Pith. sign in

REVIEW 4 major objections 6 minor 41 references

Remember-R1: Mitigating Long-Context Visual Forgetting through Reinforcement Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Three process-level rewards—visual-keyword coverage, late-stage attention persistence, and key-region focus—added to answer-level RL mitigate long-context visual forgetting and improve multimodal reasoning across seven benchmarks.

desk verdict A solid, novel RL recipe for visual forgetting whose central evidence is partly circular — worth serious review, but needs a behavior-based test of grounding. read the letter →

arxiv 2608.01314 v1 pith:RQ6DGBCJ submitted 2026-08-02 cs.CV

Jianmin Chen , Jiaqi Tang , Wei Wei , Xiaogang Xu , Jiafei Wu , Zhe Liu , Qianzhou Wang , Yingying Yan
show 4 more authors
Botong Geng Yuyang Xia Lei Zhang Qifeng Chen
This is my paper · ORCID
classification cs.CV
keywords visualforgettingmultimodallargelanguagemodelschain-of-thoughtreasoningreinforcementlearningprocess-levelrewardgroundingattentionpersistenceGRPO
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

When reasoning chains get long, multimodal models tend to lean more on the text they have already generated and less on the image, a failure the paper calls visual forgetting. Remember-R1 is a reinforcement-learning framework that attacks this drift by rewarding three things on the original reasoning trajectory: explicit coverage of annotated visual keywords, persistence of visual attention into later reasoning steps, and concentration of that attention on question-relevant image regions. The authors train Qwen2.5-VL models at 3B and 7B scales with these rewards added to the answer-correctness reward, and report consistent gains across seven benchmarks spanning math, logic, general multimodal ability, and visual perception. They also show that the trained models exhibit a slower decline in visual attention during generation, which is the training-time signal that forgetting is being mitigated. The appeal of the approach is that inference is unchanged: the model becomes more visually grounded purely from the reward signal.

What carries the argument

The central machinery is the attention proxy $\Omega(t)$, the aggregate attention mass assigned to visual tokens at generation step $t$ using final-layer attention heads, which the paper uses to quantify visual forgetting via the early-to-late change $\Delta_{\text{vis}} = \mathbb{E}_{t\in T_{\text{early}}}[\Omega(t)] - \mathbb{E}_{t\in T_{\text{late}}}[\Omega(t)]$. This proxy feeds two of the three process-level rewards: the Visual Memory Reward $r_{\text{mem}} = 1 + (\mu_{\text{end}} - \mu_{\text{start}})$, comparing average visual attention over the first and last windows of the response, and the Visual Key-Region Reward $r_{\text{region}}$, a temporally weighted average of $R_B(t) = \Ome

What would settle it

Run the trained Remember-R1 model on the same question with the image replaced by a blank or scrambled image (or with the key-region boxes shifted to irrelevant regions). If the model still produces high late-stage visual attention and answer rewards stay similar while correctness collapses, it would show the $r_{\text{mem}}$ and $r_{\text{region}}$ rewards are being gamed rather than reflecting real visual evidence use. A second check: measure whether late-stage visual attention predicts the correctness of the specific visual claim being made at that token; if attention to image tokens stays

Watch

Extended reading notes

Core claim

The paper's central claim is that long-context visual forgetting—the gradual weakening of a model's reliance on visual evidence as generation proceeds—can be reduced by applying process-level reward directly to the untouched reasoning trajectory, without any change at inference time. It identifies three complementary aspects of sustained visual grounding: the response should keep mentioning visual keywords that are annotated in the image, should keep allocating attention to visual tokens in late reasoning steps, and should keep that attention concentrated on the image regions that matter for the question. For each aspect it defines a reward: $r_{\text{voc}}$ rewards the position-weighted cov

Load-bearing premise

The load-bearing premise is that the attention weights at the final transformer layer accurately reflect how much the model actually relies on the image, so that rewarding higher late-stage visual attention and key-region attention genuinely improves visual grounding rather than merely inflating an attention signal.

Editorial extensions

If this is right

  • Long chain-of-thought reasoning in vision-language models can stay aligned with image facts, improving performance on reasoning-heavy benchmarks such as MathVista and LogicVista without extra compute at inference.
  • Process-level rewards are complementary to answer-level RL: the accuracy-only GRPO baseline improves less, and removing any of the three rewards degrades results on multiple benchmarks.
  • The method scales: consistent gains appear at both 3B and 7B model sizes, and on general multimodal and perception benchmarks as well as reasoning tasks, indicating no perception trade-off.
  • Attention dynamics are changed by training: the model's visual attention ratio declines more slowly over generation steps on both LogicVista and MMStar, directly supporting the forgetting-mitigation claim.

Reading between the lines

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

  • The attention-proxy assumption could be probed by a decoupling test: if the trained model is run with a blank or shuffled image while the text history is unchanged, and it still maintains high late-stage visual attention, then the $r_{\text{mem}}$ and $r_{\text{region}}$ rewards may be inflating attention without genuine visual evidence use; that would separate the attention-gaming explanation fro
  • Because $r_{\text{voc}}$ uses exact phrase matching with no synonym expansion, a paraphrased evaluation (same visual facts, different words) would show whether the vocabulary reward teaches flexible visual reference or encourages lexical copying of annotation phrases.
  • The annotation pipeline (automated keyword and box generation plus manual verification) is a scalability bottleneck; whether the method survives without the human verification step, or with noisy automatic annotations, is a testable extension that determines its practical reach.
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 / 6 minor

Summary. Remember-R1 proposes a GRPO-based reinforcement learning method for multimodal LLMs that adds three process-level rewards to answer correctness: a visual vocabulary reward for exact matching of pre-annotated image keywords during the rollout, a visual memory reward that penalizes late-stage decline of final-layer attention to visual tokens, and a visual key-region reward that up-weights attention to question-relevant image regions in later generation steps. The training set is ViRL39K augmented with Qwen-VL-Max-extracted and manually verified keyword/box annotations. Experiments on seven benchmarks at 3B and 7B scales show consistent accuracy gains over the Qwen2.5-VL base models and most compared baselines, and ablations indicate each reward contributes. Attention-ratio analyses on LogicVista and MMStar show a slower decline in visual attention after training, which the paper presents as evidence of reduced visual forgetting.

Significance. The proposal is practically valuable if the mechanism holds: it is a training-only intervention that changes no inference procedure and reportedly improves reasoning accuracy across diverse multimodal benchmarks. The paper's strengths are its clear decomposition of visual forgetting into three observable reward signals, the manual verification of the annotation pipeline, the two-scale evaluation, the per-reward ablations, and the code release. The main risk is that the two attention-based rewards directly optimize the same attention statistics later used as evidence of forgetting mitigation, so the causal attribution in the abstract is not yet established; missing hyperparameter reporting and lack of statistical uncertainty also weaken the quantitative claims.

major comments (4)
  1. [§3.2.2, §3.2.3, §4.3 (Eqs. 9, 13, 14; Fig. 6)] The forgetting-mitigation evidence is partly circular. r_mem directly rewards μ_end−μ_start and r_region directly rewards a time-weighted average of R_B, both computed from the same final-layer attention Ω(t) used in Fig. 6 to show slower attention decline and in Fig. 7(b)/(c) to validate the ablations. Eq. (13) also has an incentive to inflate R_B by reducing total visual attention Ω(t) in the denominator. The manuscript's caveat that Ω(t) "does not fully characterize model behavior" (§3) is never backed by a behavioral test. I ask for an independent grounding test—for example, occluding/perturbing key regions and measuring answer sensitivity, counterfactual image swaps, or comparing text-only versus image-grounded rollouts—and a discussion of the denominator incentive.
  2. [§3.2, §3.3, §4.1 (Eqs. 5, 7, 13, 16)] Key hyperparameters are not reported: c, L_min, γ, ε_den, β, and the reward combination weights in Eq. (16). These determine the relative scale of the four rewards; r_mem can range roughly [0,2], whereas r_voc and r_region are at most 1, so the unweighted sum may be dominated by r_mem. The ablation results in Table 2 could reflect this scale imbalance rather than complementary benefits. Please report the final values and at least a small sensitivity analysis (or justify the chosen scales).
  3. [Tables 1 and 2] No error bars, multiple seeds, or significance tests are provided. Several reported gains are small (e.g., MMMB +0.35/+0.95, RealWorldQA +0.66/+0.39, LogicVista +2.46/+1.79), and the ablation differences in Table 2 are often within a few points. With single runs, "consistently improves" is not statistically supported. Please report means/std over at least three seeds, or paired significance tests across benchmark items.
  4. [§4.1] The manuscript does not analyze overlap between the ViRL39K training set and the seven evaluation benchmarks. If ViRL39K contains examples from MathVista, MMStar, or related visual-reasoning sources, the large gains on those benchmarks could be inflated. Please report a substring/UUID-level overlap check and, if any overlap exists, re-run the main comparison on contamination-free subsets.
minor comments (6)
  1. [Table 1] In the VLAA-Thinking-7B row, "58.8383.6364.33" appears to be missing spaces; it should be "58.83 83.63 64.33".
  2. [Figure 4] The rendered figure contains unreadable placeholder text ("???? ?????? ..."), likely a font-embedding issue; please replace with a clean export.
  3. [§3.3] r_acc is used in Eq. (16) but never formally defined; please state whether it is 0/1 exact-match or partial credit.
  4. [Eq. (5)] Please clarify the role of L_min when L_min > T, since T_eff then exceeds the response length and reduces the position weight; also report the chosen value.
  5. [§4.3] The y-axis label of Fig. 6 and the exact definition of "attention ratio" (normalized by total attention to all tokens or by something else) should be stated in the caption.
  6. [Table 2] Ablations are reported only for the 3B model; if 7B ablations were omitted due to compute constraints, this should be stated explicitly.

Circularity Check

2 steps flagged · score 6.0 of 10

Attention-based rewards directly optimize the visual-attention metrics later presented as evidence of mitigated forgetting; benchmark gains are external, but the forgetting-mitigation claim is partly circular.

  1. self definitional [Section 3 'Quantifying Visual Forgetting'; Section 3.2.2 Eqs. (7)-(9); Section 4.3 'Attention Ratio Analysis' / Figure 6]
    "Section 3: 'We capture visual forgetting through the early-to-late change in visual attention, Δvis = E_{t∈T_early}[Ω(t)] − E_{t∈T_late}[Ω(t)].' Eq. (9): 'r_mem(Y)=1+(μ_end−μ_start).' Section 4.3: 'Compared with the base model, however, Remember-R1 shows a consistently slower decline ... This pattern suggests that Remember-R1 more effectively preserves reliance on visual evidence.'"

    μ_start and μ_end (Eq. 8) are averages of Ω(t) over the first and last windows, so r_mem = 1 − Δvis up to the same windowing. Training with r_mem directly maximizes late-window minus early-window visual attention, i.e., directly minimizes the paper's own forgetting measure. Figure 6 then plots Ω(t) and reports a slower decline; this is a check of the training objective, not independent evidence. The abstract's mechanistic claim ('slows the decline of visual attention ... supporting its effectiveness') restates the reward. Benchmark accuracy is external, but this part of the evidence reduces by construction.

  2. self definitional [Section 3.2.3 Eqs. (12)-(14); Section 4.4 'Effectiveness of r_region' / Figure 7(c)]
    "Eq. (13)-(14): 'R_B(t)=Ω_B(t)/(Ω(t)+ε_den)' and 'r_region(Y)=1/Z ∑_{t=1}^T (t/T) R_B(t).' Section 4.4: 'Removing r_region substantially reduces attention aligned with question-relevant regions in Figure 7(c).'"

    r_region is a temporally weighted average of R_B(t), and the ablation evidence is the same R_B(t) quantity: removing the reward lowers exactly the metric the reward was defined to raise. The 'key-region attention' result is therefore a tautological manipulation check, not an independent test of whether attention is behaviorally grounded. Additionally, R_B(t) is normalized by total visual attention Ω(t), so the reward can be inflated by suppressing the denominator; no non-attention behavioral verification of grounding is provided.

full rationale

The benchmark comparisons in Tables 1-2 are genuine external evidence: accuracy on seven held-out benchmarks improves under the full reward, and dropping rewards lowers accuracy, so the performance claim is not circular. The circularity is confined to the forgetting-mitigation mechanism. The paper defines visual forgetting as the early-to-late decline of visual-attention mass Ω(t) (Section 3). r_mem is literally 1 + (late-window mean Ω − early-window mean Ω), i.e., the negative of that decline, and r_region is a weighted average of a ratio of the same attention weights. Reporting post-training that Ω(t) declines more slowly (Figure 6) and that removing r_mem/r_region lowers the corresponding attention quantities (Figure 7) validates the objective rather than the hypothesis. The paper acknowledges Ω(t) 'does not fully characterize model behavior' but does not supply a behavioral, non-attention test that the model actually uses image content more. Thus the central mechanistic claim 'mitigates long-context visual forgetting' is partially supported by construction. No load-bearing self-citation is involved; the issue is reward-metric identity, not citation chains. Score reflects partial circularity, not total.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several unreported hyperparameters and on domain assumptions about attention as a proxy and the quality of automatically generated annotations. No new physical entities are introduced.

free parameters (6)
  • c = not reported
    Scaling coefficient in Visual Vocabulary Reward (Eq. 5); controls reward magnitude before clipping.
  • gamma = not reported
    Fraction of response length defining early/late windows in Visual Memory Reward (Eq. 7).
  • L_min = not reported
    Minimum effective response length used to normalize keyword position rewards (Eq. 5).
  • epsilon_den = not reported
    Stability constant in key-region attention ratio (Eq. 13).
  • reward weights = all equal to 1
    Eq. 16 combines r_acc, r_voc, r_mem, r_region with equal weight; no scaling reported despite different ranges (r_mem in [0,2], others in [0,1]).
  • beta (KL coefficient) = not reported
    GRPO KL regularization strength in Eq. 20 not specified.
assumptions (4)
  • domain assumption Attention weights from the final transformer layer are a valid proxy for visual evidence reliance.
    Stated in Section 3, Eq. (6), and used as both reward signal and as evidence of forgetting mitigation in Figure 6.
  • domain assumption Exact phrase matching of Qwen-VL-Max extracted keywords provides useful supervision for visual grounding.
    Section 3.1 and Eq. (4); the keywords are generated by an external proprietary model and manually verified, but no independent evidence that keyword coverage causally improves reasoning.
  • domain assumption Annotated key-region boxes identify regions relevant to the question.
    Section 3.1; boxes come from Qwen-VL-Max prompted with (I,Q,A) and manual correction. If boxes are noisy, r_region rewards attending to the wrong regions.
  • domain assumption Process-level rewards on the original rollout generalize beyond the training distribution.
    This is the central hypothesis; the paper assumes that optimizing these proxies will transfer to held-out benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Remember-R1: Mitigating Long-Context Visual Forgetting through Reinforcement Learning." pith.science (2026). https://pith.science/paper/RQ6DGBCJ

@misc{pith2026260801314,
  author       = {Pith},
  title        = {Pith review of: Remember-R1: Mitigating Long-Context Visual Forgetting through Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RQ6DGBCJ}},
  note         = {Machine review of arXiv:2608.01314}
}
read the original abstract

Multimodal large language models (MLLMs) increasingly rely on long chain-of-thought reasoning for complex tasks. However, as reasoning sequences lengthen, models may gradually rely less on visual evidence and more on accumulated textual context, leading to visual forgetting. Existing approaches do not directly constrain how visual evidence is used and maintained along the original reasoning trajectory, leaving long-context visual forgetting insufficiently addressed. To address this issue, we propose Remember-R1, a reinforcement learning framework that mitigates long-context visual forgetting by applying process-level supervision directly on the original reasoning trajectory. Specifically, Remember-R1 introduces rewards that encourage broader coverage of matched visual keywords, stronger persistence of visual dependence in later reasoning steps, and greater focus on question-relevant image regions. Experiments across multiple model scales and diverse multimodal benchmarks demonstrate that Remember-R1 consistently improves reasoning performance. Additional analyses further show that it slows the decline of visual attention during generation, supporting its effectiveness in mitigating long-context visual forgetting.

Figures

Figures reproduced from arXiv: 2608.01314 by the authors.

Figure 1
Figure 1. Example of visual forgetting in long-context mul [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of three approaches to mitigating vi [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Benchmark-level evidence for the three-aspect decomposition underlying our reward design. (a) Correct responses [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Overview of Remember-R1. Three process-level rewards supervise matched annotated visual-keyword coverage, [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison on geometric reasoning. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Effects of removing the three process-level rewards: [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 18 canonical work pages

  1. [1]

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

    arXiv 2025

  2. [2]

    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)

    work page 2025

  3. [3]

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al . 2024. Are we on the right way for evaluating large vision-language models?Advances in Neural Information Processing Systems37 (2024), 27056–27087

    2024

  4. [4]

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. 2024. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 24185–24198

    2024

  5. [5]

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. 2024. Openai o1 system card.arXiv preprint arXiv:2412.16720(2024)

    arXiv 2024

  6. [6]

    Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. 2025. Your large vision-language model only needs a few attention heads for visual grounding. InProceedings of the Computer Vision and Pattern Recognition Conference. 9339– 9350

    work page 2025

  7. [7]

    Muhammad Khalifa, Rishabh Agarwal, Lajanugen Logeswaran, Jaekyeom Kim, Hao Peng, Moontae Lee, Honglak Lee, and Lu Wang. 2026. Process Reward Models That Think.Transactions on Machine Learning Research(2026)

    work page 2026

  8. [8]

    Yantao Li, Qiang Hui, Chenyang Yan, Kanzhi Cheng, Fang Zhao, Chao Tan, Huanling Gao, Jianbing Zhang, Kai Wang, Xinyu Dai, et al. 2026. PaLMR: Towards Faithful Visual Reasoning via Multimodal Process Alignment.arXiv preprint arXiv:2603.06652(2026)

    arXiv 2026

Show all 41 references
  1. [9]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437(2024)

    2024 arXiv

  2. [10]

    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. In International Conference on Learning Repre...

    2024

  3. [11]

    Ke Ma, Jiaqi Tang, Bin Guo, Xueting Han, Ruonan Xu, Qingfeng He, Ziheng Wang, Xu Wang, Qifeng Chen, Zhiwen Yu, and Yunhao Liu. 2026. Response-G1: Explicit Scene Graph Modeling for Proactive Streaming Video Understanding. InProceedings of the 64th Annual Meeting of the Associat...

    2026

  4. [12]

    Lingfeng Ming, Yadong Li, Song Chen, Jianhua Xu, Zenan Zhou, and Weipeng Chen. 2025. Ocean-R1: An Open and Generalizable Large Vision-Language Model enhanced by Reinforcement Learning. https://github.com/VLM-RL/Ocean-R1. Accessed: 2025-04-03

    2025

  5. [13]

    Yingzhe Peng, Gongrui Zhang, Miaosen Zhang, Zhiyuan You, Jie Liu, Qipeng Zhu, Kai Yang, Xingzhong Xu, Xin Geng, and Xu Yang. 2025. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl.arXiv preprint arXiv:2503.07536(2025)

    2025 arXiv

  6. [14]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems36 (2023), 53728–53741

    2023

  7. [15]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  8. [16]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

    2024 arXiv

  9. [17]

    Hai-Long Sun, Zhun Sun, Houwen Peng, and Han-Jia Ye. 2025. Mitigating visual forgetting via take-along visual conditioning for multi-modal long cot reasoning. arXiv preprint arXiv:2503.13360(2025)

    2025 arXiv

  10. [18]

    Hai-Long Sun, Da-Wei Zhou, Yang Li, Shiyin Lu, Chao Yi, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, De-Chuan Zhan, and Han-Jia Ye. 2025. Parrot: Multilingual Visual Instruction Tuning. InForty-second International Conference on Machine Learning

    2025

  11. [19]

    Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liangyan Gui, Yu-Xiong Wang, Yiming Yang, et al . 2024. Aligning large multimodal models with factually augmented rlhf. InFindings of the Association for Computational Linguistics: ACL 20...

    2024

  12. [20]

    Jiaqi Tang, Jianmin Chen, Wei Wei, Xiaogang Xu, Runtao Liu, Xiangyu Wu, Qipeng Xie, Jiafei Wu, Lei Zhang, and Qifeng Chen. 2026. Robust-r1: Degradation- aware reasoning for robust visual understanding. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 9421–9429

    2026

  13. [21]

    Jiaqi Tang, Jianmin Chen, Youyang Zhai, Wei Wei, Runtao Liu, Mengjie Zhao, Xiangyu Wu, Qingfa Xiao, and Qifeng Chen. 2026. Robust-U1: Can MLLMs Self- Recover Corrupted Visual Content for Robust Understanding?arXiv preprint arXiv:2606.08063(2026)

    2026 arXiv

  14. [22]

    Jiaqi Tang, Hao Lu, Ruizheng Wu, Xiaogang Xu, Ke Ma, Cheng Fang, Bin Guo, Jiangbo Lu, Qifeng Chen, and Ying-Cong Chen. 2024. HAWK: Learning to Understand Open-World Video Anomalies. InAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. F...

    2024 doi

  15. [23]

    Jiaqi Tang, Yu Xia, Yi-Feng Wu, Yuwei Hu, Chen Yuhui, Qing-Guo Chen, Xi- aogang Xu, Xiangyu Wu, Hao LU, Yanqing Ma, Shiyin Lu, and Qifeng Chen

  16. [24]

    Jiaqi Tang, Yingying Yan, Qianzhou Wang, Yuyang Xia, Botong Geng, Jianmin Chen, Ke Ma, Youyang Zhai, Qingfeng He, Weigeng Shao, et al. 2026. Intelligent Remote Sensing Agents: A Survey. (2026)

    2026

  17. [25]

    Xinyu Tian, Shu Zou, Zhaoyuan Yang, Mengqi He, Fabian Waschkowski, Lukas Wesemann, Peter Tu, and Jing Zhang. 2025. More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models.arXiv preprint arXiv:2509.25848(2025)

    2025

  18. [26]

    Haozhe Wang, Chao Qu, Zuming Huang, Wei Chu, Fangzhen Lin, and Wenhu Chen. 2025. Vl-rethinker: Incentivizing self-reflection of vision-language models with reinforcement learning.arXiv preprint arXiv:2504.08837(2025)

    2025 arXiv

  19. [27]

    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.Advances in Neural Information Processing Systems37 (2024), 95095–95169

    2024

  20. [28]

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

    2022

  21. [29]

    Jinjian Wu, Jiaqi Tang, Wei Wei, Yingying Yan, Jianmin Chen, Botong Geng, Lei Zhang, and Qifeng Chen. 2026. IQA-T1: Tool-based Visual Evidence Reasoning for Image Quality Assessment. arXiv:2607.12375 [cs.CV] https://arxiv.org/abs/ 2607.12375

    2026 arXiv

  22. [30]

    xAI. 2024. Grok-1.5 Vision Preview. https://x.ai/blog/grok-1.5v

    2024

  23. [31]

    Yijia Xiao, Edward Sun, Tianyu Liu, and Wei Wang. 2024. Logicvista: Mul- timodal llm logical reasoning benchmark in visual contexts.arXiv preprint arXiv:2407.04973(2024)

    2024 arXiv

  24. [32]

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

    2025 arXiv

  25. [33]

    Shuo Yang, Yuwei Niu, Yuyang Liu, Yang Ye, Bin Lin, and Li Yuan. 2026. Look- back: Implicit visual re-focusing in mllm reasoning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 11694–11702

    2026

  26. [34]

    Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. 2024. MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities. InForty-first International Conference on Machine Learning

    2024

  27. [35]

    Chi Zhang, Haibo Qiu, Qiming Zhang, Zhixiong Zeng, Lin Ma, and Jing Zhang

  28. [36]

    Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara, and Filip Ilievski. 2025. MLLMs Know Where to Look: Training-free Perception of Small Visual Details with Multimodal LLMs. InThe Thirteenth International Conference on Learning Representations. https://arxiv.org/abs/2502.17422

    2025 arXiv

  29. [37]

    Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. 2023. Multimodal chain-of-thought reasoning in language models.arXiv preprint arXiv:2302.00923(2023)

    2023 arXiv

  30. [38]

    Thinking with Images

    Ziwei Zheng, Michael Yang, Jack Hong, Chenxiao Zhao, Guohai Xu, Le Yang, Chao Shen, and Xing Yu. 2025. DeepEyes: Incentivizing" Thinking with Images" via Reinforcement Learning.arXiv preprint arXiv:2505.14362(2025)

    2025 arXiv

  31. [2017]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347 (2017)

    2017 arXiv

  32. [2025]

    arXiv preprint arXiv:2509.25866(2025)

    Deepsketcher: Internalizing visual manipulation for multimodal reasoning. arXiv preprint arXiv:2509.25866(2025)

    2025 arXiv

  33. [2026]

    InFindings of the Association for Computational Linguistics: ACL 2026, Maria Liakata, Viviane P

    LPO: Towards Accurate GUI Agent Interaction via Location Preference Optimization. InFindings of the Association for Computational Linguistics: ACL 2026, Maria Liakata, Viviane P. Moreira, Jiajun Zhang, and David Jurgens (Eds.). Association for Computational Linguistics, San Di...

    2026 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.