跳转至

Video reasoning literature and citation audit: from Zero-Shot Learners to VBVR

  • Audit date: 2026-08-29
  • Purpose: support the standalone review in docs/video-reasoning.md, using Video models are zero-shot learners and reasoners as the narrative origin and VBVR as a later infrastructure node
  • Scope: reasoning through video generation, latent/diffusion visual reasoning, process-aware evaluation, verifiable training, collaborative inference, and the boundary with world models. “Reasoning about an existing video” is included only when it changes how the generation-side evidence should be interpreted.

1. Search and verification method

The review used a layered search rather than treating one citation database as complete.

  1. Start from the official v2 HTML/PDF of Video models are zero-shot learners and reasoners, and separate its stated hypothesis, system setup, quantitative results, qualitative demonstrations, and limitations.
  2. Trace backward to controlled world-model scaling, continuous visual transformation, and visualized-thought work that made the hypothesis plausible.
  3. Trace forward to papers that turn the original phenomenon into benchmarks, supervised or reward-based training, mechanism probes, inference-time scaling, and closed-loop systems.
  4. Audit the current VBVR arXiv record and v3 HTML/PDF, its project page, benchmark page, and EvalKit as a major scale-up node rather than the field origin.
  5. Query VBVR forward citations from Semantic Scholar and OpenAlex, then search arXiv HTML/PDF full text for 2602.20159, the exact title, and VBVR.
  6. Open each candidate paper and distinguish a reference-list hit from actual use of VBVR data, checkpoints, tasks, or scores.
  7. Independently scan the official Awesome Video Reasoning list for important parallel papers that do not cite VBVR.
  8. Use each original paper as the authority for dataset sizes and metrics. When a comparison table in another paper disagrees, retain both values and label the mismatch.

Focal-paper verification checkpoint

  • The reviewed focal version is arXiv:2509.20328v2, revised 2025-09-29.
  • The paper reports 62 qualitative tasks and 7 quantitative tasks. Its 18,384 generated videos decompose into 744 qualitative generations (62 tasks × 12 samples) and 17,640 quantitative generations.
  • Its operational meaning of zero-shot is no task-specific fine-tuning or task head. Because Veo's training data and post-training are closed, this is not evidence that related patterns or demonstrations were absent from training.
  • The tested object is a black-box system: closed-model Veo 2 or Veo 3 accessed through a publicly callable Google Vertex AI API, including its LLM prompt rewriter. Input is an initial image plus a text instruction; output is an 8-second, 16:9, 720p, 24 FPS video.
  • Chain-of-Frames is introduced with cautious analogy language: successive frames can apply changes through space and time, paralleling the visible role of language tokens in Chain-of-Thought. The paper does not causally establish that one output frame equals one internal reasoning step.
  • The authors explicitly judge that the simple Sudoku solution likely came from the LLM rather than the video model. Their standalone Gemini 2.5 Pro controls neither identify the hidden rewriter nor reproduce its actual rewritten prompts, so they reduce but do not eliminate the hidden-component attribution problem.
  • The paper's pass@k protocols are task-specific: maze success is binary best-of-\(k\); edge and segmentation take the best metric over candidates after within-video frame selection; visual analogy uses majority voting for $k\gt1$. They share an added-sampling budget, but do not have one universal probabilistic interpretation.
  • The Veo 2 identifier is internally inconsistent: the main methods text uses veo-2.0-generate-001, while several appendix passages use veo-2.0-generate-preview-001. Veo 2 versus Veo 3 is a product-version comparison, not a controlled scaling law.
  • The focal paper is therefore treated as a broad phenomenon-discovery and agenda-setting paper. MME-CoF converts breadth into a multidimensional benchmark; VR-Bench narrows CoF to programmable maze trajectories and training; VBVR scales to many tasks and rule scorers; Demystifying Video Reasoning challenges the frame-level mechanism with Chain-of-Steps.

VBVR citation-index snapshot

  • Semantic Scholar paper page returned 22 citing papers on 2026-08-29.
  • API lookup used:
  • https://api.semanticscholar.org/graph/v1/paper/ARXIV:2602.20159?fields=paperId,title
  • https://api.semanticscholar.org/graph/v1/paper/d3e893d3d9a722aea0373ed05c282a81f310a52a/citations?fields=title,year,externalIds,contexts,intents,isInfluential&limit=100
  • OpenAlex split the arXiv/DOI identities and showed a stale count of zero on the main record, while full-text search still located citing papers. It was therefore used for discovery, not as the authoritative count.
  • Crossref did not expose a useful arXiv citation graph; DataCite metadata counts were also stale.
  • Combining databases and direct PDF checks produced a lower bound of 28 formal citing works: 27 arXiv papers and one OpenReview workshop paper.

The number is a dated snapshot, not a permanently closed citation set.


2. VBVR version and numerical audit

Primary source: A Very Big Video Reasoning Suite, arXiv:2602.20159v3, revised 2026-08-27 and accepted to ICML 2026.

2.1 Current main-text figures

Item v3 main-text value Interpretation
Curated and publicly released tasks 150 Current main data-generation pipeline
Images 2,015,000 Initial and target states used to form clips
Clips 1,007,500 1,000,000 train + 7,500 test pool
Training tasks 100 10,000 clips per task
Training clips 1,000,000 Full released training pool
Test-pool tasks 150 50 clips per task
Test-pool clips 7,500 Distinct from the benchmark evaluation subset
VBVR-Bench tasks 100 50 ID + 50 OOD
Evaluated cases 500 Five samples per benchmark task

2.2 Internal version mismatches

  • The arXiv abstract metadata and one stale appendix sentence still say 200 tasks. The current method/data main text says 150.
  • The main text says task design began from more than 500 proposals; an appendix passage says an initial pool of more than 300.
  • The full 7,500 test pool is not the same as the 500-case VBVR-Bench evaluation subset.
  • The main scaling curve stops at 500K training clips even though the release contains 1M. It is incorrect to describe the plot as a complete 1M scaling curve.

2.3 Scorer and human-alignment boundary

  • VBVR uses task-specific 0–1 rule scorers and reports Spearman $\rho\gt0.9$ against human judgments.
  • The reported correlation is computed on nine model-level win-ratio points, not on all 4,500 video-level judgments.
  • Scorer completeness differs by task. Some scorers verify action order and legal transitions; others mainly test final location, color retention, or absence of large jumps.
  • The result supports more repeatable, interpretable model ranking than a generic VLM judge. It does not prove every intermediate causal constraint is checked.

2.4 Main reported results

Model/reference VBVR overall
Human 0.974
Wan2.2 base 0.371
Sora 2 0.546
Veo 3.1 0.480
VBVR-Wan2.2 0.685
VBVR-LTX2.3 0.516
  • VBVR-Wan2.2 improves 84.6% relative to its base score.
  • Across the displayed 0K–500K curve, ID rises from 0.412 to 0.760 and OOD from 0.329 to 0.610.
  • A persistent gap of roughly 15% remains; returns visibly flatten after approximately 200K–400K.
  • VBench-I2V overall changes little, 0.8816 to 0.8835. Camera-motion consistency rises from 0.5444 to 0.6592, while dynamic degree falls from 0.5285 to 0.4106. Reasoning score, motion, and general generation quality therefore need separate reporting.

2.5 Safest one-sentence interpretation

VBVR shows that large-scale, rule-verifiable visual task supervision can substantially improve structured visual state transformation and transfer some behavior to held-out task families; it does not show that scale alone yields general causal reasoning, a faithful frame-by-frame internal algorithm, or a closed-loop world model.


3. Backward genealogy: what VBVR consolidated

VBVR is infrastructure and consolidation, not the origin of video reasoning. The modern sequence is best described as parallel waves rather than a single chain.

3.1 Controlled world-model scaling precursor

Work Role before VBVR Main evidence Boundary
How Far Is Video Generation from World Model: A Physical Law Perspective / PhyWorld Controlled ID/compositional/OOD scaling precedent Box2D tasks; data 30K→3M; DiT about 22M→310M; ID and coverage improve, physical OOD does not reliably close Controlled 2D physics, not general logic or planning

3.2 Capability discovery and Chain-of-Frames framing

Work Role Key facts Evidence boundary
Video models are zero-shot learners and reasoners Modern empirical trigger and CoF naming 62 qualitative + 7 quantitative tasks; 18,384 generated videos; Veo 3 5×5 maze pass@10 78% vs Veo 2 14% Closed model and prompt rewriter; qualitative tasks have few samples; best-frame/pass@10 is search upper bound
MME-CoF Compact multidimensional diagnosis 12 dimensions, 59 curated entries, six displayed model variants Gemini 2.5 Pro judge; VBVR Table 1 says 120, conflicting with the native paper’s 59
Thinking with Video Extends video reasoning to multimodal/text-centric tasks VideoThinkBench 4,149 samples; studies ICL, self-consistency and test-time scaling Prompt rewriting can dominate: Wan 2.5 without rewriter fell to zero on GSM8K/MMLU in the reported test; visible process often remained wrong/unreadable

3.3 Benchmark diversification in late 2025

Work What it added Native scale or result Critical caution
TiViBench Four reasoning dimensions, three difficulty levels, VideoTPO 595 image-prompt samples, 24 scenarios Improvement partly comes from external LLM critique and prompt revision
Gen-ViRe Six cognitive dimensions and mixed process/final-state judging 24 subtasks, 72 prompts, over 2,500 generated videos Small prompt set; heavy dependence on Gemini judge
VR-Bench Programmatic maze supervision, SFT, trajectory metrics, sampling scaling Native paper reports 7,920 videos, five maze families VBVR train/test entries sum to 7,874; the mismatch is unresolved
V-ReasonBench Deterministic final-frame evaluation across four dimensions 326 instances, 652 initial/target images, 9,780 generated videos Explicitly reveals “right endpoint, wrong process” cases
RULER-Bench Rule categories and checklist judging 622 instances, 40 tasks, six rule classes; judge agreement 0.8512 “Rule-based” describes tested rules; scoring still relies on GPT-o3
MMGR Unified image/video generative reasoning across abstract, embodied and physical domains 1,853 samples; ARC-AGI below 10%, Sudoku below 7% Output modalities and judges vary across tasks
SVBench Social intention and interaction Main benchmark uses 15 tasks, 135 prompts Seed suite is 30 tasks, but only 15 are in the short-video main evaluation; agent/judge dependence

3.4 From evaluation to intervention and training

Work Intervention Why it matters
TiViBench / VideoTPO Multi-sample critique and prompt refinement, no VGM weight update Separates base capability from system-level test-time compute
VR-Bench / Wan-R1 line Programmatic trajectory SFT First narrow but clean video-reasoning training testbed
NewtonRewards Physics-grounded, verifiable post-training rewards Turns optical-flow/physical proxies into an optimizable signal
VIPER Process-Oriented Correctness Shows final-state scoring can reward outcome hacking

3.5 Historical placement warning

Demystifying Video Reasoning was first submitted on 2026-03-17, after VBVR v1 on 2026-02-23. It appears in the August v3 references because VBVR was revised later. It is a post-VBVR mechanism paper, not a precursor.


4. Forward citations that actually use or inherit VBVR assets

The following 12 papers do more than mention VBVR. They use its tasks, data, checkpoints, scores, or derived benchmark material.

# Work VBVR dependency New contribution
1 Demystifying Video Reasoning Uses VBVR-Wan/LTX models and samples 200 cases mainly from VBVR-Bench Chain-of-Steps, denoising interventions, Training-Free Ensemble; VBVR 0.685→0.716
2 CollabVR Directly evaluates VBVR-Wan2.2 on VBVR-Bench VLM plan → short VGM clip → VLM verify loop; VBVR-Wan 0.671→0.757 at matched compute
3 SenseNova-U1 Adds VBVR-Image preview evaluation and VBVR-derived comparisons Unifies multimodal understanding and generation under NEO-unify
4 Video Models Can Reason with Verifiable Rewards / VideoRLVR Evaluates VBVR-OOD transfer against VBVR-Wan2.2 SDE-GRPO, dense decomposed rewards, Early-Step Focus; about 40% lower training latency
5 VLMs are Good Teachers Uses VBVR domain-adaptation SFT and benchmark reward VLM-derived differentiable reward + test-time LoRA; VBVR/RULER average +16.7 points
6 World Model Self-Distillation Appendix VBVR OOD puzzle evaluation VLM creates tasks/solutions, Demonstrator generates traces, Executor is distilled and RL-refined
7 OpenCoF OpenCoF-17K includes 30 VBVR subtasks and 7,750 VBVR-derived videos Visual/text reasoning tokens and organized intermediate-state supervision
8 Apple-\(\pi\) Evaluates VBVR-Wan2.2 and discusses VBVR reasoning supervision Law-grounded physical intelligence: perception → formulation → deduction
9 Articulated Object Reconstruction from Rest-State Observation Uses Wan2.2 + VBVR LoRA to generate articulation hypotheses Applies a video-reasoning checkpoint as a dynamic prior for a 3D inverse problem
10 ChronoVision Builds Vbvr-VQA from VBVR-Dataset/Bench-Data Latent final-state reconstruction, ROI evidence localization, joint RL reward
11 VGI-Bench Evaluates multiple VBVR-tuned checkpoints and synthetic-transfer claims 27 tasks, 810 instances; process validity, input sensitivity, limited late correction
12 VBVR-Pro Rewrites 150 VBVR generators and adds 150 300 tasks, verifiable reward, multimodal comparison, multi-task RL, judge audit

Direct-use conclusions

  • The first wave uses VBVR mainly as a diagnostic substrate for mechanism and closed-loop inference.
  • The second wave converts deterministic scorers into RL or test-time rewards.
  • Later work restructures data around intermediate states, creates derived QA tasks, or uses VBVR checkpoints as reusable dynamic priors.
  • VBVR-Pro closes the loop from tasks and scoring to optimization, modality comparison, and mechanism probing.

5. Formal citations where VBVR is context rather than dependency

These 16 works formally cite VBVR, but their central experiment does not depend on VBVR assets. They are still important to the broader research route.

# Work Main contribution Actual role of VBVR
1 EndoCoT Iterative latent thought guidance and terminal grounding; four-task average 92.1%, +8.3 points Related-work evidence that video priors can express logical transformations
2 MME-CoF-Pro 303 samples, 16 categories, necessary-step Reasoning Score, text/visual hints Background example of large-scale verifiable evaluation
3 Video Models Reason Early Early plan commitment, ~12-step threshold, ChEaP; long maze 7%→67% Benchmark background; experiments use VR-Bench
4 How Far Are Video Models from True Multimodal Reasoning? CLVG-Bench, 1,000+ metadata, six categories, 47 subcategories Benchmark taxonomy/comparison
5 Evaluating Spatial World Modeling in Video Generators via 3D Camera Trajectory Generation Floor-plan navigation with jointly generated video and 3D camera pose Related-work comparison
6 Do multimodal models imagine electric sheep? Decodes latent visual states from a VLM; 16 visual tokens per step improve 83%→about 89% Broad emergent-generalization evidence
7 WorldReasonBench 436 open-world future-state cases, 22 subcategories, about 6K preference pairs Synthetic-puzzle benchmark background
8 Entity-Centric World Models / IA-JEPA Interaction-aware masking for collision/momentum; CLEVRER causal 3.22%→14.26% Zero-shot puzzle benchmark background
9 PaintBench 20 precise editing classes and pixel-level deterministic scoring Borrows rule-based human-aligned scorer philosophy
10 Physics-IQ Verified Audits physics benchmark; changes 57.6% samples/34.8% prompts; ranking \(\tau=0.46\) Synthetic reasoning benchmark background
11 Video-MME-Logical Controlled temporal-logical operations and intermediate-state diagnosis for MLLMs Generation-oriented reasoning comparison
12 The Seriality Gap in Video Diffusion Models Shows degradation with dependency-chain length; denoising steps do not provide arbitrary serial depth World-simulator/reasoner motivation
13 Hierarchical Denoising for Multi-Step Visual Reasoning Coarse-to-fine latent tree; success 34.22→60.29; 54.2× faster than bidirectional diffusion search Related benchmark/method lineage
14 Visual prompt engineering for video models Edits visual problem presentation; VPCT Veo 3.1 41.3%→59.3% Broad video-foundation/reasoning context
15 Deferred Exposure of Future Trajectories Avoids future-trajectory leakage in autonomous-driving RLVR Very broad benchmark context
16 Visual General Intelligence: A White Paper Visual-centric general-intelligence agenda Survey/background citation

Why this distinction matters

A citation edge is not a method-dependency edge. For example, Reason Early is central to the video-reasoning mechanism story, but it cites VBVR only as a new benchmark and runs its core experiments on VR-Bench. Conversely, Articulated Object Reconstruction is not a central reasoning paper, but it directly uses a VBVR LoRA checkpoint in its pipeline.


6. Important parallel work not recoverable from the VBVR citation graph

Filtering only by “cites VBVR” would miss several important branches:

Work Independent contribution
From Perception to Action / CHAIN Interactive physics-driven 3D reasoning benchmark; submitted one day after VBVR v1 and currently does not cite it
NEWTON Physics-aware agent toolchain with planner, scientific computation, keyframes, VGM and verifier
World Reasoning Arena Evaluates action simulation fidelity, long-horizon forecast, simulative reasoning and planning
V-Bridge Transfers video-generation priors into few-shot image restoration
ImagiNav VLM subgoals + imagined video + inverse dynamics for robot navigation
Wan-R1 GRPO for flow-based VGM with trajectory/embedding-level verifiable rewards
UniVR Global + step rewards across manipulation, spatial puzzles and physical reasoning
Thinking in Video: Can Video Generators Really Reason About the Real World? Measures a Perception–Prediction Gap between explicit causal answers and plausible generated futures
RuleMaze Natural-language rule generation, logical forms and executable validators for unseen-rule generalization

There is also a naming collision: Lumos-Nexus calls its own 208-case, eight-dimension benchmark “VR-Bench.” It is not the 7,920-video maze benchmark in arXiv:2511.15065. Reviews should disambiguate by paper title or arXiv ID.


7. Evidence-led development route

Phase 1 — emergence, September 2025

  • Question: can a pretrained video generator perform a task without task-specific training?
  • Evidence: qualitative behaviors, best-frame scores, pass@\(k\).
  • Unresolved: black-box prompt rewriting, single-run reliability, mechanism.

Phase 2 — benchmark diversification, October–December 2025

  • Question: which spatial, physical, logical, embodied, and social abilities are present?
  • Evidence: MME-CoF, TiViBench, Gen-ViRe, VR-Bench, V-ReasonBench, RULER, MMGR, SVBench.
  • Unresolved: VLM-judge calibration, final-state shortcuts, small prompt sets.

Phase 3 — scalable supervised learning, February 2026

  • Question: do task diversity and data scale create cross-task generalization?
  • Evidence: VBVR’s million-clip training pool, deterministic scorers, ID/OOD curve.
  • Unresolved: saturation, persistent OOD gap, process completeness, dynamics trade-off.

Phase 4 — mechanism and process, March–April 2026

  • Question: does reasoning happen along output frames, denoising steps, network depth, or early plan states?
  • Evidence: MME-CoF-Pro, Demystifying, Reason Early, Seriality Gap, later VGI-Bench.
  • Unresolved: causal generality across architectures and tasks.

Phase 5 — verifiable optimization and closed loops, May–June 2026

  • Question: how can the model be made reliable at inference time?
  • Evidence: VideoRLVR, CollabVR, VLM-as-Teacher, World Model Self-Distillation.
  • Unresolved: external-teacher attribution, reward hacking, matched compute.

Phase 6 — structure, multimodality, and real-world laws, July–August 2026

  • Question: what architecture and substrate support longer chains and real-world transfer?
  • Evidence: OpenCoF, Hierarchical Denoising, Apple-\(\pi\), ChronoVision, VGI-Bench, VBVR-Pro.
  • Unresolved: long serial dependencies, causal faithfulness, sim-to-real, image/video/interleaved compute trade-offs.

8. Claim ledger used by the review

Claim Supported by Safe wording Unsafe extension
Video models show zero-shot reasoning-like behavior Video-Zero-Shot, MME-CoF, TiViBench Non-zero success on specified tasks under stated sampling budget General reasoning or reliable one-shot solver
Visual generation can help some reasoning VisWorld-Eval, MVoT, visual planning work Helps tasks whose state is naturally spatial/visual Video is universally superior to text
Training scale helps VR-Bench, VBVR, PhyWorld Improves ID and some held-out performance More data will close OOD by itself
Intermediate frames can be useful CoF benchmarks, process scores Output frames expose a checkable state trajectory One frame is necessarily one internal thought
Denoising contains plan information Demystifying, Reason Early, VGI-Bench Intermediate denoising states predict and causally affect outcomes in tested settings All video architectures reason the same way
More frames/steps increase compute Thinking in Frames, Seriality Gap Can help when they enable additional effective dependency computation More frames or denoising steps always mean deeper reasoning
External VLM feedback improves systems VideoTPO, CollabVR, VLMs-as-Teachers Improves the combined planner/generator/verifier system Proves the base VGM independently reasons better
Physics-like output indicates some learned prior NewtonRewards, PhyWorld, Apple-\(\pi\) Shows task-specific physical consistency under selected checks Learned general physical laws or calibrated causal model
Closed-loop video can support planning CHAIN, CollabVR, NEWTON, world-model work Short rollouts plus feedback can improve task success Open-loop generation alone is a decision world model

9. Remaining limitations of this audit

  • The field is moving quickly and papers are revised after initial submission; citation direction can appear cyclic across versions.
  • Closed commercial systems hide prompt rewriters, audio/text solvers, safety filters, and rerankers.
  • Many papers use “reasoning” for different objects: final state, visible process, latent state, or a complete agent pipeline.
  • A paper’s reported metric may have multiple aggregation levels; this audit preserves the native wording where a denominator is not fully recoverable.
  • Full independent reproduction of the reported model scores was outside this literature audit. The review verifies claims against primary papers and official repositories, not by rerunning proprietary models.
  • Citation counts should be refreshed before publication or submission; the 28-paper set is a verified lower bound as of the audit date.