pith. sign in

arxiv: 2606.26631 · v1 · pith:3UCNR7XFnew · submitted 2026-06-25 · 💻 cs.CV

Position Rebinding Cache Reuse: Replay-Free Visual Revisiting for Interleaved Multimodal Reasoning

Mengzhao Wang , Yanli Ji , Wangmeng Zuo , Peng Ye , Chongjun Tu This is my paper

Pith reviewed 2026-06-26 05:06 UTC · model grok-4.3

classification 💻 cs.CV
keywords Position RebindingCache ReuseMultimodal ReasoningVisual RevisitingKV CacheInterleaved GenerationAttention Preservation
0
0 comments X

The pith

Rebinding positions to cached visual keys enables replay-free visual revisiting without attention distortion.

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

The paper establishes that directly copying historical visual KV caches fails in interleaved multimodal reasoning because the keys remain bound to their original positions, distorting attention in later decoding steps. Instead, PRCR stores the raw caches with their spatial coordinates, then reassigns new position-compatible coordinates and rebinds the keys before injection. This preserves both textual continuity and relative visual structure. A sympathetic reader would care because replaying visual tokens at every step incurs prohibitive compute costs as reasoning chains lengthen, while this cache-level shortcut promises equivalent accuracy at far lower cost.

Core claim

PRCR stores raw visual KV cache entries together with their original spatial coordinates, then reassigns position-compatible coordinates to selected entries and rebinds their keys before injecting the reconstructed cache into the active decoder cache, thereby reusing historical visual evidence while avoiding the autoregressive collapse that occurs with stale positional bindings.

What carries the argument

Position rebinding, which reconstructs visual evidence by reassigning coordinates to cached keys and rebinding them to match the current decoding state.

If this is right

  • Multimodal models can revisit visual evidence during generation without repeating forward passes on the same tokens.
  • Average accuracy on reasoning benchmarks rises by about 5 percent relative to replay baselines.
  • Visual-revisiting computation drops by up to tens of thousands of times while matching replay-level performance.
  • Textual positional continuity is maintained even as visual caches are reused across multiple reasoning steps.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same rebinding idea might extend to audio or other sensory caches in long-horizon agents.
  • PRCR could be combined with eviction policies to keep only the most reusable visual entries in memory.
  • If coordinate reassignment proves robust, similar cache reuse might apply to non-visual modalities that suffer positional drift.

Load-bearing premise

Reassigning position-compatible coordinates to selected cached visual entries and rebinding their keys will preserve relative visual structure and avoid attention distortion under new decoding contexts.

What would settle it

A controlled comparison on an interleaved multimodal benchmark where PRCR produces lower accuracy or visible attention collapse than full token replay under identical model and prompt conditions.

Figures

Figures reproduced from arXiv: 2606.26631 by Chongjun Tu, Mengzhao Wang, Peng Ye, Wangmeng Zuo, Yanli Ji.

Figure 1
Figure 1. Figure 1: Motivation and efficiency of PRCR. (a) Category-level results on M3CoT show that Direct KV Cache Reuse collapses compared with the baseline. Here, NS, SC, PC, and ALL denote the reported M3CoT subsets and the overall score, respectively. (b) Overall M3CoT accuracy shows that PRCR avoids this collapse and achieves Token-Replay or better performance. (c)Under the same selected token budget, PRCR (32B model) … view at source ↗
Figure 2
Figure 2. Figure 2: Decoding collapse under direct historical KV reuse. (a) Example generation trajectories for Token-Replay and Direct KV Cache Reuse (b) Failure-type pie charts, showing direct KV reuse mainly causes stuck decoding. (c) Token probabilities over generation steps, with repeated-token loops after failure onset. (d) Token entropy over generation steps; direct KV reuse collapses to low-entropy states, while Token… view at source ↗
Figure 3
Figure 3. Figure 3: Attention perturbation under direct historical KV cache reuse. The heatmap shows how Direct KV Cache Reuse disturbs attention across the active decoding context. Green shading indicates attention weight (low to high). where 𝑅(·) denotes the relative rota￾tion induced by RoPE. Eq. (5) shows that Direct KV Cache Reuse computes attention with a stale relative offset be￾tween the current text position and the … view at source ↗
Figure 4
Figure 4. Figure 4: Overview of Position Rebinding Cache Reuse (PRCR). PRCR enables replay-free visual revisiting through three stages: (a) store pre-RoPE visual KV cache with original positions; (b) reassign compatible coordinates to selected entries; (c) rebind keys with RoPE and inject the reconstructed cache for subsequent decoding. key/value projections as 𝑘 raw 𝑚,ℓ = 𝑊ℓ 𝐾 ℎ𝑚,ℓ , 𝑣raw 𝑚,ℓ = 𝑊ℓ 𝑉 ℎ𝑚,ℓ , 𝑚 = 1, . . . , 𝑀, … view at source ↗
Figure 5
Figure 5. Figure 5: Effect of left extension length 𝐿 in PRCR. Accuracy on M3CoT and MathVista peaks at 𝐿 = 2. We further evaluate the computational and memory cost of PRCR. Token￾Replay requires forwarding the se￾lected visual tokens through the full decoder whenever visual evidence is revisited, whereas PRCR reconstructs selected visual cache entries by re￾binding their keys and injecting the resulting KV cache into the act… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results of baseline and PRCR. The baseline predicts 0 inches, whereas PRCR revisits the visual evidence and outputs the correct answer of 2. Instruct and 192MB on Qwen3-VL-32B-Instruct, corresponding to only 0.64% and 0.30% overhead, respectively. This indicates that the computational advantage of PRCR is achieved with a moderate and practical memory cost. 4.4 Ablation Study Effect of position … view at source ↗
Figure 7
Figure 7. Figure 7: Effect of PRCR on next-token generation. (a) Next-token probabilities after inserting the visual KV cache at token 80. (b) Corresponding output entropy. Shaded areas show changes due to the cache [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Attention visualization for PRCR. (a) Token-Replay: attention over inserted visual tokens. (b) PRCR: attention over reconstructed visual KV cache. Green shading indicates attention weight (low to high). Interleaved multimodal CoT (ICoT) inserts visual evidence into the rea￾soning trajectory rather than relying only on textual rationales. Early ICoT methods revisit visual informa￾tion through interleaved vi… view at source ↗
read the original abstract

Interleaved multimodal reasoning improves visual grounding by revisiting visual evidence during multi-step generation, yet existing methods typically rely on token replay, repeatedly forwarding selected visual tokens. A natural shortcut is to reuse the historical visual key-value (KV) cache directly. However, we identify a critical failure mode of this strategy: cached visual keys are already bound to their original positional context. Such stale positional binding distorts attention under later decoding contexts and can trigger severe autoregressive decoding collapse. This failure suggests that effective cache reuse requires reconstructing visual evidence under positions compatible with the current decoding state, rather than directly copying position-bound historical cache entries. To this end, we propose Position Rebinding Cache Reuse (PRCR), a cache-level framework for replay-free visual revisiting. PRCR stores raw visual KV cache together with their original spatial coordinates, then reassigns position-compatible coordinates to select entries and rebinds their keys before injecting the reconstructed cache into the active decoder cache. This design reuses historical visual evidence while preserving textual positional continuity and relative visual structure. Experiments across multiple multimodal reasoning benchmarks show that PRCR achieves replay-level or better performance, improving average accuracy by 5 percent and reducing visual-revisiting computation by up to tens of thousands of times.

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

1 major / 1 minor

Summary. The paper proposes Position Rebinding Cache Reuse (PRCR) to enable replay-free visual revisiting during interleaved multimodal reasoning. It identifies that direct reuse of historical visual KV cache entries fails because the keys remain bound to their original positional context, distorting attention and risking autoregressive collapse under new decoding states. PRCR stores raw visual KV entries with their original spatial coordinates, then reassigns position-compatible coordinates to selected entries and rebinds their keys before injection into the active decoder cache. The authors claim this preserves textual positional continuity and relative visual structure, achieving replay-level or superior performance with a 5% average accuracy gain and up to tens-of-thousands-fold reduction in visual-revisiting computation across multiple benchmarks.

Significance. If the rebinding step is shown to correctly update key interactions with the current positional context, the result would be significant for efficient long-context multimodal models. The approach targets a practical bottleneck in vision-language reasoning by eliminating repeated visual token forwarding while maintaining accuracy, and the reported efficiency gains could enable more complex interleaved tasks. The identification of the stale positional binding failure mode is a clear contribution independent of the specific fix.

major comments (1)
  1. [Abstract] Abstract: The rebinding operation is described only as 'rebinds their keys' after coordinate reassignment, with no equation, algorithm, pseudocode, or implementation detail on whether keys are recomputed via the projection matrix with new positional embeddings or via another adjustment. This is load-bearing for the central claim because the replay-free guarantee and the reported accuracy/compute gains rest entirely on the rebinding succeeding where direct cache reuse fails, as noted in the stress-test concern.
minor comments (1)
  1. [Abstract] Abstract: The experimental claims (5% accuracy gain, 'multiple multimodal reasoning benchmarks', 'tens of thousands of times' reduction) are stated without naming the benchmarks, baselines, number of runs, or error bars; these details belong in the main text but their absence from the abstract reduces verifiability.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the need for greater clarity on the rebinding mechanism, which is indeed central to the PRCR contribution. We agree that the abstract's brevity leaves the precise implementation of key rebinding underspecified and will revise to address this.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The rebinding operation is described only as 'rebinds their keys' after coordinate reassignment, with no equation, algorithm, pseudocode, or implementation detail on whether keys are recomputed via the projection matrix with new positional embeddings or via another adjustment. This is load-bearing for the central claim because the replay-free guarantee and the reported accuracy/compute gains rest entirely on the rebinding succeeding where direct cache reuse fails, as noted in the stress-test concern.

    Authors: We acknowledge that the abstract provides only a high-level description of the rebinding step. In the full manuscript (Section 3.2 and Algorithm 1), the rebinding is implemented by re-computing the key projections using the original value vectors and the new position-compatible rotary embeddings (i.e., K' = W_k * V_original with updated RoPE angles derived from the reassigned coordinates), rather than a simple additive adjustment. This ensures the keys reflect the current positional context while preserving the cached visual content. We will revise the abstract to include a concise statement of this mechanism and add an explicit equation for the rebinding operation. We will also expand the stress-test discussion to quantify the failure of direct reuse versus PRCR. revision: yes

Circularity Check

0 steps flagged

No circularity: PRCR is an independent algorithmic design with separate empirical validation

full rationale

The paper proposes PRCR as a cache-reuse framework that stores raw visual KV entries with spatial coordinates, reassigns compatible positions, and rebinds keys. This design is stated directly in the abstract without reduction to fitted parameters, self-citations, or prior author results. No equations appear in the provided text, and the reported accuracy/efficiency gains are attributed to benchmark experiments rather than any quantity defined by construction from the method inputs. The central assumption about rebinding preserving structure is presented as a design hypothesis, not a self-referential derivation. This matches the default expectation of a non-circular paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the effectiveness of the newly introduced PRCR design and standard transformer KV cache mechanics; no explicit free parameters or additional axioms are stated in the abstract.

axioms (1)
  • standard math Standard properties of key-value caches and positional encodings in decoder-only transformer models
    The method builds directly on existing KV cache reuse concepts without additional unstated mathematical assumptions.
invented entities (1)
  • Position Rebinding Cache Reuse (PRCR) no independent evidence
    purpose: Cache-level framework for replay-free visual revisiting
    Newly proposed algorithmic construct introduced to solve the identified positional binding failure mode.

pith-pipeline@v0.9.1-grok · 5763 in / 1283 out tokens · 40759 ms · 2026-06-26T05:06:20.357983+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

42 extracted references · 8 linked inside Pith

  1. [1]

    Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35: 23716–23736, 2022

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35: 23716–23736, 2022

    2022

  2. [2]

    LLaVA-OneVision-1.5: Fully open framework for democratized multimodal training

    Xiang An, Yin Xie, Kaicheng Yang, Wenkang Zhang, Xiuwei Zhao, Zheng Cheng, Yirui Wang, Songcen Xu, Changrui Chen, Chunsheng Wu, Huajie Tan, Chunyuan Li, Jing Yang, Jie Yu, XiyaoWang,BinQin,YumengWang,ZizhenYan,ZiyongFeng,ZiweiLiu,BoLi,andJiankang Deng. LLaVA-OneVision-1.5: Fully open framework for democratized multimodal training. arXiv preprint arXiv:250...

    Pith/arXiv arXiv 2025

  3. [3]

    Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966, 2023

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966, 2023

    Pith/arXiv arXiv 2023

  4. [4]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Perceptiontokensenhancevisualreasoninginmultimodallanguagemodels

    Mahtab Bigverdi, Zelun Luo, Cheng-Yu Hsieh, Ethan Shen, Dongping Chen, Linda G Shapiro, andRanjayKrishna. Perceptiontokensenhancevisualreasoninginmultimodallanguagemodels. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 3836–3845, 2025

    2025

  6. [6]

    Reasoning in the dark: Interleaved vision-text reasoning in latent space.arXiv preprint arXiv:2510.12603, 2025

    Chao Chen, Zhixin Ma, Yongqi Li, Yupeng Hu, Yinwei Wei, Wenjie Li, and Liqiang Nie. Reasoning in the dark: Interleaved vision-text reasoning in latent space.arXiv preprint arXiv:2510.12603, 2025

    arXiv 2025

  7. [7]

    Are we on the right way for evaluating large vision-language models?Advances in Neural Information Processing Systems, 37:27056–27087, 2024

    LinChen,JinsongLi,XiaoyiDong,PanZhang,YuhangZang,ZehuiChen,HaodongDuan,Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models?Advances in Neural Information Processing Systems, 37:27056–27087, 2024

    2024

  8. [8]

    M3CoT: A novel benchmark for multi-domain multi-step multi-modal chain-of-thought

    Qiguang Chen, Libo Qin, Jin Zhang, Zhi Chen, Xiao Xu, and Wanxiang Che. M3CoT: A novel benchmark for multi-domain multi-step multi-modal chain-of-thought. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 8199–8221, 2024

    2024

  9. [9]

    MINT-CoT:Enablinginterleavedvisualtokensinmathematicalchain-of-thought reasoning

    Xinyan Chen, Renrui Zhang, Dongzhi Jiang, Aojun Zhou, Shilin Yan, Weifeng Lin, and HongshengLi. MINT-CoT:Enablinginterleavedvisualtokensinmathematicalchain-of-thought reasoning. InAdvances in Neural Information Processing Systems, 2025

    2025

  10. [10]

    Visual thoughts: A unified perspective of understanding multimodal chain-of-thought

    Zihui Cheng, Qiguang Chen, Xiao Xu, Jiaqi Wang, Weiyun Wang, Hao Fei, Yidong Wang, Alex Jinpeng Wang, Zhi Chen, Wanxiang Che, and Libo Qin. Visual thoughts: A unified perspective of understanding multimodal chain-of-thought. InAdvances in Neural Information Processing Systems, 2025

    2025

  11. [11]

    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. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 23678–23686, 2025

    2025

  12. [12]

    Drivingvqa: A dataset for interleaved visual chain-of-thought in real-world driving scenarios

    Charles Corbière, Simon Roburin, Syrielle Montariol, Antoine Bosselut, and Alexandre Alahi. Drivingvqa: A dataset for interleaved visual chain-of-thought in real-world driving scenarios. InFindings of the Association for Computational Linguistics: EACL 2026, pages 3309–3333, 2026

    2026

  13. [13]

    Instructblip: Towards general-purpose vision-language models with instruction tuning

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructblip: Towards general-purpose vision-language models with instruction tuning. volume 36, pages 49250–49267, 2023

    2023

  14. [14]

    Interleaved-modal chain-of-thought

    Jun Gao, Yongqi Li, Ziqiang Cao, and Wenjie Li. Interleaved-modal chain-of-thought. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 19520–19529, 2025. 12

    2025

  15. [15]

    Beyond static visual tokens: Structured sequential visual chain-of-thought reasoning.arXiv preprint arXiv:2603.26737, 2026

    Guangfu Guo, Xiaoqian Lu, Yue Feng, and Mingming Sun. Beyond static visual tokens: Structured sequential visual chain-of-thought reasoning.arXiv preprint arXiv:2603.26737, 2026

    arXiv 2026

  16. [16]

    Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding

    Yongxin Guo, Jingyu Liu, Mingda Li, Dingxin Cheng, Xiaoying Tang, Dianbo Sui, Qingbin Liu, Xi Chen, and Kevin Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 3302–3310, 2025

    2025

  17. [17]

    GLM-4.1V-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006, 2025

    Wenyi Hong et al. GLM-4.1V-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Vlm-R3: Region recognition, reasoning, and refinement for enhanced multimodal chain-of-thought.arXiv preprint arXiv:2505.16192, 2025

    Chaoya Jiang, Yongrui Heng, Wei Ye, Han Yang, Haiyang Xu, Ming Yan, Ji Zhang, Fei Huang, and Shikun Zhang. Vlm-R3: Region recognition, reasoning, and refinement for enhanced multimodal chain-of-thought.arXiv preprint arXiv:2505.16192, 2025

    arXiv 2025

  19. [19]

    MME-CoT: Benchmarking chain-of-thought in large multimodal models for reasoning quality, robustness, and efficiency

    DongzhiJiang,RenruiZhang,ZiyuGuo,YanweiLi,YuQi,XinyanChen,LiuhuiWang,Jianhan Jin, Claire Guo, Shen Yan, Bo Zhang, Chaoyou Fu, Peng Gao, and Hongsheng Li. MME-CoT: Benchmarking chain-of-thought in large multimodal models for reasoning quality, robustness, and efficiency. InProceedings of the 42nd International Conference on Machine Learning, pages 27793–27...

    2025

  20. [20]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: Bootstrapping language- image pre-training with frozen image encoders and large language models. InProceedings of the 40th International Conference on Machine Learning, pages 19730–19742, 2023

    2023

  21. [21]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, volume 36, pages 34892–34916, 2023

    2023

  22. [22]

    Improved baselines with visual instructiontuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instructiontuning. InProceedingsoftheIEEE/CVFConferenceonComputerVisionandPattern Recognition, pages 26296–26306, 2024

    2024

  23. [23]

    Let’s think with images efficiently! an interleaved-modal chain-of-thought reasoning framework with dynamic and precise visual thoughts

    Xu Liu, Yongheng Zhang, Qiguang Chen, Yao Li, Sheng Wang, and Libo Qin. Let’s think with images efficiently! an interleaved-modal chain-of-thought reasoning framework with dynamic and precise visual thoughts. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 32213–32221, 2026

    2026

  24. [24]

    Learn to explain: Multimodal reasoning via thought chains for science question answering.Advances in neural information processing systems, 35: 2507–2521, 2022

    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: 2507–2521, 2022

    2022

  25. [25]

    MathVista: Evaluating mathematical reasoning of foundation models in visual contexts

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyang Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. MathVista: Evaluating mathematical reasoning of foundation models in visual contexts. InInternational Conference on Learning Representations, 2024

    2024

  26. [26]

    Visual CoT: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning

    Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual CoT: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. InAdvances in Neural Information Processing Systems, volume 37, pages 8612–8642, 2024

    2024

  27. [27]

    Kimi-vl technical report.arXiv preprint arXiv:2504.07491, 2025

    Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chenzhuang Du, Chu Wei, et al. Kimi-vl technical report.arXiv preprint arXiv:2504.07491, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    Mitigating low-quality reasoning in mllms: Self-driven refined multimodal cot with selective thinking and step-wise visualenhancement

    Chongjun Tu, Peng Ye, Dongzhan Zhou, Tao Chen, and Wanli Ouyang. Mitigating low-quality reasoning in mllms: Self-driven refined multimodal cot with selective thinking and step-wise visualenhancement. InProceedingsoftheAAAIConferenceonArtificialIntelligence,volume40, pages 9576–9584, 2026. 13

    2026

  29. [29]

    Disentangling inter- and intra-video relations for multi-event video-text retrieval and grounding

    Mengzhao Wang, Huafeng Li, Yafei Zhang, Jinxing Li, Dapeng Tao, and Zhengtao Yu. Disentangling inter- and intra-video relations for multi-event video-text retrieval and grounding. IEEE Transactions on Image Processing, 34:7558–7571, 2025

    2025

  30. [30]

    Internvl3

    WeiyunWang,ZhangweiGao,LixinGu,HengjunPu,LongCui,XingguangWei,ZhaoyangLiu, 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

    Pith/arXiv arXiv 2025

  31. [31]

    Timerefine: Temporal grounding with time refining video llm

    XiziWang,FengCheng,ZiyangWang,HuiyuWang,MdMohaiminulIslam,LorenzoTorresani, Mohit Bansal, Gedas Bertasius, and David Crandall. Timerefine: Temporal grounding with time refining video llm. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5067–5078, 2026

    2026

  32. [32]

    Self-consistencyimproveschainofthoughtreasoninginlanguage models

    XuezhiWang,JasonWei,DaleSchuurmans, QuocV.Le,EdH.Chi,SharanNarang, Aakanksha Chowdhery,andDennyZhou. Self-consistencyimproveschainofthoughtreasoninginlanguage models. InInternational Conference on Learning Representations, 2023

    2023

  33. [33]

    Chi, Quoc V

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. InAdvancesinNeuralInformationProcessingSystems,volume35,pages24824–24837, 2022

    2022

  34. [34]

    Llava-cot: Let vision language models reason step-by-step

    Guowei Xu, Peng Jin, Ziang Wu, Hao Li, Yibing Song, Lichao Sun, and Li Yuan. Llava-cot: Let vision language models reason step-by-step. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2087–2098, 2025

    2087

  35. [35]

    Chain-of-thought provably enables learn- ing the (otherwise) unlearnable

    Chenxiao Yang, Zhiyuan Li, and David Wipf. Chain-of-thought provably enables learn- ing the (otherwise) unlearnable. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  36. [36]

    Timeexpert: An expert- guided video llm for video temporal grounding

    Zuhao Yang, Yingchen Yu, Yunqing Zhao, Shijian Lu, and Song Bai. Timeexpert: An expert- guided video llm for video temporal grounding. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 24286–24296, 2025

    2025

  37. [37]

    Mmmu: Amassivemulti-disciplinemultimodal understanding and reasoning benchmark for expert agi

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, DongfuJiang,WeimingRen,YuxuanSun,etal. Mmmu: Amassivemulti-disciplinemultimodal understanding and reasoning benchmark for expert agi. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9556–9567, 2024

    2024

  38. [38]

    Latent sketchpad: Sketching visual thoughts to elicit multimodal reasoning in MLLMs.arXiv preprint arXiv:2510.24514, 2025

    Huanyu Zhang, Wenshan Wu, Chengzu Li, Ning Shang, Yan Xia, Yangyu Huang, Yifan Zhang, Li Dong, Zhang Zhang, Liang Wang, Tieniu Tan, and Furu Wei. Latent sketchpad: Sketching visual thoughts to elicit multimodal reasoning in MLLMs.arXiv preprint arXiv:2510.24514, 2025

    arXiv 2025

  39. [39]

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

    Pith/arXiv arXiv 2023

  40. [40]

    Unsupervised visual chain-of- thought reasoning via preference optimization

    Kesen Zhao, Beier Zhu, Qianru Sun, and Hanwang Zhang. Unsupervised visual chain-of- thought reasoning via preference optimization. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2303–2312, 2025

    2025

  41. [41]

    Cot-vla: Visual chain-of-thought reasoning for vision-language-action models

    Qingqing Zhao, Yao Lu, Moo Jin Kim, Zipeng Fu, Zhuoyang Zhang, Yecheng Wu, Zhaoshuo Li, Qianli Ma, Song Han, Chelsea Finn, et al. Cot-vla: Visual chain-of-thought reasoning for vision-language-action models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 1702–1713, 2025

    2025

  42. [42]

    MiniGPT-4: Enhancing vision-language understanding with advanced large language models.arXiv preprint arXiv:2304.10592, 2023

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models.arXiv preprint arXiv:2304.10592, 2023. 14

    Pith/arXiv arXiv 2023