pith. machine review for the scientific record. sign in

arxiv: 2604.14888 · v2 · submitted 2026-04-16 · 💻 cs.CL · cs.AI· cs.CV· cs.LG

Recognition: unknown

Reasoning Dynamics and the Limits of Monitoring Modality Reliance in Vision-Language Models

Danae S\'anchez Villegas , Samuel Lewis-Lim , Nikolaos Aletras , Desmond Elliott
Authors on Pith no claims yet

Pith reviewed 2026-05-10 10:55 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.CVcs.LG
keywords vision-language modelschain-of-thought reasoningmodality relianceanswer inertiamultimodal reasoningreasoning dynamicsmodel transparency
0
0 comments X

The pith

Vision-language models reinforce early predictions during reasoning and stay influenced by misleading text even when images suffice.

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

The paper tracks confidence across Chain-of-Thought steps in 18 VLMs to measure how reasoning integrates visual and textual input. It shows models exhibit answer inertia that locks in early commitments instead of revising them with later evidence. Reasoning-trained models correct more than instruction-tuned ones, but the improvement varies sharply between text-dominant and vision-only conditions. Interventions with misleading textual cues reveal that text influence persists despite adequate visual evidence, and CoT traces only partially expose this reliance depending on model type and monitoring method.

Core claim

Models are prone to answer inertia, in which early commitments to a prediction are reinforced, rather than revised during reasoning steps. While reasoning-trained models show stronger corrective behavior, their gains depend on modality conditions, from text-dominant to vision-only settings. Using controlled interventions with misleading textual cues, models are consistently influenced by these cues even when visual evidence is sufficient, and assess whether this influence is recoverable from CoT. Although this influence can appear in the CoT, its detectability varies across models and depends on what is being monitored. Reasoning-trained models are more likely to explicitly refer to the cues

What carries the argument

Answer inertia during Chain-of-Thought steps, measured through controlled interventions that insert misleading textual cues to isolate reliance on text versus vision.

Load-bearing premise

Controlled interventions with misleading textual cues accurately isolate modality reliance without introducing artifacts that change how the models normally reason.

What would settle it

Track whether models revise an initial wrong prediction when contradictory visual evidence is supplied after the first reasoning step; consistent failure to revise would support the inertia claim.

Figures

Figures reproduced from arXiv: 2604.14888 by Danae S\'anchez Villegas, Desmond Elliott, Nikolaos Aletras, Samuel Lewis-Lim.

Figure 1
Figure 1. Figure 1: Overview of reasoning dynamics and modality reliance in VLMs. In Section [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representative examples from datasets used in our evaluation (illustrative content; [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Answer probability trajectories over reasoning steps on MathVerse for instruction [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy as a function of CoT truncation depth on MathVerse and ScienceQA for [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Net gain of CoT reasoning on MathVerse across problem versions. Net gain [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Total effect (TE) by hint type and model. TE measures the in￾crease in hint-following behavior at￾tributable to the intervention. Error bars show 95% bootstrap confidence intervals. VLMs consistently follow textual hints. Fig￾ure 6 shows the aggregated total effect of each hint type across models. All except for the rewardhack framing variant consistently pro￾duced an effect, confirming that the injected t… view at source ↗
Figure 7
Figure 7. Figure 7: Monitorability scores (G 2 mean) for hint influence detection (blue) and modality attribution (orange) on MathVerse Vision-Only, shown for the unethical and professor sycophancy hints. Results on all hints can be found in Appendix C. hint, it is often interpreted as legitimate visual reasoning. As a result, longer and more fluent CoTs effectively provide cover for textual over-reliance. Effect of Hint Moda… view at source ↗
Figure 8
Figure 8. Figure 8: Monitorability (G 2 mean) for reward hacking hints on QVL3-32B￾Thinking (text vs. image). × = below intervention causal threshold. effect on the model when placed in the image. Implications for monitoring CoTs in multimodal LLMs. These results suggest that monitorability of modality reliance depends both on the model behav￾ior and on the information available to the monitor. A monitor that knows the form o… view at source ↗
Figure 9
Figure 9. Figure 9: Effect of increased decoding temperature on reasoning dynamics for InternVL3.5- [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Answer probability trajectories over reasoning steps on MathVerse (MV), PhyX [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Answer probability trajectories over reasoning steps on MathVerse (MV), PhyX [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Accuracy versus truncation depth of CoT reasoning on MathVerse and ScienceQA [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Net gain of CoT reasoning in PhyX across problem versions. Net gain measures [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Commitment dynamics during reasoning. Cumulative proportion of examples that have reached a stable commitment to their final predicted answer by each reasoning step. Step 0 denotes the prediction before CoT generation. Results are aggregated across model sizes and shown by model family, training paradigm (instruct:I, thinking:R), and final correctness. Across models and datasets, many predictions are comm… view at source ↗
Figure 15
Figure 15. Figure 15: Paired example. Both models incorrectly assume that angles 1 and 2 form a linear [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Example of InternVL instruction-tuned models. Both models initially predict [PITH_FULL_IMAGE:figures/full_fig_p020_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative example from Qwen2.5-VL-7B. The hint monitor classifies the hint [PITH_FULL_IMAGE:figures/full_fig_p021_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Qualitative example from Qwen3-VL-32B-Thinking with the professor syco [PITH_FULL_IMAGE:figures/full_fig_p022_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Monitorability scores (G 2 mean) for hint influence detection (blue) and modality attribution (orange) on MathVerse Vision-Only, shown for the rewardhack and user syco￾phancy hints. Chat Template Formatting and Monitorability [PITH_FULL_IMAGE:figures/full_fig_p023_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Monitorability score (G 2 mean) comparison for the hint monitor when the chat template is used for the prompt and when it isn’t. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Monitorability scores (G 2 mean) for hint influence detection (blue) and modality attribution (red) on MathVerse Vision-Only, shown for the rewardhack and user sycophancy hints. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Monitorability scores (G 2 mean) for hint influence detection (blue) and modality attribution (red) on MathVerse Vision-Only, shown for differnt rewardhack variants. The commitdiff hint clearly references reward answer 26 [PITH_FULL_IMAGE:figures/full_fig_p026_22.png] view at source ↗
read the original abstract

Recent advances in vision language models (VLMs) offer reasoning capabilities, yet how these unfold and integrate visual and textual information remains unclear. We analyze reasoning dynamics in 18 VLMs covering instruction-tuned and reasoning-trained models from two different model families. We track confidence over Chain-of-Thought (CoT), measure the corrective effect of reasoning, and evaluate the contribution of intermediate reasoning steps. We find that models are prone to answer inertia, in which early commitments to a prediction are reinforced, rather than revised during reasoning steps. While reasoning-trained models show stronger corrective behavior, their gains depend on modality conditions, from text-dominant to vision-only settings. Using controlled interventions with misleading textual cues, we show that models are consistently influenced by these cues even when visual evidence is sufficient, and assess whether this influence is recoverable from CoT. Although this influence can appear in the CoT, its detectability varies across models and depends on what is being monitored. Reasoning-trained models are more likely to explicitly refer to the cues, but their longer and fluent CoTs can still appear visually grounded while actually following textual cues, obscuring modality reliance. In contrast, instruction-tuned models refer to the cues less explicitly, but their shorter traces reveal inconsistencies with the visual input. Taken together, these findings indicate that CoT provides only a partial view of how different modalities drive VLM decisions, with important implications for the transparency and safety of multimodal systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper examines reasoning dynamics across 18 VLMs (instruction-tuned and reasoning-trained from two families). It tracks confidence trajectories over CoT steps, quantifies corrective effects of reasoning, and deploys controlled interventions inserting misleading textual cues to test whether models remain influenced by text even when visual evidence suffices. Core claims are that models exhibit answer inertia (early predictions reinforced rather than revised), that textual cues exert persistent influence, and that CoT monitoring recovers modality reliance only partially—reasoning-trained models explicitly reference cues yet produce fluent, apparently grounded traces, while instruction-tuned models show shorter but inconsistent traces.

Significance. If the empirical patterns hold under scrutiny, the work is significant for multimodal interpretability and safety research. It provides large-scale evidence that CoT is an incomplete window into modality reliance, with direct consequences for deploying VLMs in settings where undetected textual bias could produce errors. The comparison between model training regimes and the focus on both inertia and cue detectability add useful granularity to the literature on VLM reasoning.

major comments (2)
  1. [§4 and §5.2] §4 (Intervention Design) and §5.2 (Cue Influence Results): The central claim that misleading textual cues isolate modality reliance (and thereby demonstrate limits of CoT monitoring) rests on the assumption that the interventions do not themselves alter normal reasoning dynamics. Inserting misleading text may increase text attention, lengthen or restructure CoT, or induce a different inference regime not present in standard VLM prompting. Without reported controls comparing CoT statistics (length, token distribution, explicit cue mentions) in neutral vs. cued conditions, the observed inertia and cue influence risk being confounded with the manipulation, weakening the inference that CoT monitoring is inherently limited in real deployments.
  2. [§5.3] §5.3 (Reasoning-Trained vs. Instruction-Tuned Comparison): The claim that reasoning-trained models produce longer, fluent CoTs that obscure cue following while instruction-tuned models reveal inconsistencies requires quantitative backing. The manuscript should report effect sizes, statistical tests, and inter-annotator agreement for any manual CoT analysis, plus ablation on monitoring methods (keyword vs. semantic) to support the differential detectability conclusion.
minor comments (2)
  1. [Abstract and §3] Abstract and §3: Clarify the exact metrics used to track 'confidence over CoT' and 'corrective effect' (e.g., probability of final answer at each step, or external judge scores).
  2. [Results tables] Table 2 or equivalent results table: Include standard errors or confidence intervals alongside reported percentages for cue influence and inertia rates to allow assessment of variability across the 18 models.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the robustness of our experimental claims. We address each major point below and will incorporate revisions to provide additional controls and quantitative analyses where appropriate. These changes will strengthen the evidence that CoT monitoring offers only a partial view of modality reliance without altering our core findings on answer inertia and cue influence.

read point-by-point responses

  1. Referee: [§4 and §5.2] The central claim that misleading textual cues isolate modality reliance rests on the assumption that the interventions do not themselves alter normal reasoning dynamics. Inserting misleading text may increase text attention, lengthen or restructure CoT, or induce a different inference regime. Without reported controls comparing CoT statistics (length, token distribution, explicit cue mentions) in neutral vs. cued conditions, the observed inertia and cue influence risk being confounded with the manipulation.

    Authors: We agree that explicit controls are valuable to rule out confounds from the intervention itself. In the revised manuscript, we will add a new subsection reporting CoT length, token distribution statistics, and rates of explicit cue mentions across neutral and cued conditions for representative models. Preliminary checks in our data indicate that cue insertion does not systematically lengthen traces or shift token distributions beyond the expected addition of cue-related tokens, but we will quantify this fully. This will support that the observed answer inertia and persistent cue influence reflect genuine modality reliance patterns rather than artifacts of the manipulation. revision: yes

  2. Referee: [§5.3] The claim that reasoning-trained models produce longer, fluent CoTs that obscure cue following while instruction-tuned models reveal inconsistencies requires quantitative backing. The manuscript should report effect sizes, statistical tests, and inter-annotator agreement for any manual CoT analysis, plus ablation on monitoring methods (keyword vs. semantic) to support the differential detectability conclusion.

    Authors: We accept that the differential detectability claims would benefit from stronger quantification. In the revision, we will add effect sizes (Cohen's d) and statistical tests (paired t-tests or Wilcoxon tests) for CoT length and fluency differences between reasoning-trained and instruction-tuned models. For the manual analysis of cue references and inconsistencies, we will report inter-annotator agreement (Cohen's kappa) based on dual annotation of a subset of traces. We will also include an ablation comparing keyword-based cue detection against semantic similarity (using sentence embeddings) to quantify how monitoring method affects recoverability of cue influence. These additions will provide rigorous support for the observed patterns without changing the qualitative conclusions. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical measurements of VLM reasoning

full rationale

The paper conducts controlled experiments on 18 VLMs, tracking confidence over CoT steps, applying misleading textual cue interventions, and measuring corrective effects and modality influence. No equations, derivations, fitted parameters renamed as predictions, or self-citation chains appear in the load-bearing claims. All results follow directly from the described interventions and observations without reducing to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The study relies on standard assumptions in VLM evaluation without introducing new free parameters or invented entities.

axioms (1)
  • domain assumption Chain-of-Thought outputs provide observable traces of internal reasoning processes that can be used to infer modality reliance
    Central to interpreting whether influence is recoverable from CoT

pith-pipeline@v0.9.0 · 5574 in / 1015 out tokens · 28309 ms · 2026-05-10T10:55:17.265126+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 21 canonical work pages · 5 internal anchors

  1. [1]

    Evaluating vision-language models as evaluators in path planning

    Mohamed Aghzal, Xiang Yue, Erion Plaku, and Ziyu Yao. Evaluating vision-language models as evaluators in path planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 6886--6897, 2025

    2025

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    A closer look at bias and chain-of-thought faithfulness of large (vision) language models

    Sriram Balasubramanian, Samyadeep Basu, and Soheil Feizi. A closer look at bias and chain-of-thought faithfulness of large (vision) language models. In Findings of the Association for Computational Linguistics: EMNLP 2025, pp.\ 13406--13439, Suzhou, China, November 2025. Association for Computational Linguistics. ISBN 979-8-89176-335-7. doi:10.18653/v1/20...

    work page doi:10.18653/v1/2025.findings-emnlp.723 2025

  4. [4]

    Quantifying and mitigating unimodal biases in multimodal large language models: A causal perspective

    Meiqi Chen, Yixin Cao, Yan Zhang, and Chaochao Lu. Quantifying and mitigating unimodal biases in multimodal large language models: A causal perspective. In Findings of the Association for Computational Linguistics: EMNLP 2024, pp.\ 16449--16469, Miami, Florida, USA, November 2024. Association for Computational Linguistics. doi:10.18653/v1/2024.findings-em...

    work page doi:10.18653/v1/2024.findings-emnlp.960 2024

  5. [5]

    Reasoning Models Don't Always Say What They Think

    Yanda Chen, Joe Benton, Ansh Radhakrishnan, Jonathan Uesato, Carson Denison, John Schulman, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, et al. Reasoning models don't always say what they think. arXiv preprint arXiv:2505.05410, 2025

    work page internal anchor Pith review arXiv 2025

  6. [6]

    Comt: A novel benchmark for chain of multi-modal thought on large vision-language models

    Zihui Cheng, Qiguang Chen, Jin Zhang, Hao Fei, Xiaocheng Feng, Wanxiang Che, Min Li, and Libo Qin. Comt: A novel benchmark for chain of multi-modal thought on large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 23678--23686, 2025

    2025

  7. [7]

    A re D eepseek R 1 A nd O ther R easoning M odels M ore F aithful? In ICLR 2025 Workshop on Foundation Models in the Wild, 2025

    James Chua and Owain Evans. A re D eepseek R 1 A nd O ther R easoning M odels M ore F aithful? In ICLR 2025 Workshop on Foundation Models in the Wild, 2025. URL https://openreview.net/forum?id=rI38nonvF5

    2025

  8. [8]

    arXiv preprint arXiv:2504.05294 , year=

    Pedro Ferreira, Wilker Aziz, and Ivan Titov. Truthful or fabricated? using causal attribution to mitigate reward hacking in explanations, 2025. URL https://arxiv.org/abs/2504.05294

    work page arXiv 2025

  9. [9]

    Guan, Miles Wang, Micah Carroll, Zehao Dou, Annie Y

    Melody Y Guan, Miles Wang, Micah Carroll, Zehao Dou, Annie Y Wei, Marcus Williams, Benjamin Arnav, Joost Huizinga, Ian Kivlichan, Mia Glaese, et al. Monitoring monitorability. arXiv preprint arXiv:2512.18311, 2025

    work page arXiv 2025

  10. [10]

    Zhitao He, Sandeep Polisetty, Zhiyuan Fan, Yuchen Huang, Shujin Wu, and Yi R. Fung. MMB oundary: Advancing MLLM knowledge boundary awareness through reasoning step confidence calibration. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 16427--16444, Vienna, Austria, July 2025. Associ...

    work page doi:10.18653/v1/2025.acl-long.802 2025

  11. [11]

    To trust or not to trust? enhancing large language models' situated faithfulness to external contexts

    Yukun Huang, Sanxing Chen, Hongyi Cai, and Bhuwan Dhingra. To trust or not to trust? enhancing large language models' situated faithfulness to external contexts. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=K2jOacHUlO

    2025

  12. [12]

    MME - C o T : B enchmarking C hain-of- T hought in L arge M ultimodal M odels for R easoning Q uality, R obustness, and E fficiency

    Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, Bo Zhang, Chaoyou Fu, Peng Gao, and Hongsheng Li. MME - C o T : B enchmarking C hain-of- T hought in L arge M ultimodal M odels for R easoning Q uality, R obustness, and E fficiency. In Forty-second International Conference on Machine Lear...

    2025

  13. [13]

    Chatbug: A c ommon v ulnerability of a ligned llm s i nduced by c hat t emplates

    Fengqing Jiang, Zhangchen Xu, Luyao Niu, Bill Yuchen Lin, and Radha Poovendran. Chatbug: A c ommon v ulnerability of a ligned llm s i nduced by c hat t emplates. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 27347--27355, 2025 b

    2025

  14. [14]

    Measuring Faithfulness in Chain-of-Thought Reasoning

    Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, Kamilė Lukošiūtė, Karina Nguyen, Newton Cheng, Nicholas Joseph, Nicholas Schiefer, Oliver Rausch, Robin Larson, Sam McCandlish, Sandipan Kundu, Saurav Kadavath, Shannon Yang, Thomas Henighan, Timothy Maxwel...

    work page Pith review arXiv 2023

  15. [15]

    Samuel Lewis-Lim, Xingwei Tan, Zhixue Zhao, and Nikolaos Aletras. Analysing chain of thought dynamics: Active guidance or unfaithful post-hoc rationalisation? In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp.\ 29838--29853, Suzhou, China, November 2025. Association for Computational Linguistics. ISBN 979-8-8917...

    work page doi:10.18653/v1/2025.emnlp-main.1516 2025

  16. [16]

    arXiv:2504.04974 [cs.CV] https://arxiv.org/abs/2504.04974

    Ming Li, Ruiyi Zhang, Jian Chen, Chenguang Wang, Jiuxiang Gu, Yufan Zhou, Franck Dernoncourt, Wanrong Zhu, Tianyi Zhou, and Tong Sun. Towards visual text grounding of multimodal large language model. arXiv preprint arXiv:2504.04974, 2025

    work page arXiv 2025

  17. [17]

    Learn to explain: Multimodal reasoning via thought chains for science question answering

    Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35: 0 2507--2521, 2022

    2022

  18. [18]

    Keeping LLM s aligned after fine-tuning: The crucial role of prompt templates

    Kaifeng Lyu, Haoyu Zhao, Xinran Gu, Dingli Yu, Anirudh Goyal, and Sanjeev Arora. Keeping LLM s aligned after fine-tuning: The crucial role of prompt templates. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=xNlQjS0dtO

    2024

  19. [19]

    Are self-explanations from large language models faithful? In Findings of the Association for Computational Linguistics: ACL 2024, pp.\ 295--337, Bangkok, Thailand, August 2024

    Andreas Madsen, Sarath Chandar, and Siva Reddy. Are self-explanations from large language models faithful? In Findings of the Association for Computational Linguistics: ACL 2024, pp.\ 295--337, Bangkok, Thailand, August 2024. Association for Computational Linguistics. doi:10.18653/v1/2024.findings-acl.19. URL https://aclanthology.org/2024.findings-acl.19/

    work page doi:10.18653/v1/2024.findings-acl.19 2024

  20. [20]

    Deepseek-r1 thoughtology: Let s think about LLM reasoning

    Sara Vera Marjanovic, Arkil Patel, Vaibhav Adlakha, Milad Aghajohari, Parishad BehnamGhader, Mehar Bhatia, Aditi Khandelwal, Austin Kraft, Benno Krojer, Xing Han L \`u , Nicholas Meade, Dongchan Shin, Amirhossein Kazemnejad, Gaurav Kamath, Marius Mosbach, Karolina Stanczak, and Siva Reddy. Deepseek-r1 thoughtology: Let s think about LLM reasoning. Transac...

    2026

  21. [21]

    MM - SHAP : A performance-agnostic metric for measuring multimodal contributions in vision and language models & tasks

    Letitia Parcalabescu and Anette Frank. MM - SHAP : A performance-agnostic metric for measuring multimodal contributions in vision and language models & tasks. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 4032--4059, Toronto, Canada, July 2023. Association for Computational Linguis...

    work page doi:10.18653/v1/2023.acl-long.223 2023

  22. [22]

    On measuring faithfulness or self-consistency of natural language explanations

    Letitia Parcalabescu and Anette Frank. On measuring faithfulness or self-consistency of natural language explanations. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 6048--6089, Bangkok, Thailand, August 2024. Association for Computational Linguistics. doi:10.18653/v1/2024.acl-long....

    work page doi:10.18653/v1/2024.acl-long.329 2024

  23. [23]

    Do vision & language decoders use images and text equally? how self-consistent are their explanations? In The Thirteenth International Conference on Learning Representations, 2025

    Letitia Parcalabescu and Anette Frank. Do vision & language decoders use images and text equally? how self-consistent are their explanations? In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=lCasyP21Bf

    2025

  24. [24]

    Phyx: Does your model have the” wits” for physical reasoning?arXiv preprint arXiv:2505.15929, 2025

    Hui Shen, Taiqiang Wu, Qi Han, Yunta Hsieh, Jizhou Wang, Yuyue Zhang, Yuxin Cheng, Zijian Hao, Yuansheng Ni, Xin Wang, et al. Phyx: Does your model have the" wits" for physical reasoning? arXiv preprint arXiv:2505.15929, 2025

    work page arXiv 2025

  25. [25]

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting

    Miles Turpin, Julian Michael, Ethan Perez, and Samuel Bowman. Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting. Advances in Neural Information Processing Systems, 36: 0 74952--74965, 2023

    2023

  26. [26]

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265, 2025 a

    work page internal anchor Pith review arXiv 2025

  27. [27]

    Multimodal chain-of-thought reasoning: A comprehensive survey.arXiv preprint arXiv:2503.12605, 2025

    Yaoting Wang, Shengqiong Wu, Yuecheng Zhang, Shuicheng Yan, Ziwei Liu, Jiebo Luo, and Hao Fei. Multimodal chain-of-thought reasoning: A comprehensive survey. arXiv preprint arXiv:2503.12605, 2025 b

    work page arXiv 2025

  28. [28]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 0 24824--24837, 2022

    2022

  29. [29]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, pp.\ 169--186

    Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, pp.\ 169--186. Springer, 2024

    2024

  31. [31]

    Evaluating and steering modality preferences in multimodal large language model

    Yu Zhang, Jinlong Ma, Yongshuai Hou, Xuefeng Bai, Kehai Chen, Yang Xiang, Jun Yu, and Min Zhang. Evaluating and steering modality preferences in multimodal large language model. arXiv preprint arXiv:2505.20977, 2025

    work page arXiv 2025

  32. [32]

    Pan, and Hanjie Chen

    Xiongtao Zhou, Jie He, Lanyu Chen, Jingyu Li, Haojing Chen, Victor Gutierrez Basulto, Jeff Z. Pan, and Hanjie Chen. M i CE val: Unveiling multimodal chain of thought ' s quality via image description and reasoning steps. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Lan...

    work page doi:10.18653/v1/2025.naacl-long.504 2025

  33. [33]

    InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025

    work page internal anchor Pith review arXiv 2025

  34. [34]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  35. [35]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  36. [36]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  37. [37]

    B6 '& &G ѡUtf! -Ff92

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

    work page arXiv 1999