Pith. sign in

REVIEW 4 major objections 4 minor 17 references

AURA: A Fine-Grained Benchmark and Decomposed Metric for Audio-Visual Reasoning

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read High-scoring audio-visual AI models often reach the right answer through flawed logic, a new benchmark and metric show.

desk verdict A promising benchmark idea undercut by a false headline claim and a circular evaluation protocol; worth engaging with, but the current version overclaims. read the letter →

arxiv 2508.07470 v2 pith:U47DZTEF submitted 2025-08-10 cs.CV

Siminfar Samakoush Galougah , Rishie Raj , Sanjoy Chowdhury , Sayan Nag , Ramani Duraiswami This is my paper
classification cs.CV
keywords audio-visualreasoningbenchmarkAuraScorefidelityfactualconsistencycoreinferencemultimodallargelanguagemodelsquestionanswering
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that current audio-visual large language models (AV-LLMs) and omni-modal language models (OLMs) are not reasoning faithfully, even when they choose the correct multiple-choice answer. It introduces AURA, a benchmark of over 1,600 question-answer pairs across six cognitive tasks that are designed to be unanswerable from a single modality, and AuraScore, a metric that separates factual grounding from logical validity in a model's reasoning trace. Evaluations on state-of-the-art models report high accuracy, up to 92% on some tasks, but Factual Consistency and Core Inference scores that often fall below 45%. If the claim holds, accuracy-based evaluation overstates multimodal comprehension, and reasoning fidelity should be part of standard model assessment.

What carries the argument

The load-bearing machine is AuraScore's two-part decomposition plus the benchmark's cross-modal design. FCS is computed by asking GPT-4o to decompose the model's and ground-truth reasoning into atomic facts and compute the fraction of ground-truth facts correctly represented; CIS is computed by asking GPT-4o to sanitize both reasoning traces of specific entities and then running a natural-language-inference cross-encoder to measure whether the ground-truth logic entails the model's logic. The benchmark side is built from an automated pipeline: AuroraCap visual captions, FusionAudio audio captions, and Whisper transcripts feed GPT-4o, which writes questions, options, and gold reasoning under

What would settle it

Present a random sample of AURA video clips with the audio removed (or the video muted) to human annotators; if annotators can pick the correct answer above chance on questions the paper classifies as cross-modal, the claim that AURA questions are unanswerable from a single modality is falsified. Alternatively, if replacing GPT-4o-generated gold reasoning with human-written traces leaves FCS/CIS scores essentially unchanged, the metric is measuring something other than agreement with a generator-specific reasoning style.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a measurable dissociation between answer accuracy and reasoning quality in audio-visual models. AURA asks models questions about causality, timbre and pitch, tempo and audio-visual synchronization, unanswerability, implicit distractions, and performer skill profiling, with each question constructed so that neither audio nor video alone supplies the answer. AuraScore then scores the model's explanation in two parts: Factual Consistency, the fraction of ground-truth factual elements present in the explanation, and Core Inference, an entailment probability between sanitized logic traces. Across eleven models, including Qwen2.5-Omni, VideoLLaMA2

Load-bearing premise

The benchmark's ground-truth questions, answers, and gold reasoning are produced entirely by an automated captioning-and-GPT-4o pipeline, and the paper assumes these targets are correct, complete, and genuinely require both modalities, with no human validation or uni-modal shortcut check.

Editorial extensions

If this is right

  • Accuracy-only leaderboards can be misleading: a model with 80-90% accuracy can still have reasoning-fidelity scores below 45%, so reported capability should be qualified by reasoning metrics.
  • Models that answer by shallow heuristics or hallucination are exposed when they must justify the causal, spatial, or temporal link between modalities.
  • Fine-grained audio attributes such as tempo, synchronization, timbre, and pitch remain largely unsolved; every evaluated model fails the tempo/synchronization task by most metrics.
  • Unanswerability questions are answered correctly at high rates, but correct refusals are often supported by weak or absent reasoning, so honesty and grounding are separate abilities.
  • Training and evaluation should reward grounded, stepwise inference rather than final correctness, or models will continue to exploit answer shortcuts.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that because the gold reasoning is generated by GPT-4o from captions, low Factual Consistency may partly measure disagreement with GPT-4o's reasoning style, not only flawed logic; adding human-validated gold traces would let the field separate the two.
  • The same prompt-driven QA pipeline could be adapted to measure reasoning fidelity in other modality pairs, such as image-text or video-caption, where final accuracy masks unfaithful explanations.
  • A direct testable extension is to train on a reasoning-fidelity reward combining FCS and CIS; if the dissociation is real, such training should improve FCS/CIS without necessarily changing accuracy.
  • The task taxonomy (causality, unanswerability, synchronization, skill profiling) could serve as diagnostic axes for temporal and cross-modal alignment research, potentially exposing which architectural components handle rhythm and pitch versus spatial grounding.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces AURA, a benchmark of roughly 1,600 multiple-choice audio-visual QA pairs across six fine-grained reasoning categories (causal reasoning, unanswerability, timbre/pitch, tempo/AV synchronization, implicit distractions, skill profiling), generated automatically from AuroraCap/FusionAudio captions, Whisper transcripts, and GPT-4o. It also proposes AuraScore, a two-part metric that scores a model's reasoning trace only when the final answer is correct: Factual Consistency Score (FCS) is computed by GPT-4o against a gold reasoning trace, and Core Inference Score (CIS) is computed by GPT-4o-based sanitization followed by an NLI entailment check. The authors evaluate several open and closed AV-LLMs/OLMs and report high answer accuracy but low FCS/CIS, concluding that current models often answer correctly through flawed reasoning.

Significance. The motivation is timely and important: most AV benchmarks measure only final-answer accuracy, and a reliable way to assess reasoning fidelity would be a real contribution. The six task categories are well chosen, and the automated pipeline is a practical way to scale the benchmark. The decomposed FCS/CIS metric is a plausible design direction, and the paper provides a broad comparison across 11 models. If the metric were properly validated and the headline claims corrected, the benchmark could be useful to the community. However, as presented, the central claim—that SOTA models show a large gap between accuracy and reasoning fidelity—is not supported because the metric is uncalibrated and the abstract's quantitative claim is contradicted by the paper's own table.

major comments (4)
  1. [Abstract; Table 2] The abstract states that 'Factual Consistency and Core Inference scores fall below 45%.' Table 2 contains multiple counterexamples: Qwen2.5-Omni 7B reaches FCS 67.95 and CIS 45.23 on Implicit Distractions, FCS 68.00 and CIS 61.03 on Unanswerability; Ola reaches FCS 73.46 and CIS 61.10 on Unanswerability; Qwen2.5-Omni 3B reaches CIS 60.2 on Unanswerability. The claim is therefore false as an overall summary. The authors should either report per-task/per-model values faithfully or define a different aggregate (e.g., accuracy-conditioned averages) and check that the claim holds under that definition.
  2. [§4, Algorithm 1] Algorithm 1 returns FCS=0 and CIS=0 whenever the model's answer is incorrect. This means the reported FCS and CIS conflate answer accuracy with reasoning quality and are not direct measures of reasoning fidelity. If a model answers correctly on 80% of items and reasons well on all of those, its unconditional mean FCS/CIS can still appear low because the 20% incorrect items are scored zero. The paper should report conditional scores (FCS/CIS given a correct answer) or justify why zero-for-wrong-answer is the intended definition. This choice materially affects the headline 'high accuracy, low reasoning' finding.
  3. [§3.2, §4, §6] The evaluation is circular in an important way: gold reasoning is generated by GPT-4o from captioning inputs, and FCS is computed by GPT-4o comparing model reasoning to that gold reasoning. A model that reaches the correct answer through a different but equally valid reasoning path will be penalized for lexical/semantic mismatch with the GPT-4o target. Section 6 acknowledges GPT-4o dependence for FCS but does not acknowledge that the same model also defines the gold standard. The authors need human-validated gold reasoning, an analysis of acceptable alternative reasoning paths, and a comparison of FCS under different judge models (e.g., open-source LLMs) to show the metric measures grounding rather than style similarity.
  4. [§3.1, A.1] The paper's central premise is that AURA questions are 'unanswerable from a single modality,' but no evidence for this is provided. Since QA pairs are generated from separate visual captions, audio captions, and transcripts, some questions may be answerable from one modal description alone or even from the text captions. The benchmark should include uni-modal baselines (audio-only, video-only, text-only) for the evaluated models, and ideally a human study confirming that the questions require both modalities. Without this, the claim that the benchmark 'forces models to construct a valid logical path grounded in both audio and video' is unverified.
minor comments (4)
  1. [References] The reference list contains the placeholder '???? TODO.' This must be fixed before publication.
  2. [Figures 4–9 and tables] Multiple figure captions and example boxes contain 'Grand Truth' instead of 'Ground Truth' (e.g., Figures 7–9). Please correct these typos throughout.
  3. [Table 2] The table is dense and the column grouping is hard to read; consider splitting by task or using clearer spacing/header notation.
  4. [§5.2] The discussion is verbose and repeats qualitative claims that are not directly tied to the table. A concise summary or per-task table with effect sizes would be more informative.

Circularity Check

2 steps flagged · score 5.0 of 10

AuraScore's FCS/CIS reference standard and judge are both GPT-4o; low reasoning scores measure agreement with GPT-4o-generated gold reasoning, so the central 'flawed logic' claim is not independently established.

  1. other [Section 3.2 (QA Generation Pipeline) and Section 4 (Factual Consistency Score)]
    "The core of our pipeline involves feeding these annotations into a detailed Prompt Template for our generator LLM, GPT-4o. ... the prompt includes in-context examples to guide the model in generating a high-quality question, multiple-choice options, and the precise 'gold reasoning' for the correct answer. ... The FCS is calculated as the fraction of factual elements of the GT reasoning that are correctly and completely represented in the generated reasoning."

    FCS is defined as the fraction of the GPT-4o-generated 'gold reasoning' that appears in the model's reasoning. The same GPT-4o model produces both the gold reasoning and the FCS judgment. Thus a low FCS formally means 'low textual overlap with GPT-4o's generated reasoning trace,' not 'not grounded in perceptual evidence.' The paper's headline conclusion that models 'arrive at correct answers through flawed logic' depends on treating the GPT-4o-generated gold reasoning as an independent ground truth. That equivalence is not established; it is an unvalidated, self-referential reduction. The paper's own Limitations concede GPT-4o's stochasticity and bias but only for the evaluator, not for the gold reasoning itself.

  2. other [Section 4 (Core Inference Score) and Section 3.2]
    "We first employ GPT-4o to sanitize both the generated and GT reasoning traces. ... The sanitized GT reasoning (as the premise) and the sanitized generated reasoning (as the hypothesis) are input to a pre-trained cross-encoder NLI model ... The final CIS is the entailment probability score produced by the model."

    CIS inherits the same circularity as FCS. The GT reasoning premise is generated by GPT-4o from machine captions; then GPT-4o 'sanitizes' both texts to their abstract logical claims; then an NLI model measures entailment relative to that GPT-4o-derived premise. Consequently, CIS measures logical agreement with a GPT-4o-authored reasoning template, not an independent standard of logical validity. The central finding that models 'reason poorly' despite high accuracy is therefore at least partly an artifact of using the same model family to create the reference and to judge against it.

full rationale

The paper does not contain a formal self-citation chain or a fitted-parameter-called-prediction step. However, its central empirical claim—that SOTA models achieve high accuracy but low reasoning fidelity—rests on AuraScore's FCS and CIS. Both metrics compare model reasoning against 'gold reasoning' produced by GPT-4o in the automated QA-generation pipeline. FCS is explicitly the fraction of GPT-4o's gold facts reproduced, and CIS is an NLI entailment score against GPT-4o-sanitized gold logic. Because the reference standard and the judge are both GPT-4o, the reported reasoning gap partly measures disagreement with GPT-4o's style and content, not independently validated reasoning quality. The paper provides no human validation, error analysis, or uni-modal shortcut tests to establish that the gold reasoning is correct, complete, or uniquely valid. The Limitations section acknowledges GPT-4o's stochasticity and bias for the FCS evaluator but does not extend this concern to the gold reasoning. Additionally, the abstract's claim that FCS and CIS 'fall below 45%' is contradicted by Table 2 (e.g., Qwen2.5-Omni 7B FCS 67.95 on Implicit Distractions; Ola FCS 73.46 and CIS 61.10 on Unanswerability), which further weakens the headline generalization. These are correctness/validity concerns rather than pure derivation-equivalence, so the score is moderate, not extreme.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the unverified correctness of a fully automated, model-generated ground truth and on the validity of an LLM-based scoring metric that is not validated against humans. There are no free parameters fitted to data, but the assumptions above are load-bearing. No new physical or conceptual entities are introduced, so the invented-entities list is empty.

assumptions (4)
  • domain assumption The automated pipeline (AuroraCap, FusionAudio, Whisper captions plus GPT-4o generation) produces correct question-answer pairs and gold reasoning.
    Section 3.2 and Appendix A.2; no human validation or error analysis is provided, and the entire benchmark rests on the fidelity of these model-generated annotations.
  • domain assumption Every AURA question is answerable only by combining audio and visual information, with no uni-modal shortcuts.
    Section 3.1 and Appendix A.1 state this as a design principle, but no experiments or analyses demonstrate that uni-modal input fails on these questions.
  • domain assumption FCS as computed by GPT-4o and CIS via NLI on GPT-4o-sanitized traces measure the intended qualities of reasoning fidelity.
    Section 4; the metric is neither calibrated against human judgments nor compared with existing reasoning-evaluation baselines, and Section 6 admits stochasticity and bias concerns.
  • domain assumption GPT-4o is a stable scorer; the reported single-pass scores are representative.
    Section 4 and Section 6; the authors acknowledge stochasticity, and no repeated evaluations or confidence intervals are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AURA: A Fine-Grained Benchmark and Decomposed Metric for Audio-Visual Reasoning." pith.science (2026). https://pith.science/paper/U47DZTEF

@misc{pith2026250807470,
  author       = {Pith},
  title        = {Pith review of: AURA: A Fine-Grained Benchmark and Decomposed Metric for Audio-Visual Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U47DZTEF}},
  note         = {Machine review of arXiv:2508.07470}
}
read the original abstract

Current audio-visual (AV) benchmarks focus on final answer accuracy, overlooking the underlying reasoning process. This makes it difficult to distinguish genuine comprehension from correct answers derived through flawed reasoning or hallucinations. To address this, we introduce AURA (Audio-visual Understanding and Reasoning Assessment), a benchmark for evaluating the cross-modal reasoning capabilities of Audio-Visual Large Language Models (AV-LLMs) and Omni-modal Language Models (OLMs). AURA includes questions across six challenging cognitive domains, such as causality, timbre and pitch, tempo and AV synchronization, unanswerability, implicit distractions, and skill profiling, explicitly designed to be unanswerable from a single modality. This forces models to construct a valid logical path grounded in both audio and video, setting AURA apart from AV datasets that allow uni-modal shortcuts. To assess reasoning traces, we propose a novel metric, AuraScore, which addresses the lack of robust tools for evaluating reasoning fidelity. It decomposes reasoning into two aspects: (i) Factual Consistency - whether reasoning is grounded in perceptual evidence, and (ii) Core Inference - the logical validity of each reasoning step. Evaluations of SOTA models on AURA reveal a critical reasoning gap: although models achieve high accuracy (up to 92% on some tasks), their Factual Consistency and Core Inference scores fall below 45%. This discrepancy highlights that models often arrive at correct answers through flawed logic, underscoring the need for our benchmark and paving the way for more robust multimodal evaluation.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 5 canonical work pages

  1. [5]

    EgoAdapt: Adaptive Multisensory Distillation and Policy Learning for Efficient Egocentric Perception

    Zero-Shot Video Question Answering with Procedu- ral Programs. Chowdhury, S.; Biswas, S.; Nag, S.; Nagarajan, T.; Mur- dock, C.; Ananthabhotla, I.; Qian, Y .; Ithapu, V . K.; Manocha, D.; and Gao, R. 2025a. EgoAdapt: Adaptive Mul- tisensory Distillation and Policy Learning for Efficient Ego- centric Perception. arXiv preprint arXiv:2506.21080. Chowdhury, ...

    work page Pith review arXiv 2021

  2. [6]

    Girdhar, R.; El-Nouby, A.; Liu, Z.; Singh, M.; Alwala, K

    Exploring the frontier of vision-language models: A survey of current methodologies and future directions.arXiv preprint arXiv:2404.07214. Girdhar, R.; El-Nouby, A.; Liu, Z.; Singh, M.; Alwala, K. V .; Joulin, A.; and Misra, I. 2023. Imagebind: One embedding space to bind them all. InProceedings of the IEEE/CVF con- ference on computer vision and pattern ...

    arXiv 2023

  3. [9]

    arXiv preprint arXiv:2011.11760

    Multimodal pretraining for dense video captioning. arXiv preprint arXiv:2011.11760. Huang, L.; Yu, W.; Ma, W.; Zhong, W.; Feng, Z.; Wang, H.; Chen, Q.; Peng, W.; Feng, X.; Qin, B.; et al. 2023. A sur- vey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. arXiv preprint arXiv:2311.05232. Huang, Y .; and Huang,...

    arXiv 2011

  4. [10]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 13700–13710

    Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 13700–13710. Karpukhin, V .; O˘guz, B.; Min, S.; Lewis, P.; Wu, L.; Edunov, S.; Chen, D.; and Yih, W.-t. 2020. Dense Passage Retrieval for Open-Domain Question An...

    arXiv 2020

  5. [11]

    arXiv preprint arXiv:2310.02255

    MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts. arXiv preprint arXiv:2310.02255. Lu, P.; Xia, T.; Shi, W.; Kholy, A. E.; Lin, X. V .; Gao, J.; Chen, X.; and Chang, K.-W. 2022b. Learn to Explain: Mul- timodal Reasoning via Thought Chains for Science Question Answering. In Advances in Neural Information Processing Syste...

    arXiv 2023

  6. [12]

    arXiv preprint arXiv:2306.05424

    Video-chatgpt: Towards detailed video understand- ing via large vision and language models. arXiv preprint arXiv:2306.05424. Mangalam, K.; Akshulakov, R.; and Malik, J. 2023. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Infor- mation Processing Systems, 36: 46212–46244. Marino, K.; Rastegari, M.; Fa...

    arXiv 2023

  7. [14]

    In Proceedings of the IEEE/CVF inter- national conference on computer vision, 1686–1697

    Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF inter- national conference on computer vision, 1686–1697. Yang, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Li, C.; Liu, D.; Huang, F.; Wei, H.; et al. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115. Yang, P.; Wang, X.; Duan, ...

    arXiv 2024

  8. [15]

    UniMD: Towards Unifying Moment Retrieval and Temporal Action Detection

    Adaptive edge-aware semantic interaction network for salient object detection in optical remote sensing images. IEEE Transactions on Geoscience and Remote Sensing. Zeng, Y .; Zhong, Y .; Feng, C.; and Ma, L. 2024. UniMD: Towards Unifying Moment Retrieval and Temporal Action Detection. arXiv preprint arXiv:2404.04933. Zhai, X.; Mustafa, B.; Kolesnikov, A.;...

    work page Pith review arXiv 2024

Show all 17 references
  1. [16]

    arXiv preprint arXiv:2203.01225

    Video question answering: Datasets, algorithms and challenges. arXiv preprint arXiv:2203.01225. Zhou, J.; Shu, Y .; Zhao, B.; Wu, B.; Xiao, S.; Yang, X.; Xiong, Y .; Zhang, B.; Huang, T.; and Liu, Z. 2024. Mlvu: A comprehensive benchmark for multi-task long video under- standi...

    2024 arXiv

  2. [17]

    In Inter- national Conference on Learning Representations (ICLR)

    MiniGPT-4: Enhancing Vision-Language Under- standing with Advanced Large Language Models. In Inter- national Conference on Learning Representations (ICLR). Zhuang, J.; Hu, J.; Mu, L.; Hu, R.; Liang, X.; Ye, J.; and Hu, H. 2024. Falip: Visual prompt as foveal attention boosts c...

    2024

  3. [2019]

    In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 3195–3204

    OK-VQA: A Visual Question Answering Benchmark Requiring External Knowledge. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 3195–3204. Masry, A.; Long, D. X.; Tan, J. Q.; Joty, S.; and Hoque, E. 2022. ChartQA: A Benchmark for Question Answeri...

    2022 arXiv

  4. [2020]

    In Proceedings of the 37th International Confer- ence on Machine Learning (ICML), 3929–3938

    REALM: Retrieval-Augmented Language Model Pre- Training. In Proceedings of the 37th International Confer- ence on Machine Learning (ICML), 3929–3938. PMLR. Gygli, M.; Grabner, H.; Riemenschneider, H.; and Van Gool, L. 2014. Creating summaries from user videos. InComputer Visio...

    2014 arXiv

  5. [2021]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 11287–11297

    Agqa: A benchmark for compositional spatio- temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 11287–11297. Guo, R.; Ying, X.; Chen, Y .; Niu, D.; Li, G.; Qu, L.; Qi, Y .; Zhou, J.; Xing, B.; Yue, W.; et al. 2025. Audio-v...

    2025

  6. [2022]

    In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), 5558–5570

    MuRAG: Multimodal Retrieval-Augmented Genera- tor for Open Question Answering over Images and Text. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), 5558–5570. As- sociation for Computational Linguistics. Chen, X.; Lin, Y .; Zh...

    2022 arXiv

  7. [2023]

    See https://vicuna

    Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality. See https://vicuna. lmsys. org (accessed 14 April 2023), 2(3): 6. Choudhury, R.; Niinuma, K.; Kitani, K. M.; and Jeni, L. A

    2023

  8. [2024]

    arXiv preprint arXiv:2406.07476

    Videollama 2: Advancing spatial-temporal model- ing and audio understanding in video-llms. arXiv preprint arXiv:2406.07476. Chiang, W.-L.; Li, Z.; Lin, Z.; Sheng, Y .; Wu, Z.; Zhang, H.; Zheng, L.; Zhuang, S.; Zhuang, Y .; Gonzalez, J. E.; et al

    arXiv

  9. [2025]

    InProceedings of the Com- puter Vision and Pattern Recognition Conference , 24817– 24826

    Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documents. InProceedings of the Com- puter Vision and Pattern Recognition Conference , 24817– 24826. Chen, J.; Zhu, D.; Shen, X.; Li, X.; Liu, Z.; Zhang, P.; Kr- ishnamoorthi, R.; Chandra, V .; Xiong, Y .; and E...

    2025 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.