Pith. sign in

REVIEW 3 major objections 4 minor 39 references

ReGround: Restoring Visual Grounding in Multi-Step Reasoning through Self-Diagnosis and Visual Re-Examination

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper claims that visual re-examination only improves multi-step VLM reasoning when paired with targeted self-diagnosis, and that diagnostic quality, not re-injection alone, is the key variable.

desk verdict Solid empirical paper with a genuinely new cue-quality ablation; the central claim is plausible and worth building on, though the oracle/generic comparison could be tightened. read the letter →

arxiv 2608.04385 v1 pith:QNZVLCY2 submitted 2026-08-05 cs.CV

Lei Peng , Shuai Lv , Wei Hu This is my paper · ORCID
classification cs.CV
keywords vision-languagemodelsvisualgroundingmulti-stepreasoningself-diagnosisre-injectionattentionentropyself-correctioncapabilitybootstrapping
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

Vision-language models (VLMs) progressively lose attention to image evidence as they reason over longer chains, a pattern the paper calls visual grounding decay. The paper argues that simply re-injecting the image for a second look is not enough: without a targeted cue saying what to re-examine, re-checking can even override correct answers. It introduces ReGround, a two-stage training recipe that teaches the model to diagnose its own grounding failures and then revisit the image with a specific cue, without external tools. If the claim holds, the practical lesson is that self-correction in VLMs should be diagnosis-driven rather than format-driven, and that a stronger model's diagnostic skill can be distilled into a smaller one that keeps most of the benefit at inference time.

What carries the argument

The load-bearing device is a two-stage training pipeline: a re-examination trigger routes each sample to one of four trajectory modes—corrective, grounding, verification, or no re-grounding—and a visual self-diagnosis module writes a first-person <reground> cue that says what to look at without revealing the answer. In Stage 1, a stronger same-family model provides those cues only during data construction (capability bootstrapping), and the policy learns to produce them autonomously; in Stage 2, GRPO with an asymmetric reward calibrates when to trigger re-examination, rewarding successful recovery, efficient precision, and attempted correction, while penalizing blind failure.

What would settle it

A human-annotation study of the Step 2 grounding judge: if human raters disagree with the judge's pass/fail decisions on a few hundred correct first-round answers, the data-construction premise is in doubt. Alternatively, if a model from the same family with similar visual ability but weaker instruction-following generates cues that recover the same 79–90%, the recovery could be explained by general capability rather than diagnostic skill.

Watch

Extended reading notes

Core claim

ReGround claims that effective visual self-correction in multi-step reasoning needs two ingredients: re-injecting the image after reasoning has started, and a targeted self-diagnosis cue that tells the model what to re-check. The central evidence is a cue-quality ablation in which a generic cue ('look again carefully') performs at or below no re-examination on several benchmarks, while oracle cues from a larger same-family model give large gains, and the policy model's self-generated cues recover 79–90% of those gains across four benchmarks. The paper further shows that a balanced triggering policy matters: removing negative examples causes the trigger rate to jump to about 85% and hurts accuracy, while over-triggering produces negative flips. Across 2,510 samples, attention entropy over image tokens narrows in 98.0% of cases and rebroadens after re-injection in 94.4%, which the authors treat as a mechanism signature rather than a direct grounding metric.

Load-bearing premise

The approach assumes the stronger model's diagnostic cues and the judge that flags 'correct but ungrounded' answers are genuinely identifying visual grounding failures—an assumption the paper states it has not yet checked against human annotations.

Editorial extensions

If this is right

  • A model trained with ReGround can recover 79–90% of the accuracy gain that oracle diagnostic cues provide, without needing the larger model at inference time.
  • Generic re-examination can actively hurt: it drops accuracy on some benchmarks below the no-re-examination baseline, so any re-injection mechanism should be coupled with diagnostic guidance.
  • Both image re-injection and targeted diagnosis are necessary; text-only reflection with the same training recipe yields much smaller gains, and even text-only reflection plus RL trails visual re-injection without RL.
  • A calibrated trigger rate (about 45% on the studied setting) balances positive and negative flips; over-triggering degrades accuracy while conservative triggering forgoes recoverable gains.
  • The recipe transfers to a second backbone and improves accuracy on all eight benchmarks, with gains concentrated on visually intensive multi-step reasoning tasks and at 1.76x latency relative to baseline.

Reading between the lines

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

  • The paper does not test this, but the attention-entropy dip before re-injection could serve as a cheap online signal for deciding when to trigger re-examination, since the paper already shows entropy correlates with the benefit of re-visiting.
  • A human-agreement study of the grounding judge would settle whether the 'ungrounded' label is itself reliable; the paper states this agreement has not yet been measured.
  • For tasks with fine-grained or spatially distributed evidence, the paper's own results suggest combining ReGround with region-level crop or zoom tools could close the remaining gap on benchmarks like VisuLogic and V*Bench.
  • A natural next experiment is to feed the same diagnostic cue into several models from the same family that share a vision encoder; if gains persist across the whole family, diagnostic-skill transfer is broader than the current two-point comparison shows.
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

3 major / 4 minor

Summary. The paper proposes ReGround, a two-stage, tool-free framework for improving visual grounding in multi-step VLM reasoning. Stage 1 bootstraps training trajectories using a stronger model to generate self-diagnostic cues and re-injects the image, then trains the policy with SFT; Stage 2 uses GRPO with an asymmetric reward to calibrate when re-examination should be triggered. The authors report gains on eight benchmarks across Qwen2.5-VL-7B and Qwen3-VL-8B, provide an attention-entropy analysis of grounding decay and recovery, and claim that diagnostic quality—not the re-examination format alone—is the key variable governing whether visual re-examination helps or hurts self-correction. The main supporting evidence is a cue-quality ablation (Table 5) showing oracle/self-generated targeted cues outperform a generic 'look again' cue, with self-generated cues recovering 79-90% of the oracle's gains.

Significance. If the central claim holds, ReGround is a useful contribution: it identifies a plausible mechanism (targeted self-diagnosis) underlying effective visual re-examination and offers a practical bootstrapping recipe that avoids architectural changes and external tools. The paper is well-structured, ships code, and includes several careful ablations (Tables 3-5, Table 9/10) plus a mechanism-level attention analysis (Table 11) that is explicitly framed as proxy evidence. The empirical documentation of attention entropy decay/recovery across 2,510 samples is a solid observational contribution. However, the headline mechanistic claim about 'diagnostic quality' is not yet cleanly established, and the exploration-preference parameter is tuned directly on the test benchmarks; both issues need to be addressed before the paper's main interpretation can be accepted.

major comments (3)
  1. [Table 5 / §4.3.3] The cue-quality ablation confounds three variables: whether the cue names any concrete re-examination target, whether that target is accurate, and how much answer-relevant information the cue leaks. The generic cue ('Please look at the image again carefully') is targetless, whereas oracle and self cues name specific regions and are prohibited from giving numeric answers but may still reveal the critical reasoning step (e.g., 're-check whether DE is parallel to BC' essentially states the key relation). Thus the 4.6-8.9 point oracle-vs-generic gap could be driven by the mere presence of a specific target—even a wrong one—rather than by diagnostic accuracy. Since the paper's central claim is that diagnostic quality, not the re-examination format, is the key variable, the experiment must include a control with targeted-but-incorrect cues of matched specificity and length, or an analysis measuring the accuracy of self-generated cues against an independent ground truth (e.g., region overlap with the oracle or human-annotated error regions). Without such a control, the reported 79-90% recovery rates do not cleanly measure learned diagnostic skill; they may measure the ability to emit any concrete cue.
  2. [Appendix C / Table 9] The exploration-preference (EP) parameter is swept directly on the four test benchmarks, and the 'Ours' results in Table 1 and the recovery rates in Table 5 use the EP value (+0.2) selected on these same test sets. This is a form of test-set tuning that can inflate the reported gains. The manuscript should select EP on a held-out validation split, or at minimum report the full EP sweep in the main text and demonstrate that the qualitative conclusions (ReGround improves over baselines) hold across a range of EP values rather than only at the best test-set value. The current presentation in Table 9 does show gains over baseline for several EP settings, but the headline numbers are in-sample selections.
  3. [§3.1.1 / Table 5] The Step 2 grounding judge (Qwen2.5-VL-72B-Instruct) is used to construct the Grounding trajectories and also serves as the oracle cue generator in Table 5, yet the manuscript states (in §3.1.1) that its agreement with human annotations remains future work. This is acceptable for the main method-vs-baseline comparison, but it weakens the interpretation of the oracle as a diagnostic upper bound: the oracle's advantage over generic cues may stem from the 72B model's general capability or from answer-hint leakage rather than from superior diagnosis. To support the capability-bootstrapping claim, the authors should provide an independent evaluation of self-generated cue quality (e.g., human-annotated relevance of the cue's focus region) or compare the self cues' focus regions against the oracle's focus regions on a per-sample basis.
minor comments (4)
  1. [Table 5 footnote] The statement that the recovery rate 'primarily reflects learned diagnostic reasoning rather than visual perception differences' is too strong, because the 7B and 72B models differ in language-model capacity and instruction following; please soften this claim or add a control cue from an equally capable but non-oracle source.
  2. [§4.2, Tables 1 and 5] No seed variance is reported for the main results or the ablations. Given the stochastic GRPO rollout (temperature 0.7) and the small EP-induced differences in some benchmarks, please report means and standard deviations across at least three seeds for the key tables.
  3. [Table 11, §4.4] The attention entropy and SAS analyses use a single layer (Layer 20); please state whether the decay/recovery pattern is stable across layers and heads, or at least acknowledge this limitation in the main text.
  4. [Abstract and §1] The phrase 'a swing of several points on key benchmarks' is vague; please quantify it with the actual range (e.g., the −2.1 to +8.6 spread reported in Table 5).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central claims rest on controlled ablations and external benchmarks, not on self-referential derivations.

full rationale

ReGround is an empirical method paper rather than a formal derivation, so there is no chain of equations whose output equals its input. The central claim that diagnostic quality, not re-examination format, drives gains is supported by Table 5, which holds trigger positions fixed and varies only the cue content; the generic-versus-targeted comparison is an independent manipulation rather than a fitted parameter renamed as a prediction. The oracle (72B) cues used in Table 5 are generated by the same model family that provides the capability-bootstrapping teacher for Stage-1 SFT data, so the reported 79–90% recovery rate is best understood as a distillation-transfer measurement rather than an independent upper-bound test. This overlap is a real interpretive limitation, but it does not make the stated gain definitional: the self-generated cues are produced at inference time by the 7B policy and scored on external benchmarks, and the generic-cue harm is demonstrated separately. The paper also explicitly discloses that the Step-2 grounding judge's agreement with human annotations is unmeasured and that the judge is not used during evaluation, which further separates data-construction choices from benchmark scoring. No load-bearing self-citations, imported uniqueness theorems, or ansatz-smuggling citations were found; references to prior work are used as background and comparison, not as the justification for the paper's own conclusions.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The framework rests on several unvalidated domain assumptions: the 72B judge's grounding decisions, the entropy signature as a proxy for grounding decay, and the oracle's status as a diagnostic upper bound. These are all flagged in the text as limitations or future work. No physical entities are invented; the <reground> tag and SAS are methodological constructs that are directly observable in the model's outputs.

free parameters (4)
  • Exploration-preference (EP) = +0.2
    Swept from -1.0 to +1.0 on MathVista, MathVerse, MathVision, and HallusionBench test sets (Table 9); the chosen value drives the reported main results and trigger rate.
  • Reward coefficients lambda_reg, lambda_acc, lambda_form = 0.5 / 0.7 / 0.01
    Given in Table 7 without a described held-out selection procedure; they shape the four-quadrant reward and thus the trigger behavior.
  • gamma and beta in accuracy reward = 0.14 / 0.14
    Selected hyperparameters that set the penalty for re-examination with correct answers; no tuning details are provided.
  • Stochastic trigger epsilon = 0.15
    Chosen for Stage 1 trajectory construction to create verification trajectories; it affects the ratio of trajectory modes.
assumptions (5)
  • domain assumption Attention entropy narrowing over image tokens is a valid proxy for visual grounding decay.
    Section 4.4; the paper states entropy measures attention breadth, not grounding quality, yet uses it as the benchmark-level signature.
  • domain assumption The Qwen2.5-VL-72B-Instruct judge correctly identifies visually ungrounded responses.
    Section 3.1.1 and Appendix B.2; human agreement explicitly not measured, and the judge shapes the training data.
  • domain assumption Gains on the four EP-swept benchmarks generalize to the other benchmarks and to Qwen3-VL without retuning.
    Section 4.2 and Appendix C; the EP choice is made on three of the four reported math and hallucination benchmarks and then applied to all.
  • domain assumption The 72B cue is a valid oracle for diagnostic skill, and the recovery rate reflects learned diagnostic reasoning.
    Section 4.3.3; the 72B shares only the vision encoder, not the language model, so the oracle advantage may include general capability.
  • domain assumption GRPO with the asymmetric reward preserves the SFT-acquired re-grounding behavior.
    Section 3.2.2; the partial reward intentionally sustains re-grounding during exploration, but the long-run stability is not proven.
invented entities (2)
  • <reground> control token independent evidence
    purpose: Marker the policy emits to trigger image re-injection and structured re-examination; it is the core mechanism of the framework.
    The token is observable in model outputs and its effect is measured on benchmarks; it is an engineered component, not a postulated physical entity.
  • Semantic Alignment Score (SAS) independent evidence
    purpose: Quantifies whether attention mass moves toward cue-referenced image regions after re-injection; used as proxy evidence for cue-guided grounding.
    SAS is a defined metric computed from model attention; it is a new analytical tool rather than a postulated physical entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReGround: Restoring Visual Grounding in Multi-Step Reasoning through Self-Diagnosis and Visual Re-Examination." pith.science (2026). https://pith.science/paper/QNZVLCY2

@misc{pith2026260804385,
  author       = {Pith},
  title        = {Pith review of: ReGround: Restoring Visual Grounding in Multi-Step Reasoning through Self-Diagnosis and Visual Re-Examination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QNZVLCY2}},
  note         = {Machine review of arXiv:2608.04385}
}
read the original abstract

Vision-Language Models (VLMs) often lose visual grounding during multi-step reasoning: as reasoning chains grow longer, later inference steps rely increasingly on language priors rather than image evidence. We identify a consistent benchmark-level signature associated with this degradation: across 2,510 re-examined samples from four benchmarks, attention entropy over image tokens typically decreases during Round 1 and rises again after image re-injection. However, we find that effective visual re-examination requires two complementary ingredients: image re-injection and targeted self-diagnosis. Without targeted diagnosis, re-examination can even hurt performance, whereas accurate self-diagnosis yields substantial gains -- a swing of several points on key benchmarks, indicating that diagnostic quality is a key factor in whether re-examination helps or hurts in our setting. We present ReGround, a two-stage framework that teaches VLMs to self-diagnose grounding failures and selectively re-examine visual evidence, without architectural modifications or external tools. Through capability bootstrapping, a stronger variant from the same model family provides diagnostic scaffolding only during data construction, while the policy model learns to diagnose autonomously at inference time and retains most of the assisted gains. Experiments on eight benchmarks across two VLM backbones demonstrate consistent gains, especially on visually intensive multi-step reasoning tasks, while incurring only modest inference overhead relative to tool-augmented baselines. Project page: https://sespoir.github.io/reground-page/ . Code: https://github.com/sespoir/ReGround .

Figures

Figures reproduced from arXiv: 2608.04385 by the authors.

Figure 1
Figure 1. Attention distribution narrows during multi-step reasoning and is redistributed after image re-injection. Qwen2.5- [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the ReGround framework. Trajectory Construction (left): Initial reasoning is evaluated by a sequential [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Representative trajectories for each trajectory mode. Correction (left): Round 1 misidentifies the arc relationship [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Cue-guided attention redirection on a MathVerse geometry problem (Layer 20, Qwen2.5-VL-7B). Left: R1 attention [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Training dynamics. (a) SFT loss over 2.5 epochs (checkpoint at 2.0). (b) GRPO reward fluctuates 0.2–0.8 (discrete [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Reasoning pipeline prompt templates executed by the policy model. (a) Round 1 produces initial observation, reasoning, [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Trigger-condition prompt templates. Left-top: Corrective cue (Step 1) identifies what to re-examine when the answer [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Generic vs. targeted diagnostic cue on a CLEVR-Math spatial reasoning problem. Both conditions share the same [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Verification example (MathVista #744). Round 1 correctly applies the Basic Proportionality Theorem. The diagnostic [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Correction example (MMStar #671). Round 1 confuses the motor cortex (a functional sub-area) with the frontal [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: No-ReGround examples. Top: MathVista #314 — bar chart reading: the model reads all four bar values from the chart [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 20 canonical work pages

  1. [1]

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

    arXiv doi:10.48550/arxiv.2511.21631 2025

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. 2025. Qwen2.5-VL Technical Rep...

    arXiv doi:10.48550/arxiv.2502.13923 2025

  3. [3]

    Hardy Chen, Haoqin Tu, Fali Wang, Hui Liu, Xianfeng Tang, Xinya Du, Yuyin Zhou, and Cihang Xie. 2025. SFT or RL? An Early Investigation into Training R1-Like Reasoning Large Vision-Language Models.Transactions on Machine Learning Research(2025)

    2025

  4. [4]

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. 2024. Are We on the Right Way for Evaluating Large Vision-Language Models?. InAdvances in Neural Information Processing Systems

    work page 2024

  5. [5]

    Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. 2024. An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models. InComputer Vision – ECCV 2024. Springer, 19–35. doi:10.1007/978-3-031-73004-7_2

    doi:10.1007/978-3-031-73004-7_2 2024

  6. [6]

    Xu Chu, Xinrong Chen, Guanyu Wang, Zhijie Tan, Kui Huang, Wenyu Lv, Tong Mo, and Weiping Li. 2025. Qwen Look Again: Guiding Vision-Language Rea- soning Models to Re-attention Visual Information. (2025). arXiv:2505.23558 doi:10.48550/arXiv.2505.23558

    arXiv doi:10.48550/arxiv.2505.23558 2025

  7. [7]

    Yihe Deng, Hritik Bansal, Fan Yin, Nanyun Peng, Wei Wang, and Kai-Wei Chang

  8. [8]

    Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Pramu- ditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. 2024. Multi-Modal Hallucination Control by Visual Information Grounding. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14303–14312

    work page 2024

Show all 39 references
  1. [9]

    Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. 2024. HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large ...

    2024

  2. [10]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. 2025. DeepSeek-R1 incen- tivizes reasoning in LLMs through reinforcement learning.Nature645, 8081 (2025), 633–638. doi:10.1038/s41586-025-09422-z

    2025 doi

  3. [11]

    Jack Hong, Chenxiao Zhao, ChengLin Zhu, Weiheng Lu, Guohai Xu, and Xing Yu. 2026. DeepEyesV2: Toward Agentic Multimodal Model. InThe Fourteenth International Conference on Learning Representations

    2026

  4. [12]

    Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2024. Large Language Models Cannot Self-Correct Reasoning Yet. InThe Twelfth International Conference on Learning Representations (ICLR)

    2024

  5. [13]

    Pu Jian, Junhong Wu, Wei Sun, Chen Wang, Shuo Ren, and Jiajun Zhang. 2025. Look Again, Think Slowly: Enhancing Visual Reflection in Vision-Language Models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. Association for Computational L...

    2025 doi

  6. [14]

    Omri Kaduri, Shai Bagon, and Tali Dekel. 2025. What’s in the Image? A Deep- Dive into the Vision of Vision Language Models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14549–14558

    2025

  7. [15]

    Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. 2025. See What You Are Told: Visual Attention Sink in Large Multimodal Models. InThe Thirteenth International Conference on Learning Representations (ICLR)

    2025

  8. [16]

    Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin

  9. [17]

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. 2024. MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts. InThe Twelfth International Conference on Lea...

    2024

  10. [18]

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023. Self-Refine: It...

    2023

  11. [19]

    Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Tiancheng Han, Botian Shi, Wenhai Wang, Junjun He, Kaipeng Zhang, Ping Luo, Yu Qiao, Qiaosheng Zhang, and Wenqi Shao. 2025. MM-Eureka: Exploring the Frontiers of Multimodal Reasoning with Rule-bas...

    2025 doi arXiv

  12. [20]

    OpenAI. 2024. GPT-4o System Card. (2024). arXiv:2410.21276 doi:10.48550/arXiv. 2410.21276

    2024 doi arXiv

  13. [21]

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. 2025. VLM-R1: A Stable and Generalizable R1-style Large Vision-Language Model. (2025). arXiv:2504.07615 doi:10.4855...

    2025 doi arXiv

  14. [22]

    verl-project. 2025. verl: Volcano Engine Reinforcement Learning for LLMs. GitHub repository. Retrieved March 31, 2026 from https://github.com/verl- project/verl

    2025

  15. [23]

    Haozhe Wang, Chao Qu, Zuming Huang, Wei Chu, Fangzhen Lin, and Wenhu Chen. 2025. VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Mod- els with Reinforcement Learning. InAdvances in Neural Information Processing Systems

    2025

  16. [24]

    Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. 2024. Measuring Multimodal Mathematical Reasoning with MATH-Vision Dataset. InAdvances in Neural Information Processing Systems, Datasets and Benchmarks Track

    2024

  17. [25]

    Penghao Wu and Saining Xie. 2024. V*: Guided Visual Search as a Core Mecha- nism in Multimodal LLMs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 13084–13094

    2024

  18. [26]

    Guowei Xu, Peng Jin, Ziang Wu, Hao Li, Yibing Song, Lichao Sun, and Li Yuan

  19. [27]

    Weiye Xu, Jiahao Wang, Weiyun Wang, Zhe Chen, Wengang Zhou, Aijun Yang, Lewei Lu, Houqiang Li, Xiaohua Wang, Xizhou Zhu, Wenhai Wang, Jifeng Dai, and Jinguo Zhu. 2026. VisuLogic: A Benchmark for Evaluating Visual Reason- ing in Multi-modal Large Language Models. InThe Fourteen...

    2026

  20. [28]

    Shuo Yang, Yuwei Niu, Yuyang Liu, Yang Ye, Bin Lin, and Li Yuan. 2025. Look- Back: Implicit Visual Re-focusing in MLLM Reasoning. (2025). arXiv:2507.03019 doi:10.48550/arXiv.2507.03019

    2025 doi arXiv

  21. [29]

    InProceed- ings of the IEEE/CVF International Conference on Computer Vision

    LLaVA-CoT: Let Vision Language Models Reason Step-by-Step. InProceed- ings of the IEEE/CVF International Conference on Computer Vision. 2087–2098

  22. [30]

    Le Zhang, Yixiong Xiao, Xinjiang Lu, Jingjia Cao, Yusai Zhao, Jingbo Zhou, Lang An, Zikan Feng, Wanxiang Sha, Yu Shi, et al. 2026. OmegaUse: Building a General- Purpose GUI Agent for Autonomous Task Execution. (2026). arXiv:2601.20380 doi:10.48550/arXiv.2601.20380

    2026 doi

  23. [31]

    Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, Peng Gao, and Hongsheng Li. 2024. MathVerse: Does Your Multi-modal LLM Truly See the Diagrams in Visual Math Problems?. InComputer Vision – ECCV 2024. Spri...

    2024 doi

  24. [32]

    Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. 2025. R1-VL: Learning to Reason with Multimodal Large Lan- guage Models via Step-wise Group Relative Policy Optimization. InProceedings of the IEEE/CVF International Conference on C...

    2025

  25. [33]

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo

  26. [34]

    Thinking with Im- ages

    Ziwei Zheng, Michael Yang, Jack Hong, Chenxiao Zhao, Guohai Xu, Le Yang, Chao Shen, and Xing Yu. 2026. DeepEyes: Incentivizing “Thinking with Im- ages” via Reinforcement Learning. InThe Fourteenth International Conference on Learning Representations. Peng et al. Table 6: Stage...

    2026

  27. [35]

    Yi-Fan Zhang, Xingyu Lu, Shukang Yin, Chaoyou Fu, Wei Chen, Xiao Hu, Bin Wen, Kaiyu Jiang, Changyi Liu, Tianke Zhang, Haonan Fan, Kaibing Chen, Jiankang Chen, Haojie Ding, Kaiyu Tang, Zhang Zhang, Liang Wang, Fan Yang, Tingting Gao, and Guorui Zhou. 2026. Thyme: Think Beyond I...

    2026

  28. [37]

    InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)

    LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations). Association for Computational Linguistics, Bangkok, Thailand, 400–410. doi:10.18653...

    doi

  29. [39]

    motor cortex

    Remaining:10−1−1=8</think> Answer: 8✗(negative flip) Targeted Self-Diagnosis <reground> Wait, I need to re-examine the small blue sphere — is it truly matte and blue, or could it be shiny like the other spheres? Also, let me check if there’s any yellow object that might have b...

  30. [2024]

    InComputer Vision – ECCV 2024

    MMBench: Is Your Multi-modal Model an All-Around Player?. InComputer Vision – ECCV 2024. 216–233. doi:10.1007/978-3-031-72658-3_13

    2024 doi

  31. [2025]

    InAdvances in Neural Information Processing Systems

    OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles. InAdvances in Neural Information Processing Systems

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.