World Models Meet Language Models: On the Complementarity of Concrete and Abstract Reasoning

Jianbing Shen; Wei Tao; Yiwen Guo; Yucheng Zhou

arxiv: 2606.03603 · v1 · pith:ZSBPEEIMnew · submitted 2026-06-02 · 💻 cs.CV · cs.CL

World Models Meet Language Models: On the Complementarity of Concrete and Abstract Reasoning

Yucheng Zhou , Wei Tao , Yiwen Guo , Jianbing Shen This is my paper

Pith reviewed 2026-06-28 10:32 UTC · model grok-4.3

classification 💻 cs.CV cs.CL

keywords world modelsmultimodal language modelsconcrete reasoningabstract reasoningself-distillationvisual rolloutsfuture simulationcontrolled reasoning

0 comments

The pith

A training method lets models decide when to use visual future simulations alongside abstract reasoning, boosting performance on prediction tasks.

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

The paper argues that world models generating concrete visual rollouts and multimodal language models performing abstract reasoning offer complementary strengths for predicting future outcomes from images. The key difficulty is that rollouts can be stochastic or task-irrelevant, so the model must learn to invoke, verify, and integrate them selectively. The authors introduce Privileged-Future On-Policy Self-Distillation to solve this by supplying ground-truth futures and answers only to a teacher during training. This produces a student model that improves at deciding rollout use without privileged information at test time. The method delivers measurable gains on two new benchmarks for spatial lookahead and open-domain physical prediction.

Core claim

The paper establishes that controlled concrete reasoning can be achieved by training via Privileged-Future On-Policy Self-Distillation, where ground-truth future videos and answers serve as teacher-side privileged context to evaluate on-policy trajectories, enabling the deployable student model to invoke, verify, and integrate visual future simulations with abstract reasoning without access to true futures at inference time.

What carries the argument

Privileged-Future On-Policy Self-Distillation (PF-OPSD), which uses teacher-side privileged futures to score and guide the student's on-policy concrete-reasoning trajectories during training.

If this is right

The method produces 10.6% and 10.9% gains over baseline on VRQABench and OpenWorldQA.
Models gain robustness when rollouts are noisy or conflict with the question.
The student can operate effectively on open-domain physical prediction without true futures.
Concrete visual simulation becomes selectively usable rather than always applied or ignored.

Where Pith is reading between the lines

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

The distillation pattern could transfer to other settings where an expensive simulator must be invoked selectively.
It suggests a route to hybrid systems that avoid always running full visual rollouts by learning cheap rejection rules first.
Similar teacher-privileged signals might help in longer-horizon or multi-agent prediction tasks.

Load-bearing premise

That training signals from ground-truth futures available only to the teacher will enable the student to correctly evaluate and decide on its own generated rollouts when true futures are absent at test time.

What would settle it

If the trained model shows no accuracy gain or reduced robustness relative to a baseline that ignores rollouts when tested on VRQABench or OpenWorldQA with deliberately noisy or conflicting rollouts, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2606.03603 by Jianbing Shen, Wei Tao, Yiwen Guo, Yucheng Zhou.

**Figure 1.** Figure 1: Abstract reasoning organizes goals, rules, and questions in language, while concrete reasoning uses world-model rollouts to make possible futures visually explicit; reliable agents must coordinate both capabilities. or plans. Recent MLLMs can organize goals, rules, and alternatives in language (Yin et al., 2024; Yoon et al., 2025; Bai et al., 2025), while video world models can make possible futures visua… view at source ↗

**Figure 2.** Figure 2: Preliminary empirical observations showing simulation inertia and the forced-simulation paradox. Panel (a) reports the no-call probability under optional world-model use, and Panel (b) compares accuracy before and after forced Helios simulation. decides whether simulation is needed, dsim ∼ πθ(· | o, q), (1) and, if dsim = 1, writes a simulation prompt and receives a rollout: psim ∼ πθ(· | o, q, dsim), vˆ … view at source ↗

**Figure 3.** Figure 3: Inference-time controlled concrete reasoning. The student MLLM first decides whether abstract reasoning is sufficient. If simulation is needed, it writes a simulation prompt psim, queries the world model W for a candidate rollout vˆ, verifies the rollout, and decides how much to rely on it alongside abstract reasoning. Rejected rollouts trigger prompt retry up to three simulation attempts; once this per-ex… view at source ↗

**Figure 4.** Figure 4: Cross-world-model diagnostic. Replacing Helios with Wan 2.2 improves both the always-simulate variant and PF-OPSD, while PF-OPSD remains stronger than always simulating with the stronger rollout generator. VRQA overall OWQA overall OWQA hard 0 20 40 60 80 100 Call rate (%) 45.2 38.4 65.5 (a) Selective Calls Always VRQA overall OWQA overall OWQA hard 20 40 60 80 100 Score (%) 80.4 77.3 87.3 42.5 48.2 28.5 … view at source ↗

**Figure 5.** Figure 5: Simulation decision quality. PF-OPSD keeps the overall call rate well below always-simulate while increasing calls on OWQA hard samples and maintaining high agreement with privileged evaluator-derived simulation-help labels. “Reduced” denotes the fraction of unnecessary simulation calls avoided relative to an always-simulate policy, not the complement of Call Rate. prompts unchanged [PITH_FULL_IMAGE:fig… view at source ↗

**Figure 6.** Figure 6: Realized simulation depth and performance under PF-OPSD. Attempts denotes the actual number of worldmodel calls selected by the policy for an example, not an externally imposed global constraint. Rows with one to three attempts in panel (a) sum to the learned call rate; panel (b) shows that repeated retries correspond to harder examples with lower accuracy and higher misjudgment. 90 100 92.5 Rollout Veri… view at source ↗

**Figure 9.** Figure 9: summarizes representative diagnostic subsets in which abstract reasoning and rollouts are conflicting or complementary. It visualizes whether S0 Useful S1 Low-q S2 Wrong S3 Corrupt 0 25 50 75 100 Rate (%) (a) Verification Accept False accept False reject S0 Useful S1 Low-q S2 Wrong S3 Corrupt 60 70 80 90 Accuracy (%) 88.5 65.2 64.5 62.8 (b) Final Accuracy [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗

read the original abstract

World models and multimodal large language models (MLLMs) provide complementary capabilities for predicting future outcomes from static visual observations. World models can generate concrete visual rollouts of possible futures, while MLLMs can reason abstractly over questions, goals, and rules. However, generated rollouts are stochastic and may be visually plausible but task-incorrect, making it necessary to determine when visual simulation is useful, whether a rollout is credible, and how it should influence the final answer. We formulate this problem as controlled concrete reasoning, where a model learns to invoke, verify, and integrate visual future simulation alongside abstract reasoning. To study this setting, we construct two human-verified benchmarks, VRQABench for controllable spatial lookahead and OpenWorldQA for open-domain physical prediction, and propose Privileged-Future On-Policy Self-Distillation (PF-OPSD). During training, PF-OPSD uses ground-truth future videos and answers only as teacher-side privileged context to evaluate on-policy concrete-reasoning trajectories, while the deployable student never observes true futures at test time. Experimental results show that PF-OPSD outperforms baseline by 10.6% and 10.9% on VRQABench and OpenWorldQA, respectively, while increasing robustness to noisy or conflicting rollouts. Our code and dataset are available at https://github.com/yczhou001/PF-OPSD.

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

New benchmarks plus a privileged-teacher distillation method for blending world-model rollouts with MLLM reasoning, but the abstract supplies too little on baselines and verification mechanics to assess the central claim.

read the letter

The paper introduces VRQABench and OpenWorldQA as human-verified testbeds for tasks that need both visual future simulation and abstract reasoning, and it describes PF-OPSD, a distillation procedure that gives the teacher ground-truth futures only during training. That setup is a clean way to train a deployable student that must decide when to run and trust its own rollouts.

The work is clearest on the problem framing: stochastic rollouts can be visually plausible yet wrong, so the model needs an internal check before folding them into the answer. The reported 10-plus percent gains on the two benchmarks and the robustness note are the concrete results.

The main limitation is that the abstract gives no baseline architectures, rollout generation details, filtering rules, dataset sizes, or statistical tests. Without those, the 10.6 % and 10.9 % numbers cannot be weighed against standard training or against simpler prompting tricks. The stress-test worry is also live: it is not obvious from the given description whether the student acquires an independent ability to gate and verify rollouts or simply inherits a smoother overall policy from the teacher signal.

This is aimed at groups already working on world models and multimodal LLMs who want concrete testbeds for the integration problem. The new benchmarks and the explicit train/test asymmetry are useful even if the empirical section needs more controls. It is worth sending to referees.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes Privileged-Future On-Policy Self-Distillation (PF-OPSD) to combine world models for generating concrete visual rollouts with MLLMs for abstract reasoning, formulating the task as controlled concrete reasoning. It introduces two human-verified benchmarks (VRQABench and OpenWorldQA) and claims that PF-OPSD yields 10.6% and 10.9% gains over baselines on these benchmarks while improving robustness to noisy or conflicting rollouts; ground-truth futures and answers are used only as privileged teacher-side context during training, with the student model operating without them at test time.

Significance. If the empirical claims hold with proper controls, the work would demonstrate a practical mechanism for making stochastic visual simulations useful alongside abstract reasoning in a fully deployable model. The release of code and datasets supports reproducibility and is a clear strength.

major comments (2)

[Abstract] Abstract: the abstract states performance numbers and a robustness claim but supplies no information on baseline details, statistical significance, error bars, dataset sizes, or how rollouts are generated and filtered; without these the central empirical claim cannot be evaluated from the given text.
[Method] Method: the privileged-teacher setup is a standard distillation pattern, but the lack of any derivation or external benchmark comparison leaves the independence of the reported gains unverified; in particular, no evidence is supplied that the student acquires an independent capability to decide when to invoke simulation, assess rollout credibility, and integrate results rather than merely inheriting superficial improvements.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point-by-point below, clarifying the contributions of PF-OPSD and indicating where revisions will be made to improve clarity.

read point-by-point responses

Referee: [Abstract] Abstract: the abstract states performance numbers and a robustness claim but supplies no information on baseline details, statistical significance, error bars, dataset sizes, or how rollouts are generated and filtered; without these the central empirical claim cannot be evaluated from the given text.

Authors: We agree that additional context in the abstract would aid evaluation. In the revised version, we will expand the abstract to briefly note the main baselines (standard MLLM and world-model-augmented variants), dataset sizes (VRQABench: 1,200 questions; OpenWorldQA: 2,500 questions), mention of error bars and statistical significance from the reported experiments, and a short description of rollout generation (via the world model) and filtering (credibility scoring during training). revision: yes
Referee: [Method] Method: the privileged-teacher setup is a standard distillation pattern, but the lack of any derivation or external benchmark comparison leaves the independence of the reported gains unverified; in particular, no evidence is supplied that the student acquires an independent capability to decide when to invoke simulation, assess rollout credibility, and integrate results rather than merely inheriting superficial improvements.

Authors: The privileged setup is indeed related to distillation, but PF-OPSD differs through its on-policy formulation: the student generates its own concrete-reasoning trajectories at training time and receives privileged feedback only on those trajectories, enabling it to learn invocation, credibility assessment, and integration policies. The manuscript already provides supporting evidence via ablations showing selective rollout usage and robustness gains specifically under noisy rollouts (not observed in pure distillation baselines). No external benchmarks exist for this exact controlled concrete reasoning task, which is why we introduced the two new human-verified datasets; the independence of gains is verified through controlled comparisons on these benchmarks. revision: no

Circularity Check

0 steps flagged

No significant circularity; empirical method with independent experimental claims

full rationale

The paper proposes PF-OPSD as a training procedure that supplies privileged ground-truth futures only to a teacher during on-policy distillation, then deploys a student without them. Performance gains (10.6%/10.9% on the two benchmarks) are reported as experimental outcomes rather than derived by construction from any equation or self-citation. No mathematical derivation chain, fitted-parameter-as-prediction, uniqueness theorem, or ansatz smuggling appears in the provided text. The setup follows a standard privileged-information distillation pattern whose effectiveness is tested externally on human-verified benchmarks, leaving the central claim self-contained and falsifiable outside any internal reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the method description implies standard assumptions about rollout generation and distillation but does not enumerate them.

pith-pipeline@v0.9.1-grok · 5791 in / 1234 out tokens · 25642 ms · 2026-06-28T10:32:00.564013+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 4 canonical work pages · 1 internal anchor

[1]

Sundar Pichai, Demis Hassabis, and Koray Kavukcuoglu

AAAI Press. Sundar Pichai, Demis Hassabis, and Koray Kavukcuoglu. 2025. A new era of intelligence with gemini 3. Cheng Qian, Emre Can Acikgoz, Bingxuan Li, Xiusi Chen, Yuji Zhang, Bingxiang He, Qinyu Luo, Dilek Hakkani-Tür, Gokhan Tur, Yunzhu Li, and 1 others

2025
[2]

Qwen Team

Current agents fail to leverage world model as tool for foresight.arXiv preprint arXiv:2601.03905. Qwen Team. 2026a. Qwen3.5: Towards native multi- modal agents. Qwen Team. 2026b. Qwen3.6-27B: Flagship-level cod- ing in a 27B dense model. Ronan Riochet, Mario Ynocente Castro, Mathieu Bernard, Adam Lerer, Rob Fergus, Véronique Izard, and Emmanuel Dupoux. 2...

work page arXiv 2018
[3]

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R

Reasoning via video: The first evaluation of video models’ reasoning abilities through maze- solving tasks.arXiv Preprint, abs/2511.15065. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. InThe Eleventh International Conference on Learning Repre...

work page arXiv 2023
[4]

Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models

Self-distilled reasoner: On-policy self- distillation for large language models.arXiv preprint arXiv:2601.18734. Yucheng Zhou, Xiubo Geng, Tao Shen, Chongyang Tao, Guodong Long, Jian-Guang Lou, and Jianbing Shen

work page internal anchor Pith review Pith/arXiv arXiv
[5]

simulation- helps

Thread of thought unraveling chaotic contexts. arXiv preprint arXiv:2311.08734. Yucheng Zhou, Xiang Li, Qianning Wang, and Jian- bing Shen. 2024. Visual in-context learning for large vision-language models. InFindings of the Associa- tion for Computational Linguistics: ACL 2024, pages 15890–15902. 11 A Detailed Experimental Setup Datasets and splits.All t...

work page arXiv 2024
[6]

The question is presented **twice**, each time with the option order independently shuffled
[7]

category

If the probe answers correctly on ** both** attempts, the question is classified as`too_easy`and discarded without reaching the Reviewer. This filter removes questions that can be answered correctly without genuine spatial reasoning. VRQABench Step 5: Reviewer Prompt ## Stage 4 - Reviewer (VLM) **Role.** Receives the puzzle image and the QA list. Evaluate...
[8]

Produce a structured physical scene report that later agents will use to design hard questions
[9]

frame_003 to frame_007

Select the best anchor frame - the single frame that best represents the moment just BEFORE the main physical event reaches its outcome (i.e., initial-condition frame). ============================== PART 1 - SCENE REPORT ============================== Analyse ALL provided frames carefully and fill every section below. [OBJECTS] List every physically rele...
[10]

The structured scene report from SceneAnalyst (text only)
[11]

Your job is to turn each skeleton into a complete multiple-choice QA item with 4 options (A/B/C/D) where ALL four options are physically plausible in the scene context

The 6 question skeletons from QuestionDesigner (text only). Your job is to turn each skeleton into a complete multiple-choice QA item with 4 options (A/B/C/D) where ALL four options are physically plausible in the scene context. ============================== DISTRACTOR QUALITY - this is your primary responsibility ============================== Every wro...
[12]

PHYSICALLY POSSIBLE in the real world under some conditions
[13]

LOCALLY PLAUSIBLE in THIS specific scene - a viewer unfamiliar with the exact physics could reasonably believe it might happen given the visible setup
[14]

slides 10cm past the edge

CLEARLY WRONG when the anchor frame is examined carefully using correct physical reasoning. ABSOLUTELY FORBIDDEN distractor types ( instant reject): [REJECT] Objects floating, levitating, or defying gravity [REJECT] Objects freezing in midair without cause [REJECT] Objects passing through solid surfaces [REJECT] Motion reversing spontaneously without an a...
[15]

Questions that cover different categories (diversity)
[16]

Questions with the strongest distractor quality (all 3 wrong options are tempting)
[17]

selected

Questions that most clearly require future-process simulation to answer. Mark the 3 selected items with "selected": true and the other 3 with "selected ": false. ============================== OUTPUT FORMAT ============================== Return a single JSON object with key " qa_drafts" containing exactly 6 items (ALL text in English): { "qa_drafts": [ { ...
[18]

Consider: physical laws, object properties, visible forces, trajectory

ANSWER CORRECTNESS Using ALL frames (anchor + context), verify that the claimed correct answer is physically accurate. Consider: physical laws, object properties, visible forces, trajectory. Reject if the answer key is wrong or ambiguous
[19]

ANCHOR VALIDITY Look at Image 1 (the anchor frame, filename in INPUT section). Does it show sufficient initial- condition cues for a reasoning model to predict the outcome? Does it NOT already reveal the answer directly? Reject if: (a) Image 1 is a blank/cover shot with no relevant physical content, OR (b) The outcome is already fully visible in Image 1 (...
[20]

DISTRACTOR PHYSICAL PLAUSIBILITY For each wrong option independently: - Is it physically possible in the real world under some conditions? ( MUST be yes) - Could a viewer reasonably think it might happen in this specific scene? (MUST be yes) Reject immediately if ANY distractor: - Defies gravity or levitates without cause - Freezes in midair - Passes thro...
[21]

VISUAL CONSISTENCY Does the question text accurately describe what is shown in the frames? Reject if the question references objects or actions not visible anywhere in the clip
[22]

review_list

CATEGORY ALIGNMENT Does this QA item genuinely test the assigned category's type of reasoning ? (Not a blocking reject - note and suggest correction if misaligned.) ============================== 26 APPROVE if ALL of the following hold: ============================== [PASS] Answer key is correct and unambiguous [PASS] Image 1 (anchor) provides sufficient ...

[1] [1]

Sundar Pichai, Demis Hassabis, and Koray Kavukcuoglu

AAAI Press. Sundar Pichai, Demis Hassabis, and Koray Kavukcuoglu. 2025. A new era of intelligence with gemini 3. Cheng Qian, Emre Can Acikgoz, Bingxuan Li, Xiusi Chen, Yuji Zhang, Bingxiang He, Qinyu Luo, Dilek Hakkani-Tür, Gokhan Tur, Yunzhu Li, and 1 others

2025

[2] [2]

Qwen Team

Current agents fail to leverage world model as tool for foresight.arXiv preprint arXiv:2601.03905. Qwen Team. 2026a. Qwen3.5: Towards native multi- modal agents. Qwen Team. 2026b. Qwen3.6-27B: Flagship-level cod- ing in a 27B dense model. Ronan Riochet, Mario Ynocente Castro, Mathieu Bernard, Adam Lerer, Rob Fergus, Véronique Izard, and Emmanuel Dupoux. 2...

work page arXiv 2018

[3] [3]

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R

Reasoning via video: The first evaluation of video models’ reasoning abilities through maze- solving tasks.arXiv Preprint, abs/2511.15065. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. InThe Eleventh International Conference on Learning Repre...

work page arXiv 2023

[4] [4]

Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models

Self-distilled reasoner: On-policy self- distillation for large language models.arXiv preprint arXiv:2601.18734. Yucheng Zhou, Xiubo Geng, Tao Shen, Chongyang Tao, Guodong Long, Jian-Guang Lou, and Jianbing Shen

work page internal anchor Pith review Pith/arXiv arXiv

[5] [5]

simulation- helps

Thread of thought unraveling chaotic contexts. arXiv preprint arXiv:2311.08734. Yucheng Zhou, Xiang Li, Qianning Wang, and Jian- bing Shen. 2024. Visual in-context learning for large vision-language models. InFindings of the Associa- tion for Computational Linguistics: ACL 2024, pages 15890–15902. 11 A Detailed Experimental Setup Datasets and splits.All t...

work page arXiv 2024

[6] [6]

The question is presented **twice**, each time with the option order independently shuffled

[7] [7]

category

If the probe answers correctly on ** both** attempts, the question is classified as`too_easy`and discarded without reaching the Reviewer. This filter removes questions that can be answered correctly without genuine spatial reasoning. VRQABench Step 5: Reviewer Prompt ## Stage 4 - Reviewer (VLM) **Role.** Receives the puzzle image and the QA list. Evaluate...

[8] [8]

Produce a structured physical scene report that later agents will use to design hard questions

[9] [9]

frame_003 to frame_007

Select the best anchor frame - the single frame that best represents the moment just BEFORE the main physical event reaches its outcome (i.e., initial-condition frame). ============================== PART 1 - SCENE REPORT ============================== Analyse ALL provided frames carefully and fill every section below. [OBJECTS] List every physically rele...

[10] [10]

The structured scene report from SceneAnalyst (text only)

[11] [11]

Your job is to turn each skeleton into a complete multiple-choice QA item with 4 options (A/B/C/D) where ALL four options are physically plausible in the scene context

The 6 question skeletons from QuestionDesigner (text only). Your job is to turn each skeleton into a complete multiple-choice QA item with 4 options (A/B/C/D) where ALL four options are physically plausible in the scene context. ============================== DISTRACTOR QUALITY - this is your primary responsibility ============================== Every wro...

[12] [12]

PHYSICALLY POSSIBLE in the real world under some conditions

[13] [13]

LOCALLY PLAUSIBLE in THIS specific scene - a viewer unfamiliar with the exact physics could reasonably believe it might happen given the visible setup

[14] [14]

slides 10cm past the edge

CLEARLY WRONG when the anchor frame is examined carefully using correct physical reasoning. ABSOLUTELY FORBIDDEN distractor types ( instant reject): [REJECT] Objects floating, levitating, or defying gravity [REJECT] Objects freezing in midair without cause [REJECT] Objects passing through solid surfaces [REJECT] Motion reversing spontaneously without an a...

[15] [15]

Questions that cover different categories (diversity)

[16] [16]

Questions with the strongest distractor quality (all 3 wrong options are tempting)

[17] [17]

selected

Questions that most clearly require future-process simulation to answer. Mark the 3 selected items with "selected": true and the other 3 with "selected ": false. ============================== OUTPUT FORMAT ============================== Return a single JSON object with key " qa_drafts" containing exactly 6 items (ALL text in English): { "qa_drafts": [ { ...

[18] [18]

Consider: physical laws, object properties, visible forces, trajectory

ANSWER CORRECTNESS Using ALL frames (anchor + context), verify that the claimed correct answer is physically accurate. Consider: physical laws, object properties, visible forces, trajectory. Reject if the answer key is wrong or ambiguous

[19] [19]

ANCHOR VALIDITY Look at Image 1 (the anchor frame, filename in INPUT section). Does it show sufficient initial- condition cues for a reasoning model to predict the outcome? Does it NOT already reveal the answer directly? Reject if: (a) Image 1 is a blank/cover shot with no relevant physical content, OR (b) The outcome is already fully visible in Image 1 (...

[20] [20]

DISTRACTOR PHYSICAL PLAUSIBILITY For each wrong option independently: - Is it physically possible in the real world under some conditions? ( MUST be yes) - Could a viewer reasonably think it might happen in this specific scene? (MUST be yes) Reject immediately if ANY distractor: - Defies gravity or levitates without cause - Freezes in midair - Passes thro...

[21] [21]

VISUAL CONSISTENCY Does the question text accurately describe what is shown in the frames? Reject if the question references objects or actions not visible anywhere in the clip

[22] [22]

review_list

CATEGORY ALIGNMENT Does this QA item genuinely test the assigned category's type of reasoning ? (Not a blocking reject - note and suggest correction if misaligned.) ============================== 26 APPROVE if ALL of the following hold: ============================== [PASS] Answer key is correct and unambiguous [PASS] Image 1 (anchor) provides sufficient ...