Pith. sign in

REVIEW 3 major objections 53 references

Vision-language models judge physical reachability by appearance matching and linguistic belief rather than physical process, and inserting visual simulation as an intermediate step corrects much of that failure.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 07:25 UTC pith:3PRBCVFY

load-bearing objection Solid new benchmark that exposes real VLM physical-reasoning failures; the simulator half of the story is under-validated but does not sink the paper. the 3 major comments →

arxiv 2607.11044 v1 pith:3PRBCVFY submitted 2026-07-13 cs.MM

RetroHolmes: When Semantic Plausibility Fails Retrospective Physical Process Reasoning

Ruoxuan Zhang , Qiyun Zheng , Siyu Wu , Ling Zou , Hongxia Xie , Zhiyu Zhou , Jian-Yu Jiang-Lin , Zihan Li
show 7 more authors
Zhengguang Wang Bin Wen Ling Lo Jianlong Fu Meibao Yao Juncheng Hu Wen-Huang Cheng
This is my paper
classification cs.MM
keywords Vision-Language ModelsPhysical Process ReasoningAnalysis-by-SynthesisReachability JudgmentBelief-Conflict SensitivityRetrospective ReasoningWorld ModelingSycophancy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Humans can look at two sparse snapshots of an object and infer whether one could have physically become the other, and how. Current vision-language models produce fluent procedures but do not reliably capture physical realizability: they may invent steps that reverse irreversible changes. This paper defines Retrospective Physical Process Reasoning—decide reachability from an image pair, then reconstruct a physically valid causal chain if reachable—and builds RetroHolmes, a real-world object-centric benchmark spanning four transition types and fourteen domains with reachability labels and step sequences. Evaluations of leading models reveal systematic judgment bias and belief dominance: when a human assertion contradicts the images, models often flip their answer while barely changing what they attend to. A simple analysis-by-synthesis loop, Simulate-and-Verify, that plans, synthesizes intermediate video, then re-verifies, raises accuracy and sharply lowers belief-conflict sensitivity, showing that physically grounded intermediate representations matter for this form of reasoning.

Core claim

State-of-the-art vision-language models systematically fail Retrospective Physical Process Reasoning on RetroHolmes: they show strong judgment bias on whether one object state can physically reach another, and they prioritize contradictory human belief prompts over visual physical evidence, mirroring sycophancy. A Simulate-and-Verify analysis-by-synthesis method that inserts visual simulation between planning and verification improves reachability accuracy and reduces belief-conflict sensitivity, validating the benchmark’s diagnostic value.

What carries the argument

Retrospective Physical Process Reasoning: given sparse visual endpoints (IA, IB), decide whether a physically valid transition exists and, if so, reconstruct the latent causal step sequence. Operationalized by the RetroHolmes benchmark (reachability labels, causal steps, five evaluation dimensions) and diagnosed by the Simulate-and-Verify loop that uses video generation as an intermediate physical simulator.

Load-bearing premise

The method assumes that a pretrained video generation model, conditioned on start and end images plus a textual plan, produces intermediate trajectories faithful enough to real physics that a verifier can trust them for reachability and causal consistency.

What would settle it

Replace the physical simulator with non-physical generation (pure noise or unrelated video) or measure whether belief-induced answer flips still occur with unchanged attention maps on RetroHolmes; if gains vanish under non-physical simulation or bias and flips persist under stronger dynamics training without simulation, the central diagnostic claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Benchmarks that only score forward prediction or linguistic plausibility will overstate physical understanding in vision-language models.
  • Models that follow linguistic priors over visual physics remain unreliable for planning, robotics, and safety-critical visual reasoning.
  • Physically grounded intermediate representations such as simulated trajectories can measurably reduce belief-conflict sensitivity.
  • Evaluation suites need belief-conflict and irreversibility probes to expose shortcut reasoning on invariant visual features.
  • Analysis-by-synthesis with generative world models is a practical route to better retrospective physical reasoning.

Where Pith is reading between the lines

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

  • The same attention-to-invariant-patches shortcut may explain failures on other irreversible or multi-step physical tasks outside this benchmark.
  • Further language-only alignment without dynamics supervision may increase, rather than reduce, belief dominance over evidence.
  • As video simulators improve, closed-loop Simulate-and-Verify could become a standard test-time method for physical question answering and planning.
  • Endpoint reachability judgments of the RetroHolmes type could serve as a diagnostic for any world model that claims to capture physics.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The paper introduces Retrospective Physical Process Reasoning: given sparse visual endpoints (I_A, I_B), a model must judge physical reachability and, if reachable, reconstruct a causal step sequence under physical constraints. It presents RetroHolmes, a real-world object-centric benchmark of 9,530 image pairs spanning four physical-transition categories and 14 domains, with five diagnostic dimensions (ED1–ED5) and metrics for judgment accuracy (JA), belief-conflict sensitivity (BCS), and step completeness (SC via RougeL). Zero-shot evaluation of 11 VLMs reveals systematic judgment bias and belief dominance over visual evidence (sycophancy-like). The authors then propose Simulate-and-Verify, an analysis-by-synthesis loop that uses a video generator (Wan2.2-TI2V-5B) as an intermediate physical simulator, reporting gains in JA, BCS, and SC over the base Qwen3-VL-8B and ablations.

Significance. The work targets a genuine and under-evaluated gap: whether VLMs capture physical realizability rather than linguistic/procedural plausibility. RetroHolmes is a concrete, multi-domain, dual-output benchmark with explicit reachability labels and causal chains, plus a useful diagnostic suite (especially ED3 belief conflict and ED5 irreversible micro-states). The multi-model Table 2 results, attention analyses (Fig. 4), and ablations (SFT-only; w/o Stage 2) provide credible evidence of judgment bias and belief dominance. If the findings hold under stronger step-quality and simulator-fidelity checks, the benchmark would be a lasting diagnostic resource for physical reasoning in VLMs, and the analysis-by-synthesis framing would usefully connect world-model/generation work to feasibility reasoning.

major comments (3)
  1. Section 4.2 (Physical World Simulation) and Table 2: The central claim that Simulate-and-Verify improves reasoning via physically grounded intermediate representations rests on Wan2.2-TI2V-5B producing trajectories that are faithful enough to real dynamics to verify reachability/causal consistency. The paper reports no fidelity metric (human ratings of simulated videos, physical-consistency rate, or correlation between simulator-detected inconsistencies and ground-truth unreachability). The 'Ours w/o Stage 2' ablation shows degradation but does not isolate physical fidelity from the mere addition of denser intermediate visual tokens. Without such evidence, the large BCS drop (50.32% → 7.53%) and JA gains can be explained by extra visual context or the planner–verifier loop rather than genuine physical grounding. Please add a fidelity study or substantially temper claims that attribute ga
  2. Section 3.2 Evaluation Metrics and Table 2 (SC columns): Step Completeness is measured primarily by RougeL against reference text. RougeL rewards lexical/structural overlap and does not assess physical feasibility, causal order correctness, or law violations—the very failure modes the paper diagnoses (law/causality/identity/space/goal errors in Fig. 2). The manuscript mentions a human study of reconstructed sequences but does not report its protocol, scale, or results in the main evaluation. For a paper whose thesis is that semantic plausibility fails physical reasoning, SC as currently reported is too weak to support claims of improved causal reconstruction. Report human ratings (or a physics-aware automatic metric) for step quality, at least for the main models and Simulate-and-Verify.
  3. Section 3.2 (ED1 construction) and taxonomy: Unreachable pairs for ED1 are synthesized via 'intra-category goal swapping.' This is an ad-hoc axiom that is load-bearing for JA on unreachable cases. The paper does not report inter-annotator agreement, expert validation rates, or failure modes of this procedure (e.g., swapped goals that remain physically reachable via alternative processes). Please quantify how often swapped pairs are truly unreachable under the stated physical constraints, and how this affects ED1/ED5 interpretation.

Circularity Check

0 steps flagged

No significant circularity: RetroHolmes labels and metrics are external to the models under test; Simulate-and-Verify gains are measured against those fixed labels, not defined by them.

full rationale

This is an empirical benchmark-and-method paper, not a first-principles derivation. Reachability labels and causal step sequences are constructed from external procedural datasets plus human checks; Judgment Accuracy, Belief-Conflict Sensitivity, and Step Completeness are then scored against those fixed annotations. The multi-model failure modes (judgment bias, belief dominance) are comparative measurements on that held-out suite, including independent closed-source models (GPT-5.2, GPT-4o, Gemini 3 Pro). Simulate-and-Verify is an analysis-by-synthesis pipeline whose reported JA/BCS/SC improvements are likewise evaluated against the same external labels; no free parameter is fitted to a subset of RetroHolmes and then renamed a prediction, and no uniqueness theorem or ansatz is imported from overlapping-author prior work as a load-bearing premise. Mild reuse of Qwen3-VL variants as both baseline and planner/verifier, and of Qwen3-VL-Plus as one of two label verifiers for ED4, is ordinary engineering practice and does not reduce any central claim by construction. The paper is therefore self-contained against its external benchmark; circularity score is 0.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 3 invented entities

As an empirical systems paper the load-bearing content is the benchmark construction and the experimental measurements rather than free parameters or new physical entities. The main assumptions are domain conventions about what counts as a physical transition and the untested fidelity of the chosen video generator as a physics proxy.

free parameters (2)
  • number of intermediate simulation frames N
    Chosen for the latent diffusion simulator; not reported as optimized against a held-out physical metric, yet affects the density of the trajectory used by the verifier.
  • top-10% attention patch threshold for CLIP similarity analysis
    Hand-chosen cutoff used to quantify invariant-feature focus; different percentiles would change the reported similarity gap.
axioms (4)
  • domain assumption A pretrained conditional video generation model encodes sufficiently accurate physical dynamics to serve as an internal simulator for reachability verification.
    Stated as premise in Section 4.2 ('Based on the premise that generative models encapsulate physical processes'); no independent physical validation is supplied.
  • domain assumption Rouge-L against gold step sequences is an adequate proxy for causal step completeness under physical constraints.
    Adopted following prior recipe/instruction work; physical feasibility errors can still score high if wording overlaps.
  • domain assumption Human-annotated reachability labels and step sequences correctly encode real-world physical possibility and minimal causal chains.
    Labels obtained via automatic filtering plus human checks; residual annotation noise would directly affect all reported accuracies.
  • ad hoc to paper Intra-category goal swapping produces valid unreachable pairs for ED1.
    Construction method for negative examples; assumes category membership preserves visual similarity while guaranteeing physical impossibility.
invented entities (3)
  • Retrospective Physical Process Reasoning paradigm no independent evidence
    purpose: Defines the dual task of reachability judgment plus causal chain reconstruction from sparse endpoints.
    New evaluation framing introduced in Section 3.1; not an ontological claim about the world.
  • Five evaluation dimensions ED1–ED5 no independent evidence
    purpose: Diagnose complementary failure modes (basic causal, long-range, belief-conflict, reversibility, irreversible micro-states).
    Taxonomy constructed for the benchmark; diagnostic utility is internal to RetroHolmes.
  • Simulate-and-Verify framework no independent evidence
    purpose: Closed-loop planner–simulator–verifier that grounds VLM reasoning in synthesized intermediate states.
    Architectural instantiation of analysis-by-synthesis for this task; performance gains are measured only on RetroHolmes.

pith-pipeline@v1.1.0-grok45 · 22477 in / 3231 out tokens · 31742 ms · 2026-07-14T07:25:46.578977+00:00 · methodology

0 comments
read the original abstract

Humans can infer hidden physical processes from sparse observations, yet current evaluation protocols for Vision Language Models fail to assess whether such physical reasoning is genuinely captured. To address this gap, we introduce Retrospective Physical Process Reasoning, a new evaluation paradigm to reason backward from outcomes under explicit physical constraints. Building on the paradigm, we present RetroHolmes, the first real-world benchmark for Retrospective Physical Process Reasoning, comprising object-centric image pairs annotated with reachability labels and causal step sequences across diverse physical transitions. Using RetroHolmes, we analyze state of the art Vision Language Models and uncover systematic failure modes, including judgment bias in reachability assessment and belief dominance over physical evidence, mirroring sycophancy behavior observed in large language models. We further demonstrate a simple analysis-by-synthesis instantiation with visual simulation as an intermediate step, validating the diagnostic value of RetroHolmes and highlighting the importance of physically grounded intermediate representations for physical reasoning.

Figures

Figures reproduced from arXiv: 2607.11044 by Bin Wen, Hongxia Xie, Jianlong Fu, Jian-Yu Jiang-Lin, Juncheng Hu, Ling Lo, Ling Zou, Meibao Yao, Qiyun Zheng, Ruoxuan Zhang, Siyu Wu, Wen-Huang Cheng, Zhengguang Wang, Zhiyu Zhou, Zihan Li.

Figure 1
Figure 1. Figure 1: Overview of the RetroHolmes Benchmark. We propose two tasks: World Reachability Judgment and Step Recon [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Evaluation Suite. Five evaluation dimensions (ED1–ED5) are defined for the two tasks to systematically measure [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Physical Transitions (PTs) taxonomy of the RetroHolmes Benchmark. RetroHolmes categorizes real-world object [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Analysis of Cross-Modal Attention Maps. (a) The [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the Simulate-and-Verify Framework. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Comparison of Model Outputs. We com [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Model Performance Comparison on BCS and Re [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗

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

53 extracted references · 17 linked inside Pith

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  2. [2]

    Xiang An, Yin Xie, Kaicheng Yang, Wenkang Zhang, Xiuwei Zhao, Zheng Cheng, Yirui Wang, Songcen Xu, Changrui Chen, Didi Zhu, et al. 2025. Llava-onevision- 1.5: Fully open framework for democratized multimodal training.arXiv preprint arXiv:2509.23661(2025)

    Pith/arXiv arXiv 2025

  3. [3]

    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 ...

    Pith/arXiv arXiv 2025

  4. [4]

    Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez- Opazo, Hongdong Li, and Stephen Gould. 2021. The ikea asm dataset: Understand- ing people assembling furniture through actions, objects and pose. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 847–859

    2021

  5. [5]

    Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xu Huang, Shu Jiang, et al. 2025. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems. arXiv preprint arXiv:2503.06669(2025)

    Pith/arXiv arXiv 2025

  6. [6]

    Prateek Chhikara, Dhiraj Chaurasia, Yifan Jiang, Omkar Masur, and Filip Ilievski

  7. [7]

    InProceedings of the IEEE/CVF winter conference on applications of computer vision

    Fire: Food image to recipe generation. InProceedings of the IEEE/CVF winter conference on applications of computer vision. 8184–8194

  8. [8]

    Xiaowei Chi, Chun-Kai Fan, Hengyuan Zhang, Xingqun Qi, Rongyu Zhang, Anthony Chen, Chi-min Chan, Wei Xue, Qifeng Liu, Shanghang Zhang, et al

  9. [9]

    Eva: An embodied world model for future video anticipation.arXiv preprint arXiv:2410.15461(2024)

    Pith/arXiv arXiv 2024

  10. [10]

    Aditya Chinchure, Sahithya Ravi, Raymond Ng, Vered Shwartz, Boyang Li, and Leonid Sigal. 2025. Black Swan: Abductive and Defeasible Video Reasoning in Unpredictable Events. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2025

  11. [11]

    Wei Chow, Jiageng Mao, Boyi Li, Daniel Seita, Vitor Guizilini, and Yue Wang. 2025. PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding. arXiv:2501.16411 [cs.CV] https://arxiv.org/abs/2501.16411

    Pith/arXiv arXiv 2025

  12. [12]

    Yilun Du, Mengjiao Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Brian Ichter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, et al. 2023. V ideo language planning.arXiv preprint arXiv:2310.10625(2023)

    Pith/arXiv arXiv 2023

  13. [13]

    Aaron Foss, Chloe Evans, Sasha Mitts, Koustuv Sinha, Ammar Rizvi, and Justine T Kao. 2025. Causalvqa: A physically grounded causal reasoning benchmark for video models.arXiv preprint arXiv:2506.09943(2025)

    Pith/arXiv arXiv 2025

  14. [14]

    Jiayi Gao and Huaiwen Zhang. 2025. Evaluating and Mitigating Sycophancy in Large Vision-Language Models. InProceedings of the 33rd ACM International Conference on Multimedia. 11706–11715

    2025

  15. [15]

    Qiyue Gao, Xinyu Pi, Kevin Liu, Junrong Chen, Ruolan Yang, Xinqi Huang, Xinyu Fang, Lu Sun, Gautham Kishore, Bo Ai, et al. 2025. Do vision-language models have internal world models? towards an atomic evaluation.arXiv preprint arXiv:2506.21876(2025)

    Pith/arXiv arXiv 2025

  16. [16]

    Google DeepMind. 2025. A new era of intelligence with Gemini 3. https://blog. google/products-and-platforms/products/gemini/gemini-3/. Accessed: 2026-01- 29

    2025

  17. [17]

    Yanjia Huang, Xianshun Jiang, Xiangbo Gao, Mingyang Wu, and Zhengzhong Tu

  18. [18]

    arXiv preprint arXiv:2512.00041(2025)

    VISTAv2: World Imagination for Indoor Vision-and-Language Navigation. arXiv preprint arXiv:2512.00041(2025)

    arXiv 2025

  19. [19]

    Yanjia Huang, Mingyang Wu, Renjie Li, and Zhengzhong Tu. 2025. Vista: Gen- erative visual imagination for vision-and-language navigation.arXiv preprint arXiv:2505.07868(2025)

    arXiv 2025

  20. [20]

    Harsh Jhamtani and Taylor Berg-Kirkpatrick. 2018. Learning to Describe Differ- ences Between Pairs of Similar Images

    2018

  21. [21]

    Dacheng Li, Yunhao Fang, Yukang Chen, Shuo Yang, Shiyi Cao, Justin Wong, Michael Luo, Xiaolong Wang, Hongxu Yin, Joseph E Gonzalez, et al. 2025. World- ModelBench: Judging Video Generation Models As World Models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2025

  22. [22]

    Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. InText summarization branches out. 74–81

    2004

  23. [23]

    Fei Liu, Shichao Xie, Minghua Luo, Zedong Chu, Junjun Hu, Xiaolong Wu, and Mu Xu. 2025. NavForesee: A Unified Vision-Language World Model for Hi- erarchical Planning and Dual-Horizon Navigation Prediction.arXiv preprint arXiv:2512.01550(2025)

    arXiv 2025

  24. [24]

    Guoshan Liu, Hailong Yin, Bin Zhu, Jingjing Chen, Chong-Wah Ngo, and Yu- Gang Jiang. 2025. Retrieval augmented recipe generation. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). IEEE, 2453–2463

    2025

  25. [25]

    Liang Ma, Jiajun Wen, Min Lin, Rongtao Xu, Xiwen Liang, Bingqian Lin, Jun Ma, Yongxin Wang, Ziming Wei, Haokun Lin, et al. 2025. Phyblock: A progressive benchmark for physical understanding and planning via 3d block assembly. In The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track

    2025

  26. [26]

    Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. 2019. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. InProceedings of the IEEE/CVF international conference on computer vision. 2630–2640

    2019

  27. [27]

    NVIDIA. 2025. Cosmos-Reason2-8B. https://huggingface.co/nvidia/Cosmos- Reason2-8B. Accessed: 2026-01-27

    2025

  28. [28]

    OpenAI. 2025. Introducing GPT-5.2. https://openai.com/index/introducing-gpt- 5-2. Accessed: 2026-01-29

    2025

  29. [29]

    Dong Huk Park, Trevor Darrell, and Anna Rohrbach. 2019. Robust Change Captioning

    2019

  30. [30]

    Akhil Perincherry, Jacob Krantz, and Stefan Lee. 2025. Do visual imaginations improve vision-and-language navigation agents?. InProceedings of the Computer Vision and Pattern Recognition Conference. 3846–3855

    2025

  31. [31]

    Renjie Pi, Kehao Miao, Li Peihang, Runtao Liu, Jiahui Gao, Jipeng Zhang, and Xiaofang Zhou. 2025. Pointing to a llama and call it a camel: On the sycophancy of multimodal large language models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 20177–20191

    2025

  32. [32]

    Li Puyin, Tiange Xiang, Ella Mao, Shirley Wei, Xinye Chen, Adnan Masood, Li Fei-Fei, and Ehsan Adeli. 2025. QuantiPhy: A Quantitative Benchmark Evalu- ating Physical Reasoning Abilities of Vision-Language Models.arXiv preprint arXiv:2512.19526(2025)

    arXiv 2025

  33. [33]

    Tomáš Souček, Prajwal Gatti, Michael Wray, Ivan Laptev, Dima Damen, and Josef Sivic. 2025. Showhowto: Generating scene-conditioned step-by-step visual instructions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 27435–27445

    2025

  34. [34]

    Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. 2019. Coin: A large-scale dataset for comprehensive in- structional video analysis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 1207–1216

    2019

  35. [35]

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, T...

    Pith/arXiv arXiv 2025

  36. [36]

    Qineng Wang, Wenlong Huang, Yu Zhou, Hang Yin, Tianwei Bao, Jianwen Lyu, Weiyu Liu, Ruohan Zhang, Jiajun Wu, Li Fei-Fei, et al . 2026. ENACT: Evaluating Embodied Cognition with World Modeling of Egocentric Interaction. InInternational Conference on Learning Representations

    2026

  37. [37]

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

    Pith/arXiv arXiv 2025

  38. [38]

    Yifan Wang, Shiyu Li, Peiming Li, Xiaochen Yang, Yang Tang, and Zheng Wei

  39. [39]

    Render-of-Thought: Rendering Textual Chain-of-Thought as Images for Visual Latent Reasoning.arXiv preprint arXiv:2601.14750(2026)

    Pith/arXiv arXiv 2026

  40. [40]

    Xiyang Wu, Zongxia Li, Jihui Jin, Guangyao Shi, Gouthaman KV, Vishnu Raj, Nilotpal Sinha, Jingxi Chen, Fan Du, and Dinesh Manocha. 2025. MASS: Motion- Aware Spatial-Temporal Grounding for Physics Reasoning and Comprehension in Vision-Language Models. arXiv:2511.18373 [cs.CV] https://arxiv.org/abs/2511. 18373

    Pith/arXiv arXiv 2025

  41. [41]

    Xianjin Wu, Dingkang Liang, Tianrui Feng, Kui Xia, Yumeng Zhang, Xiaofan Li, Xiao Tan, and Xiang Bai. 2026. Generation Models Know Space: Unleashing Implicit 3D Priors for Scene Understanding.arXiv preprint arXiv:2603.19235 (2026)

    Pith/arXiv arXiv 2026

  42. [42]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

    Pith/arXiv arXiv 2017

  43. [43]

    Yuncong Yang, Jiageng Liu, Zheyuan Zhang, Siyuan Zhou, Reuben Tan, Jianwei Yang, Yilun Du, and Chuang Gan. 2025. MindJourney: Test-Time Scaling with World Models for Spatial Reasoning.arXiv preprint arXiv:2507.12508(2025)

    arXiv 2025

  44. [44]

    Shaofeng Yin, Jiaxin Ge, Zora Zhiruo Wang, Xiuyu Li, Michael J Black, Trevor Darrell, Angjoo Kanazawa, and Haiwen Feng. 2026. Vision-as-Inverse-Graphics Agent via Interleaved Multimodal Reasoning.arXiv preprint arXiv:2601.11109 (2026)

    Pith/arXiv arXiv 2026

  45. [45]

    Shoubin Yu, Yue Zhang, Zun Wang, Jaehong Yoon, Huaxiu Yao, Mingyu Ding, and Mohit Bansal. 2026. When and how much to imagine: Adaptive test-time scaling with world models for visual spatial reasoning.arXiv preprint arXiv:2602.08236 (2026)

    Pith/arXiv arXiv 2026

  46. [46]

    Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. 2023. Magicbrush: A manually annotated dataset for instruction-guided image editing.Advances in Neural Information Processing Systems36 (2023), 31428–31449

    2023

  47. [47]

    Ruoxuan Zhang, Jidong Gao, Bin Wen, Hongxia Xie, Chenming Zhang, Hong- Han Shuai, and Wen-Huang Cheng. 2025. RecipeGen: A Step-Aligned Multimodal Benchmark for Real-World Recipe Generation. InProceedings of the 33rd ACM International Conference on Multimedia(Dublin, Ireland)(MM ’25). Association for Computing Machinery, New York, NY, USA, 12659–12665. doi...

    doi:10.1145/3746027 2025

  48. [48]

    Ruoxuan Zhang, Bin Wen, Hongxia Xie, Yi Yao, Songhan Zuo, Jian-Yu Jiang-Lin, Hong-Han Shuai, and Wen-Huang Cheng. 2025. CookAnything: A Framework for Flexible and Consistent Multi-Step Recipe Image Generation. InProceedings of the 33rd ACM International Conference on Multimedia. 7854–7863

    2025

  49. [49]

    Wenliang Zhao, Yongming Rao, Yansong Tang, Jie Zhou, and Jiwen Lu. 2022. Videoabc: A real-world video dataset for abductive visual reasoning.IEEE Trans- actions on Image Processing31 (2022), 6048–6061

    2022

  50. [50]

    Yunpu Zhao, Rui Zhang, Junbin Xiao, Changxin Ke, Ruibo Hou, Yifan Hao, and Ling Li. 2025. Sycophancy in vision-language models: A systematic analysis and an inference-time mitigation framework.Neurocomputing(2025), 131217

    2025

  51. [51]

    Fengzhe Zhou, Jiannan Huang, Jialuo Li, Deva Ramanan, and Humphrey Shi

  52. [52]

    PAI-Bench: A Comprehensive Benchmark For Physical AI.arXiv preprint arXiv:2512.01989(2025)

    arXiv 2025

  53. [53]

    Luowei Zhou, Chenliang Xu, and Jason Corso. 2018. Towards automatic learning of procedures from web instructional videos. InProceedings of the AAAI conference on artificial intelligence, Vol. 32

    2018