arxiv: 2605.01657 · v1 · submitted 2026-05-03 · 💻 cs.CV

Recognition: 3 theorem links

· Lean Theorem

Act2See: Emergent Active Visual Perception for Video Reasoning

Aditya Agrawal, Louis-Philippe Morency, Martin Q. Ma, Paul Pu Liang, Ruslan Salakhutdinov, Willis Guo, Yuxiao Qu

Pith reviewed 2026-05-08 19:30 UTC · model grok-4.3

classification 💻 cs.CV

keywords active visual perceptionvideo reasoningvision-language modelschain-of-thoughtsupervised fine-tuningemergent capabilitiesframe synthesis

0 comments

The pith

VLMs can learn to actively retrieve or synthesize video frames mid-reasoning by fine-tuning on verified active CoT traces.

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

The paper shows that standard VLMs are limited because they start with a fixed set of frames and cannot pull new visual information as their reasoning chain develops. Act2See trains the model on carefully constructed reasoning traces that contain explicit calls to fetch existing frames or generate hypothetical ones, all checked against human-written chains. After this supervised fine-tuning, the model at test time spontaneously decides when and which frames to bring in, rather than hallucinating or stopping at the initial view. This matters for tasks where understanding evolves with new evidence, such as tracking events across a video or imagining what would happen under different conditions.

Core claim

Act2See trains VLMs through supervised fine-tuning on high-quality reasoning traces generated by a frontier model; each trace interleaves text steps with active calls to retrieve existing video frames or synthesize new ones, and every trace is verified against human-annotated chains of thought. The resulting models exhibit emergent active visual perception: during inference they autonomously determine when to search for or generate the visual evidence needed to continue or correct their reasoning, rather than remaining restricted to the initial static frames.

What carries the argument

Act2See framework of supervised fine-tuning on verified reasoning traces that embed explicit active calls to retrieve or synthesize video frames inside text CoTs.

If this is right

The model can now handle counterfactual and hypothetical video scenarios by synthesizing frames on demand.
Performance improves on benchmarks that reward dynamic evidence gathering, such as VideoEspresso and ViTIB.
The same training recipe yields gains on EgoNormia and VCR-Bench even against larger models that lack active perception.
Reasoning chains become more interpretable because each active frame call is explicit in the output trace.

Where Pith is reading between the lines

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

The approach could be extended to live video streams by letting the model request the next relevant clip instead of waiting for a full upload.
Similar active-perception training might transfer to other modalities where evidence must be acquired on the fly, such as tool-use or embodied agents.
If the verification step against human CoTs is relaxed, the method might still work but could increase the risk of inherited errors.

Load-bearing premise

That fine-tuning on frontier-generated and human-verified traces will produce reliable active frame calls at inference without copying over the teacher model's biases or hallucinations.

What would settle it

Run the fine-tuned model on a video reasoning question whose correct answer requires a frame that was never shown in the initial input; measure whether the model emits a correct retrieval or synthesis call before answering, versus hallucinating the missing content.

Figures

Figures reproduced from arXiv: 2605.01657 by Aditya Agrawal, Louis-Philippe Morency, Martin Q. Ma, Paul Pu Liang, Ruslan Salakhutdinov, Willis Guo, Yuxiao Qu.

**Figure 1.** Figure 1: The two-round construction of the interleaved video-text CoT dataset. In the first round, Gemini 2.5 Pro is prompted to perform view at source ↗

**Figure 3.** Figure 3: An example of the SFT dataset, consisting of the video, view at source ↗

**Figure 4.** Figure 4: Qualitative examples showing the emergent capability of A view at source ↗

**Figure 5.** Figure 5: Qualitative examples (full CoT) showing the emergent capability of A view at source ↗

**Figure 6.** Figure 6: ACT2SEE pushes the Pareto frontier beyond Qwen3- VL-3B baseline by achieving higher accuracy with lower latency. F. Further Analyses Below we include further analyses including latency and FLOPs, when image generation will help reasoning, categorization of image generation failure inside CoTs and their impact, and model’s robustness towards generation failure view at source ↗

read the original abstract

Vision-Language Models (VLMs) typically rely on static initial frames for video reasoning, restricting their ability to incorporate essential dynamic information as the reasoning process evolves. Existing methods that augment Chain-of-Thought (CoT) with additional frame information often exhibit suboptimal CoT quality and lack the crucial ability to synthesize visual information for hypothetical or counterfactual scenarios. We introduce Act-to-See (Act2See), a novel framework that enables active visual perception by empowering VLMs to actively interleave video frames within text CoTs. Act2See is developed via Supervised Fine-Tuning (SFT) on a high-quality dataset of reasoning traces generated by a frontier VLM. These traces integrate active calls to either retrieve existing frames or generate new ones, and are rigorously verified against human-annotated CoTs to ensure quality. This approach cultivates an emergent capability: at inference time, the model actively determines when to search for or synthesize the necessary visual evidence. Act2See establishes new state-of-the-art results on challenging benchmarks, including VideoEspresso and ViTIB, and outperforms comparable or larger models on Video-MME, EgoNormia, and VCR-Bench, demonstrating an advancement in enabling VLMs with active visual perception for video reasoning.

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.

Desk Editor's Note private letter to a colleague

Act2See trains a VLM on frontier-model traces with human-checked retrieve/generate actions so it can interleave extra frames inside its own CoT, and reports gains on several video benchmarks, but the setup leaves open how much is real active perception versus copied teacher behavior.

read the letter

The main thing to know is that this paper trains a VLM to insert calls for real or synthetic video frames right inside its chain-of-thought steps. They do it by supervised fine-tuning on traces produced by a larger model, where those traces already contain explicit retrieve or generate actions, and the traces were checked against human-annotated reasoning. At inference the fine-tuned model then decides on its own when to pull or make up visual evidence. That interleaving inside the trace is the concrete addition over earlier static-frame or post-hoc methods. The reported results show it reaching new highs on VideoEspresso and ViTIB while beating comparable or larger models on Video-MME, EgoNormia, and VCR-Bench. Those numbers are the clearest positive signal. The experimental gaps are the main concern. The abstract supplies no ablations, no controls for the contribution of retrieval versus generation, and no statistical tests, so it is hard to tell whether the gains survive without the teacher signal. The generate-new-frames route is especially exposed: human verification of the trace does not guarantee the synthesized content is accurate or that the decision to synthesize was optimal. If the student model simply reproduces the teacher’s call frequency and hallucination patterns, the claimed emergent active perception reduces to distillation. This work is aimed at groups already building VLMs for video reasoning in robotics or analysis settings. The method is straightforward enough and the benchmark numbers are strong enough that it should go to referees, where the authors can add the missing controls and test whether the active behavior holds up without the teacher traces.

Referee Report

3 major / 1 minor

Summary. The paper introduces Act2See, a framework that enables VLMs to perform active visual perception for video reasoning by interleaving explicit calls to retrieve existing frames or generate new (hypothetical/counterfactual) frames within text-based chain-of-thought traces. The model is trained via supervised fine-tuning on a dataset of such traces produced by a frontier VLM and rigorously verified against human-annotated CoTs. At inference the fine-tuned model is claimed to autonomously decide when and how to invoke these actions, yielding new state-of-the-art results on VideoEspresso and ViTIB while outperforming comparable or larger models on Video-MME, EgoNormia, and VCR-Bench.

Significance. If the reported gains are shown to arise from genuine emergent active perception rather than distillation artifacts, the work would meaningfully advance VLM capabilities for dynamic, evidence-seeking video reasoning. The use of human-verified frontier-VLM traces for dataset construction is a concrete strength that improves data quality over purely synthetic or unverified sources. The significance remains conditional on stronger evidence that the 'generate new ones' pathway produces accurate synthesized frames and that frame-selection decisions are not simply inherited from the teacher.

major comments (3)

[Abstract] Abstract: The claim of new state-of-the-art results on VideoEspresso and ViTIB (and outperformance on Video-MME, EgoNormia, VCR-Bench) is presented without any mention of statistical significance testing, multiple-run variance, or ablation controls on the active-perception components; this information is load-bearing for the central claim that the gains reflect emergent active perception.
[Method] Training and inference sections: The 'generate new ones' pathway for hypothetical or counterfactual frames is central to the active-perception claim, yet the manuscript provides no quantitative evaluation of the accuracy or utility of the synthesized frames at inference time, nor any comparison isolating generation versus retrieval; without this, it is impossible to rule out that performance improvements are artifacts of teacher-model biases propagated through SFT.
[Experiments] Experiments: No ablation is reported that compares the fine-tuned Act2See model against direct imitation of the teacher VLM's frame-selection policy or against a version trained only on retrieval actions; such a control is required to substantiate that the observed benchmark gains arise from learned active decision-making rather than supervised imitation.

minor comments (1)

[Abstract] The abstract and method sections use the term 'emergent capability' without a precise operational definition distinguishing it from behavior acquired through SFT; a short clarifying sentence would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights important areas for strengthening the evidence of emergent active perception in Act2See. We address each major comment point-by-point below, committing to targeted revisions that directly respond to the concerns while preserving the core contributions of the work.

read point-by-point responses

Referee: [Abstract] Abstract: The claim of new state-of-the-art results on VideoEspresso and ViTIB (and outperformance on Video-MME, EgoNormia, VCR-Bench) is presented without any mention of statistical significance testing, multiple-run variance, or ablation controls on the active-perception components; this information is load-bearing for the central claim that the gains reflect emergent active perception.

Authors: We agree that the abstract claims would be more robust with explicit reference to statistical support. In the revised manuscript, we will expand the Experiments section to report results averaged over multiple independent runs (with means and standard deviations), include statistical significance testing where applicable, and add ablations isolating active-perception components. The abstract will be lightly updated to reference these supporting analyses in the main text, respecting length limits while addressing the load-bearing concern. revision: partial
Referee: [Method] Training and inference sections: The 'generate new ones' pathway for hypothetical or counterfactual frames is central to the active-perception claim, yet the manuscript provides no quantitative evaluation of the accuracy or utility of the synthesized frames at inference time, nor any comparison isolating generation versus retrieval; without this, it is impossible to rule out that performance improvements are artifacts of teacher-model biases propagated through SFT.

Authors: The referee correctly notes this evaluation gap for the generation pathway. Although training traces were human-verified, we did not quantify synthesized frame fidelity or isolate generation at inference. In the revision, we will add a dedicated analysis subsection using perceptual similarity metrics and available ground-truth counterfactuals from benchmarks to evaluate generated frame accuracy and utility. We will also include an ablation comparing the full model against a retrieval-only variant to separate the contributions and mitigate concerns about teacher bias propagation. revision: yes
Referee: [Experiments] Experiments: No ablation is reported that compares the fine-tuned Act2See model against direct imitation of the teacher VLM's frame-selection policy or against a version trained only on retrieval actions; such a control is required to substantiate that the observed benchmark gains arise from learned active decision-making rather than supervised imitation.

Authors: We acknowledge that distinguishing learned active decision-making from direct imitation is essential for the 'emergent' claim. The current setup relies on verified CoT traces and autonomous inference behavior, but lacks these specific controls. In the revised Experiments section, we will add ablations comparing Act2See to (i) a model trained via direct imitation of the teacher's frame-selection policy and (ii) a retrieval-only training variant. These will help demonstrate that benchmark gains stem from the learned active perception capability rather than pure supervised imitation. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical SFT on external verified traces with independent benchmark evaluation

full rationale

The paper's core procedure is supervised fine-tuning on reasoning traces produced by an external frontier VLM and cross-checked against separate human-annotated CoTs. The claimed 'emergent' active-perception behavior at inference is the direct learned output of that training objective rather than a derived prediction that collapses back onto the inputs by construction. No equations, uniqueness theorems, self-citations, or fitted parameters are invoked to justify the results; performance is reported on held-out benchmarks (VideoEspresso, ViTIB, Video-MME, etc.) that are not used in training. This satisfies the default expectation of a non-circular empirical ML paper.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Ledger derived from abstract only; full paper would detail training hyperparameters and dataset construction. No invented physical entities; relies on standard ML assumptions.

free parameters (1)

SFT hyperparameters and dataset size
Standard fine-tuning choices and scale of generated traces not specified in abstract but required for the method.

axioms (1)

domain assumption Frontier VLM generates high-quality, verifiable reasoning traces suitable for SFT
Invoked to create the training dataset from which emergent behavior is claimed.

pith-pipeline@v0.9.0 · 5541 in / 1149 out tokens · 51566 ms · 2026-05-08T19:30:23.872708+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Cost.FunctionalEquation (J(x)=½(x+x⁻¹)−1) washburn_uniqueness_aczel unclear
J(θ) = − (1/N) Σ Σ log P(y_t | V, Q, y_<t; θ) ... We compute the token-level losses over the whole rollout, including sequences of retrieval or generation queries, only excluding retrieved or generated frames.
Foundation (parameter-free forcing chain) reality_from_one_distinction unclear
We fine-tune with LoRA with 8 A100 GPUs, using a learning rate of 2.5×10^−6, a batch size of 1, and a single epoch.

Reference graph

Works this paper leans on

36 extracted references · 22 canonical work pages · 6 internal anchors

[1]

Qwen3-VL Technical Report

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. 1, 2, 5, 6

work page Pith review arXiv 2025
[2]

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025. 1, 5, 6, 4

work page Pith review arXiv 2025
[3]

Active perception.Proceedings of the IEEE, 76(8):966–1005, 1988

Ruzena Bajcsy. Active perception.Proceedings of the IEEE, 76(8):966–1005, 1988. 1, 8

1988
[4]

Qualitative failures of image generation models and their application in detecting deepfakes.Image and Vi- sion Computing, 137:104771, 2023

Ali Borji. Qualitative failures of image generation models and their application in detecting deepfakes.Image and Vi- sion Computing, 137:104771, 2023. 7

2023
[5]

M3-embedding: Multi- linguality, multi-functionality, multi-granularity text embed- dings through self-knowledge distillation

Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. M3-embedding: Multi- linguality, multi-functionality, multi-granularity text embed- dings through self-knowledge distillation. InFindings of the association for computational linguistics: ACL 2024, pages 2318–2335, 2024. 4, 3

2024
[6]

Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling

Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhang- wei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test- time scaling.arXiv preprint arXiv:2412.05271, 2024. 1, 2, 5, 6

work page internal anchor Pith review arXiv 2024
[7]

Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blis- tein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025. 1, 4, 5, 6

work page Pith review arXiv 2025
[8]

Scaling recti- fied flow transformers for high-resolution image synthesis

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning,
[9]

Video- of-thought: Step-by-step video reasoning from perception to cognition.arXiv preprint arXiv:2501.03230, 2024

Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. arXiv preprint arXiv:2501.03230, 2024. 8

work page arXiv 2024
[10]

Video-R1: Reinforcing Video Reasoning in MLLMs

Kaituo Feng, Kaixiong Gong, Bohao Li, Zonghao Guo, Yibing Wang, Tianshuo Peng, Junfei Wu, Xiaoying Zhang, Benyou Wang, and Xiangyu Yue. Video-r1: Reinforcing video reasoning in mllms.arXiv preprint arXiv:2503.21776,

work page internal anchor Pith review arXiv
[11]

Causalvqa: A physically grounded causal reasoning benchmark for video models.arXiv preprint arXiv:2506.09943, 2025

Aaron Foss, Chloe Evans, Sasha Mitts, Koustuv Sinha, Am- mar Rizvi, and Justine T Kao. Causalvqa: A physically grounded causal reasoning benchmark for video models. arXiv preprint arXiv:2506.09943, 2025. 1, 2, 4, 5, 7

work page arXiv 2025
[12]

Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis

Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 24108–24118, 2025. 2, 5, 7, 4

2025
[13]

Famemind: Frame-interleaved video reasoning via reinforcement learning.arXiv e-prints, pages arXiv– 2509, 2025

Haonan Ge, Yiwei Wang, Kai-Wei Chang, Hang Wu, and Yujun Cai. Famemind: Frame-interleaved video reasoning via reinforcement learning.arXiv e-prints, pages arXiv– 2509, 2025. 1, 2, 5, 6, 7, 8

2025
[14]

Chain-of-Frames: Advancing Video Understanding in Multimodal LLMs via Frame-Aware Reasoning

Sara Ghazanfari, Francesco Croce, Nicolas Flammarion, Prashanth Krishnamurthy, Farshad Khorrami, and Siddharth Garg. Chain-of-frames: Advancing video understanding in multimodal llms via frame-aware reasoning.arXiv preprint arXiv:2506.00318, 2025. 1, 5, 6, 8

work page internal anchor Pith review Pith/arXiv arXiv 2025
[15]

Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives

Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 193...

2024
[16]

Videoespresso: A large-scale chain-of-thought dataset for fine-grained video reasoning via core frame selection

Songhao Han, Wei Huang, Hairong Shi, Le Zhuo, Xiu Su, Shifeng Zhang, Xu Zhou, Xiaojuan Qi, Yue Liao, and Si Liu. Videoespresso: A large-scale chain-of-thought dataset for fine-grained video reasoning via core frame selection. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 26181–26191, 2025. 1, 2, 3, 4, 5, 7

2025
[17]

LoRA: Low-Rank Adaptation of Large Language Models

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arxiv 2021.arXiv preprint arXiv:2106.09685, 10, 2021. 5

work page internal anchor Pith review arXiv 2021
[18]

Cos: Chain-of-shot prompting for long video under- standing.arXiv preprint arXiv:2502.06428, 2025

Jian Hu, Zixu Cheng, Chenyang Si, Wei Li, and Shaogang Gong. Cos: Chain-of-shot prompting for long video under- standing.arXiv preprint arXiv:2502.06428, 2025. 1, 2, 5, 6, 8

work page arXiv 2025
[19]

GPT-4o System Card

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perel- man, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli- hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 5, 6

work page Pith review arXiv 2024
[20]

Interac- tive sketchpad: A multimodal tutoring system for collabora- tive, visual problem-solving

Jimin Lee, Steven-Shine Chen, and Paul Pu Liang. Interac- tive sketchpad: A multimodal tutoring system for collabora- tive, visual problem-solving. InProceedings of the Extended Abstracts of the CHI Conference on Human Factors in Com- puting Systems, pages 1–14, 2025. 8

2025
[21]

Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InIn- ternational conference on machine learning, pages 19730– 19742. PMLR, 2023. 4, 2

2023
[22]

Eyes can deceive: Benchmarking counterfac- tual reasoning abilities of multi-modal large language mod- els.CoRR, 2024

Yian Li, Wentao Tian, Yang Jiao, Jingjing Chen, and Yu- Gang Jiang. Eyes can deceive: Benchmarking counterfac- tual reasoning abilities of multi-modal large language mod- els.CoRR, 2024. 1

2024
[23]

Foundations & trends in multimodal machine learning: Prin- ciples, challenges, and open questions.ACM Computing Sur- veys, 56(10):1–42, 2024

Paul Pu Liang, Amir Zadeh, and Louis-Philippe Morency. Foundations & trends in multimodal machine learning: Prin- ciples, challenges, and open questions.ACM Computing Sur- veys, 56(10):1–42, 2024. 8

2024
[24]

Social genome: Grounded social rea- soning abilities of multimodal models.arXiv preprint arXiv:2502.15109, 2025

Leena Mathur, Marian Qian, Paul Pu Liang, and Louis- Philippe Morency. Social genome: Grounded social rea- soning abilities of multimodal models.arXiv preprint arXiv:2502.15109, 2025. 2, 4, 5, 7, 1

work page arXiv 2025
[25]

Xuefei Ning, Guohao Dai, Haoli Bai, Lu Hou, Yu Wang, and Qun Liu

Arsha Nagrani, Sachit Menon, Ahmet Iscen, Shyamal Buch, Ramin Mehran, Nilpa Jha, Anja Hauth, Yukun Zhu, Carl V ondrick, Mikhail Sirotenko, et al. Minerva: Evaluating complex video reasoning.arXiv preprint arXiv:2505.00681,

work page arXiv
[26]

Vcr-bench: A comprehensive evaluation frame- work for video chain-of-thought reasoning.arXiv preprint arXiv:2504.07956, 2025

Yukun Qi, Yiming Zhao, Yu Zeng, Xikun Bao, Wenxuan Huang, Lin Chen, Zehui Chen, Jie Zhao, Zhongang Qi, and Feng Zhao. Vcr-bench: A comprehensive evaluation frame- work for video chain-of-thought reasoning.arXiv preprint arXiv:2504.07956, 2025. 2, 5, 7, 4

work page arXiv 2025
[27]

Egonormia: Benchmarking physical social norm understanding.arXiv preprint arXiv:2502.20490, 2025

MohammadHossein Rezaei, Yicheng Fu, Phil Cuvin, Caleb Ziems, Yanzhe Zhang, Hao Zhu, and Diyi Yang. Egonormia: Benchmarking physical social norm understanding.arXiv preprint arXiv:2502.20490, 2025. 1, 2, 5, 4

work page arXiv 2025
[28]

Openthinkimg: Learning to think with images via visual tool reinforcement learning.arXiv preprint arXiv:2505.08617, 2025

Zhaochen Su, Linjie Li, Mingyang Song, Yunzhuo Hao, Zhengyuan Yang, Jun Zhang, Guanjie Chen, Jiawei Gu, Jun- tao Li, Xiaoye Qu, et al. Openthinkimg: Learning to think with images via visual tool reinforcement learning.arXiv preprint arXiv:2505.08617, 2025. 1, 8

work page arXiv 2025
[29]

Videorft: Incentivizing video reasoning capability in mllms via reinforced fine-tuning.arXiv preprint arXiv:2505.12434, 2025

Qi Wang, Yanrui Yu, Ye Yuan, Rui Mao, and Tianfei Zhou. Videorft: Incentivizing video reasoning capability in mllms via reinforced fine-tuning.arXiv preprint arXiv:2505.12434,

work page arXiv
[30]

Videocot: A video chain-of- thought dataset with active annotation tool.arXiv preprint arXiv:2407.05355, 2024

Yan Wang, Yawen Zeng, Jingsheng Zheng, Xiaofen Xing, Jin Xu, and Xiangmin Xu. Videocot: A video chain-of- thought dataset with active annotation tool.arXiv preprint arXiv:2407.05355, 2024. 1, 8

work page arXiv 2024
[31]

Chain-of-thought prompting elicits reasoning in large lan- guage models.Advances in neural information processing systems, 35:24824–24837, 2022

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large lan- guage models.Advances in neural information processing systems, 35:24824–24837, 2022. 1

2022
[32]

VideoLLaMA 3: Frontier Multimodal Foundation Models for Image and Video Understanding

Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multi- modal foundation models for image and video understand- ing.arXiv preprint arXiv:2501.13106, 2025. 1, 2, 5, 6

work page internal anchor Pith review arXiv 2025
[33]

Rewatch- r1: Boosting complex video reasoning in large vision- language models through agentic data synthesis.arXiv preprint arXiv:2509.23652, 2025

Congzhi Zhang, Zhibin Wang, Yinchao Ma, Jiawei Peng, Yi- han Wang, Qiang Zhou, Jun Song, and Bo Zheng. Rewatch- r1: Boosting complex video reasoning in large vision- language models through agentic data synthesis.arXiv preprint arXiv:2509.23652, 2025. 1, 2, 5, 6, 7, 8

work page arXiv 2025
[34]

Vitcot: Video-text in- terleaved chain-of-thought for boosting video understanding in large language models.arXiv preprint arXiv:2507.09876,

Yongheng Zhang, Xu Liu, Ruihan Tao, Qiguang Chen, Hao Fei, Wanxiang Che, and Libo Qin. Vitcot: Video-text in- terleaved chain-of-thought for boosting video understanding in large language models.arXiv preprint arXiv:2507.09876,

work page arXiv
[35]

Training-free video temporal grounding using large-scale pre-trained models

Minghang Zheng, Xinhao Cai, Qingchao Chen, Yuxin Peng, and Yang Liu. Training-free video temporal grounding using large-scale pre-trained models. InEuropean Conference on Computer Vision, pages 20–37. Springer, 2024. 4, 5

2024
[36]

DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning

Ziwei Zheng, Michael Yang, Jack Hong, Chenxiao Zhao, Guohai Xu, Le Yang, Chao Shen, and Xing Yu. Deep- eyes: Incentivizing” thinking with images” via reinforce- ment learning.arXiv preprint arXiv:2505.14362, 2025. 1, 8 ACT2SEE: Emergent Active Visual Perception for Video Reasoning Supplementary Material A. Impact statement The ACT2SEEframework significant...

work page internal anchor Pith review arXiv 2025