%20Kei%22%2C%22lastName%22%3A%22Cheng%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Alexander%22%2C%22lastName%22%3A%22Schwing%22%7D%5D%2C%22abstractNote%22%3A%22Minibatch%20optimal%20transport%20coupling%20straightens%20paths%20in%20unconditional%20flow%20matching.%20This%20leads%20to%20computationally%20less%20demanding%20inference%20as%20fewer%20integration%20steps%20and%20less%20complex%20numerical%20solvers%20can%20be%20employed%20when%20numerically%20solving%20an%20ordinary%20differential%20equation%20at%20test%20time.%20However%2C%20in%20the%20conditional%20setting%2C%20minibatch%20optimal%20transport%20falls%20short.%20This%20is%20because%20the%20default%20optimal%20transport%20mapping%20disregards%20conditions%2C%20resulting%20in%20a%20conditionally%20skewed%20prior%20distribution%20during%20training.%20In%20contrast%2C%20at%20test%20time%2C%20we%20have%20no%20access%20to%20the%20skewed%20prior%2C%20and%20instead%20sample%20from%20the%20full%2C%20unbiased%20prior%20distribution.%20This%20gap%20between%20training%20and%20testing%20leads%20to%20a%20subpar%20performance.%20To%20bridge%20this%20gap%2C%20we%20propose%20conditional%20optimal%20transport%20C%5E2OT%20that%20adds%20a%20conditional%20weighting%20term%20in%20the%20cost%20matrix%20when%20computing%20the%20optimal%20transport%20assignment.%20Experiments%20demonstrate%20that%20this%20simple%20fix%20works%20with%20both%20discrete%20and%20continuous%20conditions%20in%208gaussians-to-moons%2C%20CIFAR-10%2C%20ImageNet-32x32%2C%20and%20ImageNet-256x256.%20Our%20method%20performs%20better%20overall%20compared%20to%20the%20existing%20baselines%20across%20different%20function%20evaluation%20budgets.%20Code%20is%20available%20at%20https%3A%5C%2F%5C%2Fhkchengrex.github.io%5C%2FC2OT%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.2503.10636%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2503.10636%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-03-18T13%3A50%3A56Z%22%7D%7D%2C%7B%22key%22%3A%223AFQPTAR%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Arora%20et%20al.%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3EArora%2C%20S.%20%3Ci%3Eet%20al.%3C%5C%2Fi%3E%20ESPnet-SDS%3A%20Unified%20Toolkit%20and%20Demo%20for%20Spoken%20Dialogue%20Systems.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2503.08533%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2503.08533%3C%5C%2Fa%3E%20%282025%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22ESPnet-SDS%3A%20Unified%20Toolkit%20and%20Demo%20for%20Spoken%20Dialogue%20Systems%22%2C%22creators%22%3A%5B%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%22Yifan%22%2C%22lastName%22%3A%22Peng%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jiatong%22%2C%22lastName%22%3A%22Shi%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%22William%22%2C%22lastName%22%3A%22Chen%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shikhar%22%2C%22lastName%22%3A%22Bharadwaj%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Hayato%22%2C%22lastName%22%3A%22Futami%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yosuke%22%2C%22lastName%22%3A%22Kashiwagi%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Emiru%22%2C%22lastName%22%3A%22Tsunoo%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Shuichiro%22%2C%22lastName%22%3A%22Shimizu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Vaibhav%22%2C%22lastName%22%3A%22Srivastav%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%22Advancements%20in%20audio%20foundation%20models%20%28FMs%29%20have%20fueled%20interest%20in%20end-to-end%20%28E2E%29%20spoken%20dialogue%20systems%2C%20but%20different%20web%20interfaces%20for%20each%20system%20makes%20it%20challenging%20to%20compare%20and%20contrast%20them%20effectively.%20Motivated%20by%20this%2C%20we%20introduce%20an%20open-source%2C%20user-friendly%20toolkit%20designed%20to%20build%20unified%20web%20interfaces%20for%20various%20cascaded%20and%20E2E%20spoken%20dialogue%20systems.%20Our%20demo%20further%20provides%20users%20with%20the%20option%20to%20get%20on-the-fly%20automated%20evaluation%20metrics%20such%20as%20%281%29%20latency%2C%20%282%29%20ability%20to%20understand%20user%20input%2C%20%283%29%20coherence%2C%20diversity%2C%20and%20relevance%20of%20system%20response%2C%20and%20%284%29%20intelligibility%20and%20audio%20quality%20of%20system%20output.%20Using%20the%20evaluation%20metrics%2C%20we%20compare%20various%20cascaded%20and%20E2E%20spoken%20dialogue%20systems%20with%20a%20human-human%20conversation%20dataset%20as%20a%20proxy.%20Our%20analysis%20demonstrates%20that%20the%20toolkit%20allows%20researchers%20to%20effortlessly%20compare%20and%20contrast%20different%20technologies%2C%20providing%20valuable%20insights%20such%20as%20current%20E2E%20systems%20having%20poorer%20audio%20quality%20and%20less%20diverse%20responses.%20An%20example%20demo%20produced%20using%20our%20toolkit%20is%20publicly%20available%20here%3A%20https%3A%5C%2F%5C%2Fhuggingface.co%5C%2Fspaces%5C%2FSiddhant%5C%2FVoice_Assistant_Demo.%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.2503.08533%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2503.08533%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-03-17T15%3A35%3A59Z%22%7D%7D%2C%7B%22key%22%3A%22FG5ZP5ZB%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22creatorSummary%22%3A%22Wilfong%20et%20al.%22%2C%22parsedDate%22%3A%222025%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3EWilfong%2C%20B.%20%3Ci%3Eet%20al.%3C%5C%2Fi%3E%20MFC%205.0%3A%20An%20exascale%20many-physics%20flow%20solver.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2503.07953%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2503.07953%3C%5C%2Fa%3E%20%282025%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22MFC%205.0%3A%20An%20exascale%20many-physics%20flow%20solver%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Benjamin%22%2C%22lastName%22%3A%22Wilfong%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Henry%20A.%20Le%22%2C%22lastName%22%3A%22Berre%22%7D%2C%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%22Ansh%22%2C%22lastName%22%3A%22Gupta%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Diego%22%2C%22lastName%22%3A%22Vaca-Revelo%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Dimitrios%22%2C%22lastName%22%3A%22Adam%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Haocheng%22%2C%22lastName%22%3A%22Yu%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Hyeoksu%22%2C%22lastName%22%3A%22Lee%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Jose%20Rodolfo%22%2C%22lastName%22%3A%22Chreim%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Mirelys%20Carcana%22%2C%22lastName%22%3A%22Barbosa%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Yanjun%22%2C%22lastName%22%3A%22Zhang%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Esteban%22%2C%22lastName%22%3A%22Cisneros-Garibay%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Aswin%22%2C%22lastName%22%3A%22Gnanaskandan%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Mauro%22%2C%22lastName%22%3A%22Rodriguez%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Reuben%20D.%22%2C%22lastName%22%3A%22Budiardja%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Stephen%22%2C%22lastName%22%3A%22Abbott%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Tim%22%2C%22lastName%22%3A%22Colonius%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%22Engineering%2C%20medicine%2C%20and%20the%20fundamental%20sciences%20broadly%20rely%20on%20flow%20simulations%2C%20making%20performant%20computational%20fluid%20dynamics%20solvers%20an%20open%20source%20software%20mainstay.%20A%20previous%20work%20made%20MFC%203.0%20a%20published%20open%20source%20source%20solver%20with%20many%20features.%20MFC%205.0%20is%20a%20marked%20update%20to%20MFC%203.0%2C%20including%20a%20broad%20set%20of%20well-established%20and%20novel%20physical%20models%20and%20numerical%20methods%20and%20the%20introduction%20of%20GPU%20and%20APU%20%28or%20superchip%29%20acceleration.%20We%20exhibit%20state-of-the-art%20performance%20and%20ideal%20scaling%20on%20the%20first%20two%20exascale%20supercomputers%2C%20OLCF%20Frontier%20and%20LLNL%20El%20Capitan.%20Combined%20with%20MFC%27s%20single-GPU%5C%2FAPU%20performance%2C%20MFC%20achieves%20exascale%20computation%20in%20practice.%20With%20these%20capabilities%2C%20MFC%20has%20evolved%20into%20a%20tool%20for%20conducting%20simulations%20that%20many%20engineering%20challenge%20problems%20hinge%20upon.%20New%20physical%20features%20include%20the%20immersed%20boundary%20method%2C%20%24N%24-fluid%20phase%20change%2C%20Euler--Euler%20and%20Euler--Lagrange%20sub-grid%20bubble%20models%2C%20fluid-structure%20interaction%2C%20hypo-%20and%20hyper-elastic%20materials%2C%20chemically%20reacting%20flow%2C%20two-material%20surface%20tension%2C%20and%20more.%20Numerical%20techniques%20now%20represent%20the%20current%20state-of-the-art%2C%20including%20general%20relaxation%20characteristic%20boundary%20conditions%2C%20WENO%20variants%2C%20Strang%20splitting%20for%20stiff%20sub-grid%20flow%20features%2C%20and%20low%20Mach%20number%20treatments.%20Weak%20scaling%20to%20tens%20of%20thousands%20of%20GPUs%20on%20OLCF%20Frontier%20and%20LLNL%20El%20Capitan%20see%20efficiencies%20within%205%25%20of%20ideal%20to%20over%2090%25%20of%20their%20respective%20system%20sizes.%20Strong%20scaling%20results%20for%20a%2016-time%20increase%20in%20device%20count%20show%20parallel%20efficiencies%20over%2090%25%20on%20OLCF%20Frontier.%20Other%20MFC%20improvements%20include%20ensuring%20code%20resilience%20and%20correctness%20with%20a%20continuous%20integration%20suite%2C%20the%20use%20of%20metaprogramming%20to%20reduce%20code%20length%20and%20maintain%20performance%20portability%2C%20and%20efficient%20computational%20representations%20for%20chemical%20reactions%20and%20thermodynamics%20via%20code%20generation%20with%20Pyrometheus.%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.2503.07953%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2503.07953%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-03-17T15%3A30%3A07Z%22%7D%7D%2C%7B%22key%22%3A%22WRYRN6MD%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22lastModifiedByUser%22%3A%7B%22id%22%3A15215433%2C%22username%22%3A%22sarahdc-Illinois%22%2C%22name%22%3A%22%22%2C%22links%22%3A%7B%22alternate%22%3A%7B%22href%22%3A%22https%3A%5C%2F%5C%2Fwww.zotero.org%5C%2Fsarahdc-illinois%22%2C%22type%22%3A%22text%5C%2Fhtml%22%7D%7D%7D%2C%22creatorSummary%22%3A%22Kacmaz%20et%20al.%22%2C%22parsedDate%22%3A%222024%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3EKacmaz%2C%20S.%2C%20Haas%2C%20R.%20%26amp%3B%20Huerta%2C%20E.%20A.%20Machine%20learning-driven%20conservative-to-primitive%20conversion%20in%20hybrid%20piecewise%20polytropic%20and%20tabulated%20equations%20of%20state.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2412.07836%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2412.07836%3C%5C%2Fa%3E%20%282024%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22Machine%20learning-driven%20conservative-to-primitive%20conversion%20in%20hybrid%20piecewise%20polytropic%20and%20tabulated%20equations%20of%20state%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Semih%22%2C%22lastName%22%3A%22Kacmaz%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Roland%22%2C%22lastName%22%3A%22Haas%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22E.%20A.%22%2C%22lastName%22%3A%22Huerta%22%7D%5D%2C%22abstractNote%22%3A%22We%20present%20a%20novel%20machine%20learning%20%28ML%29%20method%20to%20accelerate%20conservative-to-primitive%20inversion%2C%20focusing%20on%20hybrid%20piecewise%20polytropic%20and%20tabulated%20equations%20of%20state.%20Traditional%20root-finding%20techniques%20are%20computationally%20expensive%2C%20particularly%20for%20large-scale%20relativistic%20hydrodynamics%20simulations.%20To%20address%20this%2C%20we%20employ%20feedforward%20neural%20networks%20%28NNC2PS%20and%20NNC2PL%29%2C%20trained%20in%20PyTorch%20and%20optimized%20for%20GPU%20inference%20using%20NVIDIA%20TensorRT%2C%20achieving%20significant%20speedups%20with%20minimal%20accuracy%20loss.%20The%20NNC2PS%20model%20achieves%20%24%20L_1%20%24%20and%20%24%20L_%5C%5Cinfty%20%24%20errors%20of%20%24%204.54%20%5C%5Ctimes%2010%5E%7B-7%7D%20%24%20and%20%24%203.44%20%5C%5Ctimes%2010%5E%7B-6%7D%20%24%2C%20respectively%2C%20while%20the%20NNC2PL%20model%20exhibits%20even%20lower%20error%20values.%20TensorRT%20optimization%20with%20mixed-precision%20deployment%20substantially%20accelerates%20performance%20compared%20to%20traditional%20root-finding%20methods.%20Specifically%2C%20the%20mixed-precision%20TensorRT%20engine%20for%20NNC2PS%20achieves%20inference%20speeds%20approximately%20400%20times%20faster%20than%20a%20traditional%20single-threaded%20CPU%20implementation%20for%20a%20dataset%20size%20of%201%2C000%2C000%20points.%20Ideal%20parallelization%20across%20an%20entire%20compute%20node%20in%20the%20Delta%20supercomputer%20%28Dual%20AMD%2064%20core%202.45%20GHz%20Milan%20processors%3B%20and%208%20NVIDIA%20A100%20GPUs%20with%2040%20GB%20HBM2%20RAM%20and%20NVLink%29%20predicts%20a%2025-fold%20speedup%20for%20TensorRT%20over%20an%20optimally-parallelized%20numerical%20method%20when%20processing%208%20million%20data%20points.%20Moreover%2C%20the%20ML%20method%20exhibits%20sub-linear%20scaling%20with%20increasing%20dataset%20sizes.%20We%20release%20the%20scientific%20software%20developed%2C%20enabling%20further%20validation%20and%20extension%20of%20our%20findings.%20This%20work%20underscores%20the%20potential%20of%20ML%2C%20combined%20with%20GPU%20optimization%20and%20model%20quantization%2C%20to%20accelerate%20conservative-to-primitive%20inversion%20in%20relativistic%20hydrodynamics%20simulations.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222024%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2412.07836%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2412.07836%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%222GS7DZHJ%22%2C%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-02-04T17%3A01%3A58Z%22%7D%7D%2C%7B%22key%22%3A%22YNQMLUHQ%22%2C%22library%22%3A%7B%22id%22%3A5854943%7D%2C%22meta%22%3A%7B%22lastModifiedByUser%22%3A%7B%22id%22%3A15215433%2C%22username%22%3A%22sarahdc-Illinois%22%2C%22name%22%3A%22%22%2C%22links%22%3A%7B%22alternate%22%3A%7B%22href%22%3A%22https%3A%5C%2F%5C%2Fwww.zotero.org%5C%2Fsarahdc-illinois%22%2C%22type%22%3A%22text%5C%2Fhtml%22%7D%7D%7D%2C%22creatorSummary%22%3A%22Tiki%20et%20al.%22%2C%22parsedDate%22%3A%222024%22%2C%22numChildren%22%3A0%7D%2C%22bib%22%3A%22%3Cdiv%20class%3D%5C%22csl-bib-body%5C%22%20style%3D%5C%22line-height%3A%202%3B%20%5C%22%3E%5Cn%20%20%3Cdiv%20class%3D%5C%22csl-entry%5C%22%20style%3D%5C%22clear%3A%20left%3B%20%5C%22%3E%5Cn%20%20%20%20%3Cdiv%20class%3D%5C%22csl-left-margin%5C%22%20style%3D%5C%22float%3A%20left%3B%20padding-right%3A%200.5em%3B%20text-align%3A%20right%3B%20width%3A%201em%3B%5C%22%3E1.%3C%5C%2Fdiv%3E%3Cdiv%20class%3D%5C%22csl-right-inline%5C%22%20style%3D%5C%22margin%3A%200%20.4em%200%201.5em%3B%5C%22%3ETiki%2C%20V.%2C%20Pham%2C%20K.%20%26amp%3B%20Huerta%2C%20E.%20AI%20forecasting%20of%20higher-order%20wave%20modes%20of%20spinning%20binary%20black%20hole%20mergers.%20Preprint%20at%20%3Ca%20class%3D%27zp-DOIURL%27%20href%3D%27https%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2409.03833%27%3Ehttps%3A%5C%2F%5C%2Fdoi.org%5C%2F10.48550%5C%2FARXIV.2409.03833%3C%5C%2Fa%3E%20%282024%29.%3C%5C%2Fdiv%3E%5Cn%20%20%3C%5C%2Fdiv%3E%5Cn%3C%5C%2Fdiv%3E%22%2C%22data%22%3A%7B%22itemType%22%3A%22preprint%22%2C%22title%22%3A%22AI%20forecasting%20of%20higher-order%20wave%20modes%20of%20spinning%20binary%20black%20hole%20mergers%22%2C%22creators%22%3A%5B%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Victoria%22%2C%22lastName%22%3A%22Tiki%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Kiet%22%2C%22lastName%22%3A%22Pham%22%7D%2C%7B%22creatorType%22%3A%22author%22%2C%22firstName%22%3A%22Eliu%22%2C%22lastName%22%3A%22Huerta%22%7D%5D%2C%22abstractNote%22%3A%22We%20present%20a%20physics-inspired%20transformer%20model%20that%20predicts%20the%20non-linear%20dynamics%20of%20higher-order%20wave%20modes%20emitted%20by%20quasi-circular%2C%20spinning%2C%20non-precessing%20binary%20black%20hole%20mergers.%20The%20model%20forecasts%20the%20waveform%20evolution%20from%20the%20pre-merger%20phase%20through%20the%20ringdown%2C%20starting%20with%20an%20input%20time-series%20spanning%20%24%20t%20%5C%5Cin%20%5B-5000%5C%5Ctextrm%7BM%7D%2C%20-100%5C%5Ctextrm%7BM%7D%29%20%24.%20The%20merger%20event%2C%20defined%20as%20the%20peak%20amplitude%20of%20waveforms%20that%20include%20the%20%24l%20%3D%20%7Cm%7C%20%3D%202%24%20modes%2C%20occurs%20at%20%24%20t%20%3D%200%5C%5Ctextrm%7BM%7D%20%24.%20The%20transformer%20then%20generates%20predictions%20over%20the%20time%20range%20%24%20t%20%5C%5Cin%20%5B-100%5C%5Ctextrm%7BM%7D%2C%20130%5C%5Ctextrm%7BM%7D%5D%20%24.%20We%20produced%20training%2C%20evaluation%20and%20test%20sets%20using%20the%20NRHybSur3dq8%20model%2C%20considering%20a%20signal%20manifold%20defined%20by%20mass%20ratios%20%24%20q%20%5C%5Cin%20%5B1%2C%208%5D%20%24%3B%20spin%20components%20%24%20s%5Ez_%7B%5C%5C%7B1%2C2%5C%5C%7D%7D%20%5C%5Cin%20%5B-0.8%2C%200.8%5D%20%24%3B%20modes%20up%20to%20%24l%20%5C%5Cleq%204%24%2C%20including%20the%20%24%285%2C5%29%24%20mode%20but%20excluding%20the%20%24%284%2C0%29%24%20and%20%24%284%2C1%29%24%20modes%3B%20and%20inclination%20angles%20%24%5Cu03b8%5C%5Cin%20%5B0%2C%20%5Cu03c0%5D%24.%20We%20trained%20the%20model%20on%2014%2C440%2C761%20waveforms%2C%20completing%20the%20training%20in%2015%20hours%20using%2016%20NVIDIA%20A100%20GPUs%20in%20the%20Delta%20supercomputer.%20We%20used%204%20H100%20GPUs%20in%20the%20DeltaAI%20supercomputer%20to%20compute%2C%20within%207%20hours%2C%20the%20overlap%20between%20ground%20truth%20and%20predicted%20waveforms%20using%20a%20test%20set%20of%20840%2C000%20waveforms%2C%20finding%20that%20the%20mean%20and%20median%20overlaps%20over%20the%20test%20set%20are%200.996%20and%200.997%2C%20respectively.%20Additionally%2C%20we%20conducted%20interpretability%20studies%20to%20elucidate%20the%20waveform%20features%20utilized%20by%20our%20transformer%20model%20to%20produce%20accurate%20predictions.%20The%20scientific%20software%20used%20for%20this%20work%20is%20released%20with%20this%20manuscript.%22%2C%22genre%22%3A%22%22%2C%22repository%22%3A%22arXiv%22%2C%22archiveID%22%3A%22%22%2C%22date%22%3A%222024%22%2C%22DOI%22%3A%2210.48550%5C%2FARXIV.2409.03833%22%2C%22citationKey%22%3A%22%22%2C%22url%22%3A%22https%3A%5C%2F%5C%2Farxiv.org%5C%2Fabs%5C%2F2409.03833%22%2C%22language%22%3A%22%22%2C%22collections%22%3A%5B%222GS7DZHJ%22%2C%223NXZNVBX%22%5D%2C%22dateModified%22%3A%222025-02-04T16%3A55%3A20Z%22%7D%7D%5D%7D
1.
Butler, L. et al. ProxySPEX: Inference-Efficient Interpretability via Sparse Feature Interactions in LLMs. Preprint at https://doi.org/10.48550/ARXIV.2505.17495 (2025).
1.
Agarwal, S., Zhang, Z., Yuan, L., Han, J. & Peng, H. The Unreasonable Effectiveness of Entropy Minimization in LLM Reasoning. Preprint at https://doi.org/10.48550/ARXIV.2505.15134 (2025).
1.
Liu, Q. et al. Geometry-Informed Neural Operator Transformer. Preprint at https://doi.org/10.48550/ARXIV.2504.19452 (2025).
1.
Li, S. & Shao, C. Multi-Modal Fusion of In-Situ Video Data and Process Parameters for Online Forecasting of Cookie Drying Readiness. Preprint at https://doi.org/10.48550/ARXIV.2504.15599 (2025).
1.
Gauba, A. et al. AgMMU: A Comprehensive Agricultural Multimodal Understanding and Reasoning Benchmark. Preprint at https://doi.org/10.48550/ARXIV.2504.10568 (2025).
1.
Lin, L., Yu, X., Pang, Z. & Wang, Y.-X. GLUS: Global-Local Reasoning Unified into A Single Large Language Model for Video Segmentation. Preprint at https://doi.org/10.48550/ARXIV.2504.07962 (2025).
1.
Liu, Q., Koric, S., Abueidda, D., Meidani, H. & Geubelle, P. Towards Signed Distance Function based Metamaterial Design: Neural Operator Transformer for Forward Prediction and Diffusion Model for Inverse Design. Preprint at https://doi.org/10.48550/ARXIV.2504.01195 (2025).
1.
Cui, S. et al. Characterizing GPU Resilience and Impact on AI/HPC Systems. Preprint at https://doi.org/10.48550/ARXIV.2503.11901 (2025).
1.
Cheng, H. K. & Schwing, A. The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation. Preprint at https://doi.org/10.48550/ARXIV.2503.10636 (2025).
1.
Arora, S. et al. ESPnet-SDS: Unified Toolkit and Demo for Spoken Dialogue Systems. Preprint at https://doi.org/10.48550/ARXIV.2503.08533 (2025).
1.
Wilfong, B. et al. MFC 5.0: An exascale many-physics flow solver. Preprint at https://doi.org/10.48550/ARXIV.2503.07953 (2025).
1.
Kacmaz, S., Haas, R. & Huerta, E. A. Machine learning-driven conservative-to-primitive conversion in hybrid piecewise polytropic and tabulated equations of state. Preprint at https://doi.org/10.48550/ARXIV.2412.07836 (2024).
1.
Tiki, V., Pham, K. & Huerta, E. AI forecasting of higher-order wave modes of spinning binary black hole mergers. Preprint at https://doi.org/10.48550/ARXIV.2409.03833 (2024).