Paper deep dive
A Very Big Video Reasoning Suite
Maijunxian Wang, Ruisi Wang, Juyi Lin, Ran Ji, Thaddäus Wiedemer, Qingying Gao, Dezhi Luo, Yaoyao Qian, Lianyu Huang, Zelong Hong, Jiahui Ge, Qianli Ma, Hang He, Yifan Zhou, Lingzi Guo, Lantao Mei, Jiachen Li, Hanwen Xing, Tianqi Zhao, Fengyuan Yu, Weihang Xiao, Yizheng Jiao, Jianheng Hou, Danyang Zhang, Pengcheng Xu, Boyang Zhong, Zehong Zhao, Gaoyun Fang, John Kitaoka, Yile Xu, Hua Xu, Kenton Blacutt, Tin Nguyen, Siyuan Song, Haoran Sun, Shaoyue Wen, Linyang He, Runming Wang, Yanzhi Wang, Mengyue Yang, Ziqiao Ma, Raphaël Millière, Freda Shi, Nuno Vasconcelos, Daniel Khashabi, Alan Yuille, Yilun Du, Ziming Liu, Bo Li, Dahua Lin, Ziwei Liu, Vikash Kumar, Yijiang Li, Lei Yang, Zhongang Cai, Hokin Deng
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 7/20/2026, 2:58:33 PM
Summary
The paper introduces the Very Big Video Reasoning (VBVR) suite, comprising a large-scale dataset (VBVR-Dataset) and an evaluation framework (VBVR-Bench). The dataset contains over one million video clips across 200 curated reasoning tasks based on five cognitive faculties: abstraction, knowledge, spatiality, perception, and transformation. The suite enables systematic scaling studies, revealing emergent generalization in video reasoning models but highlighting persistent gaps between model and human performance.
Entities (16)
Relation Signals (16)
VBVR → contains → VBVR-Dataset
confidence 98% · VBVR suite, centered on an unprecedentedly large-scale and continually growing dataset for video reasoning, VBVR-Dataset
VBVR → contains → VBVR-Bench
confidence 98% · VBVR suite, centered on ... VBVR-Dataset, together with a verifiable, human-aligned evaluation toolkit, VBVR-Bench.
VBVR-Dataset → coverscognitivefaculty → Abstraction
confidence 95% · organize core visual reasoning capabilities into five pillars: abstraction, knowledge, spatiality, perception, and transformation.
VBVR-Dataset → coverscognitivefaculty → Knowledge
confidence 95% · organize core visual reasoning capabilities into five pillars: abstraction, knowledge, spatiality, perception, and transformation.
VBVR-Dataset → coverscognitivefaculty → Spatiality
confidence 95% · organize core visual reasoning capabilities into five pillars: abstraction, knowledge, spatiality, perception, and transformation.
VBVR-Dataset → coverscognitivefaculty → Perception
confidence 95% · organize core visual reasoning capabilities into five pillars: abstraction, knowledge, spatiality, perception, and transformation.
VBVR-Dataset → coverscognitivefaculty → Transformation
confidence 95% · organize core visual reasoning capabilities into five pillars: abstraction, knowledge, spatiality, perception, and transformation.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Rapid progress in video models has largely focused on visual quality, leaving their reasoning capabilities underexplored. Video reasoning grounds intelligence in spatiotemporally consistent visual environments that go beyond what text can naturally capture, enabling intuitive reasoning over spatiotemporal structure such as continuity, interaction, and causality. However, systematically studying video reasoning and its scaling behavior is hindered by the lack of large-scale training data. To address this gap, we introduce the Very Big Video Reasoning (VBVR) Dataset, an unprecedentedly large-scale resource spanning 200 curated reasoning tasks following a principled taxonomy and over one million video clips, approximately three orders of magnitude larger than existing datasets. We further present VBVR-Bench, a verifiable evaluation framework that moves beyond model-based judging by incorporating rule-based, human-aligned scorers, enabling reproducible and interpretable diagnosis of video reasoning capabilities. Leveraging the VBVR suite, we conduct one of the first large-scale scaling studies of video reasoning and observe early signs of emergent generalization to unseen reasoning tasks. Together, VBVR lays a foundation for the next stage of research in generalizable video reasoning. The data, benchmark toolkit, and models are publicly available at this https URL .
Tags
Links
- Source: https://arxiv.org/abs/2602.20159v2
- Canonical: https://arxiv.org/abs/2602.20159v2
Trouble viewing inline? Open PDF directly →
Full Text
139,626 characters extracted from source content.
Expand or collapse full text
VIDEO-REASON.COM A Very Big Video Reasoning Suite Maijunxian Wang * 1 Ruisi Wang * 2 Juyi Lin * 3 Ran Ji * 4 Thadd ̈ aus Wiedemer 5 Qingying Gao 6 Dezhi Luo 7 Yaoyao Qian 3 Lianyu Huang 8 Zelong Hong 9 Jiahui Ge 8 Qianli Ma 10 Hang He 11 Yifan Zhou 10 Lingzi Guo 12 Lantao Mei 12 Jiachen Li 13 Hanwen Xing 8 Tianqi Zhao 14 Fengyuan Yu 2 Weihang Xiao 15 Yizheng Jiao 16 Jianheng Hou 8 Danyang Zhang 17 Pengcheng Xu 18 Boyang Zhong 19 Zehong Zhao 4 Gaoyun Fang 20 John Kitaoka 21 Yile Xu 22 Hua Xu 23 Kenton Blacutt 24 Tin Nguyen 25 Siyuan Song 13 Haoran Sun 6 Shaoyue Wen 20 Linyang He 26 Runming Wang 6 Yanzhi Wang 3 Mengyue Yang 27 Ziqiao Ma 7 Rapha ̈ el Milli ` ere 28 Freda Shi 29 Nuno Vasconcelos 4 Daniel Khashabi 6 Alan Yuille 6 Yilun Du 30 Ziming Liu 12 Bo Li 2 Dahua Lin 31 Ziwei Liu 2 Vikash Kumar 32 Yijiang Li 4 Lei Yang 31 Zhongang Cai 2 † Hokin Deng 32 † 1 University of California, Berkeley 2 Nanyang Technological University 3 Northeastern University 4 University of California, San Diego 5 University of T ̈ ubingen 6 Johns Hopkins University 7 University of Michigan 8 University of Southern California 9 Washington University in St. Louis 10 Shanghai Jiao Tong University 11 East China Normal University 12 Stanford University 13 University of Texas at Austin 14 University of California, Los Angeles 15 Cornell University 16 University of North Carolina at Chapel Hill 17 San Jose State University 18 University of California, Irvine 19 Technical University of Munich 20 Imperial College London 21 University of Wisconsin–Madison 22 University of Edinburgh 23 Hong Kong University of Science and Technology 24 New York University 25 Auburn University 26 Columbia University 27 University of Bristol 28 University of Oxford 29 University of Waterloo 30 Harvard 31 The Chinese University of Hong Kong 32 Carnegie Mellon University * Equal contribution. † Correspondence to: Hokin Deng<hokind@andrew.cmu.edu>, Zhongang Cai<caiz0023@e.ntu.edu.sg>. Figure 1Overview of VBVR. Left: the grid shows representative tasks spanning our cognitive architecture, which are color-coded according to their underlying capability: Spatiality, Transformation, Knowledge, Abstraction, and Perception. At the center of the grids, we visualize the scale comparison between VBVR (2.015M samples) and nine other datasets combined (12.8K samples): the sizes of the circles are drawn to scale. Top-right: scaling behavior on in-domain and out-of-domain evaluations. Bottom-right: benchmark performance across five cognitive capabilities. VBVR-Wan EvalKit Score VBVR-Wan Training Data In Domain Out of Domain Human 0.974 0.760 0.610 Very Big Video Reasoning Nine Existing Datasets Combined 12.8 K (0.63% of VBVR) The sizes of the circles are drawn to scale* samples Abstraction Knowledge TransformationSpatiality Perception 1 arXiv:2602.20159v2 [cs.CV] 24 Feb 2026 VIDEO-REASON.COMA Very Big Video Reasoning Suite ABSTRACT Rapid progress in video models has largely focused on visual quality, leaving their reasoning capa- bilities underexplored. Video reasoning grounds intelligence in spatiotemporally consistent visual environments that go beyond what text can naturally capture, enabling intuitive reasoning over spatiotemporal structure, such as continuity, interaction, and causality. However, systematically studying video reasoning and its scaling behavior is hindered by the lack of large-scale video reasoning training data. To address this gap, we introduce the Very Big Video Reasoning (VBVR) Dataset, an unprecedentedly large-scale resource spanning 200 curated reasoning tasks following a principled taxonomy, and over one million video clips—approximately three orders of magnitude larger than existing datasets. We further present VBVR-Bench, a verifiable evaluation framework that moves beyond model-based judging by incorporating rule-based, human-aligned scorers, en- abling reproducible and interpretable diagnosis of video reasoning capabilities. Leveraging the VBVR suite, we conduct one of the first large-scale scaling studies of video reasoning and observe early signs of emergent generalization to unseen reasoning tasks. Together, VBVR lays a foundation for the next stage of research in generalizable video reasoning. The data, benchmark tool kit, and models are released publicly at video-reason.com. 1. Introduction Ground-breaking progress has been achieved in large language models, whose reasoning abilities now generalize across challenging tasks such as coding, mathematics, and scientific discovery (Mitchell, 2025; Rapaport, 2026). However, such capabilities remain largely confined to text-based scenarios. Meanwhile, recent advances in video generation models have predominantly emphasized visual realism, with comparatively limited focus on reasoning capabilities. Yet video models hold the potential to support a new paradigm of reasoning (Wiedemer et al., 2025), grounded in spatiotemporally consistent visual environments where spatial structure, physical dynamics, and long-range causality are naturally encoded. This makes video frames an ideal substrate for studying reasoning grounded in the physical world. Despite this promise and growing interest in video reasoning, the community still lacks several critical components required for systematic progress: (1) a large-scale and diverse dataset to enable meaningful investigation of scaling and generalization, (2) an evaluation toolkit built on verifiable and reproducible principles, and (3) an initial scaling study that examines emergent capabilities in video reasoning models. In this work, we address all three challenges by introducing the Very Big Video Reasoning (VBVR) suite. First, we introduce VBVR-Dataset, a large-scale and diverse training source designed to facilitate systematic study of video reasoning. We adopt a principled approach, grounding our task taxonomy in well-established theories of human cognitive architecture (Newell & Simon, 1972; Anderson, 2007). Specifically, we orga- nize core visual reasoning capabilities into five pillars: abstraction, knowledge, spatiality, perception, and transformation. The dataset is the result of a community-oriented, collaborative effort involving over 50 researchers and engineers from diverse disciplines worldwide, ensuring broad coverage and strong domain expertise across 200 tasks to date. Contributors are given full freedom to design the core task semantics and reasoning procedures, allowing for maximal diversity, while a unified, overarching task template is applied as a standardized wrapper for input and output specification. This separation ensures consistency for automated scaling without constraining task creativity. All tasks undergo expert human inspection to ensure quality and correctness before being processed by our automated, cloud-based pipeline, which generates large volumes of randomized training examples in a distributed manner. In total, VBVR-Dataset contains 2,015,000 images and 1,007,500 video clips, making it approximately 1,000×larger than existing alternatives. Importantly, the pipeline is immediately compatible with newly added tasks and supports scalable generation of additional examples per task, enabling continuous expansion in both dataset breadth and scale. Second, VBVR-Bench provides a systematic, reproducible, and explainable evaluation framework for video reasoning models. While VLM-as-a-judge paradigms have been widely adopted for evaluating video generation models (Peng et al., 2025), we explicitly enforce the use of verifiable, rule-based scorers to ensure that 2 VIDEO-REASON.COMA Very Big Video Reasoning Suite Table 1Comparison of VBVR-Dataset with existing video reasoning benchmarks. VBVR-Dataset surpasses all prior benchmarks by multiple orders of magnitude across every dimension and is, to our knowledge, the first to provide large-scale training data for video reasoning. Dataset#Task#Images#Videos#Train #Test Video-Zero-Shot (Wiedemer et al., 2025)691,57800 2,128 V-ReasonBench (Luo et al., 2025c)1365200326 MMGR (Cai et al., 2025)101,3235300 1,853 VideoThinkBench (Tong et al., 2025)248,29800 4,149 TiViBench (Chen et al., 2025)2459500595 VR-Bench (Yang et al., 2025a)507,9206,336 1,538 MME-CoF (Guo et al., 2025)1212000120 Gen-ViRe (Liu et al., 2025)241170072 Ruler-Bench (He et al., 2025)4010100622 VBVR-Dataset200 2,015,000 1,007,500 1,000,000 7,500 Table 2Foundational faculties of a human mind. One-sentence definitions are provided. The comprehensive set of philosophical justifications and empirical supports for each faculty are available in Sec. A. FacultyDefinition AbstractionTo find rules from observations and use rules to deduce results. KnowledgePropositional truth statements one could utter, either learned or gifted since birth. PerceptionImmediate access to sense datum, no further justification could be provided, i.e. ”Here is one hand” SpatialityThe intuition of the basic properties of our world, such as three-dimensionality and Euclidean-ness. TransformationTo simulate spatial-temporal continuities with internal models in one’s mind evaluation outcomes are clearly defined and fully reproducible. To validate that these task-specific scorers faithfully reflect model capabilities, we conduct human preference alignment experiments, observing strong agreement between automated scores and human judgments, with a Spearman’s correlation coefficient of ρ > 0.9. Leveraging VBVR-Bench, we benchmark leading proprietary models: Veo 3.1 (Google DeepMind, 2026), Sora 2 (OpenAI, 2025), Kling 2.6 (Kuaishou Technology, 2025), and Runway Gen-4 (Runway Research, 2025), alongside representative open-source models, including Wan-2.2 (WanTeam, 2025), CogVideoX- 1.5 (Yang et al., 2024), HuanyuanVideo (Kong et al., 2024), and LTX-2 (HaCohen et al., 2026). We reveal a substantial gap in video reasoning capabilities across systems; the strongest model still falls short of human performance by a large margin. Moreover, we use VBVR to analyze how different cognitive capabilities co-develop across models, revealing non-trivial dependencies and trade-offs between reasoning skills. Third, with a large-scale dataset and a reliable evaluation benchmark in place, we conduct an in-depth investigation of scaling effects in video generation models. Using Wan-2.2 as the base model, we observe concurrent performance improvements on both in-domain (ID) and out-of-domain (OOD) tasks as training scale increases, indicating the gradual emergence of generalization capabilities. Beyond these gains, our analysis yields several key insights. First, performance on both ID and OOD tasks eventually plateaus as data scale increases, leaving a persistent gap between model and human performance that cannot be bridged by data scaling alone. This suggests fundamental limitations in current video generation architectures when applied to video reasoning. Second, although OOD performance improves substantially with scale, a consistent gap remains between ID and OOD settings; narrowing this gap appears essential for robust, in-the-wild video reasoning and generation. Finally, qualitative analyses reveal emergent behaviors in instruction following, controlled editing, and semantic understanding with increased model scale, while also exposing important limitations that motivate future research. In summary, we present the VBVR suite, centered on an unprecedentedly large-scale and continually growing dataset for video reasoning, VBVR-Dataset, together with a verifiable, human-aligned evaluation toolkit, VBVR-Bench. Leveraging this suite, we conduct one of the first systematic scaling studies of video reasoning models and uncover early, encouraging evidence of emergent generalization. We believe VBVR provides a foundational infrastructure for future research toward generalizable video reasoning. 3 VIDEO-REASON.COMA Very Big Video Reasoning Suite 2. Related Works Since the inauguration of diffusion models and transformer-based scaling (Ho et al., 2020; Peebles & Xie, 2023), video generation models are rapidly proliferating, with closed models such as Sora, MovieGen, and Veo, and open-source ones like CogVideoX, HunyuanVideo, and Wan (OpenAI, 2025; Polyak et al., 2024; Google DeepMind, 2026; Yang et al., 2024; Kong et al., 2024; WanTeam, 2025). But many models are optimized for creative production rather than explicit relational, causal, or counterfactual reasoning (Peebles & Xie, 2023; Yang et al., 2024; Zheng et al., 2024). Recent research increasingly investigates video generation not only as a content-creation tool, but as a reasoning substrate (Tong et al., 2025; Guo et al., 2025; Liu et al., 2025; Wiedemer et al., 2025). A recent study has tested Veo-3 and shown early evidence that the model exhibits nontrivial zero-shot perceptual and manipulation behaviors and can solve simple tasks without task-specific training (Wiedemer et al., 2025). Later works now span generation-as-reasoning paradigms (Tong et al., 2025), multi-step Chain-of-Frame diagnosis (Guo et al., 2025; Liu et al., 2025), TI2V answer suites (Luo et al., 2025c; Chen et al., 2025), among others (He et al., 2025; Yang et al., 2025a; Cai et al., 2025). Despite sharper measurement, much of the ecosystem remains evaluation-heavy: standardized, large-scale training splits and controlled ablation protocols are missing, making it difficult to run reproducible scaling studies that directly optimize for reasoning correctness. This motivates datasets that are designed not only to test video reasoning, but to support training for reasoning at scale under consistent domain coverage and reliable supervision signals. 3. Dataset In this section, we describe the cognitive architecture underlying our systematic task design (Sec. 3.1), present the key statistics of VBVR-Dataset (Sec. 3.2), and detail the data generation pipeline (Sec. 3.3). 3.1. Cognitive Architecture Aristotle treated cognition as an organized hierarchy of dunameis, cognitive faculties, ascending from aisth ˆ esis, perception, through phantasia, imagination, and mn ̄ em ̄ e, memory, to no ˆ us, understanding, culminating in the extraction of katholou, knowledge, from empeiria, experience (Aristotle, 1984g;b). Kant further argued the mind structures experience through a priori intuitions and categories, aggregating by Einbildungskraft (Kant, 1781). Synthesizing from two millennia of philosophical inquiry and recent cognitive and neural sciences, we organize VBVR around five foundational cognitive faculties. Perception refers to the extraction of structured representations from sensory input, what Aristotle called receiving “form without matter”, where, for example, we test edge detection, color, and shape perception, discrimination (Aristotle, 1984g; Hubel & Wiesel, 1962; DiCarlo et al., 2012). Transformation is the manipulation and synthesis of mental representations, corresponding to Aristotle’s phantasia and Kant’s Einbildungskraft, where we use cases like mental rotation to test (Shepard & Metzler, 1971; Zacks, 2008). Spatiality is the representation of places and their geometric relationships in our world. Kant identified space as an a priori form of intuition prerequisite to perception itself, and we use cases like navigation to probe this ability (Kant, 1781; O’Keefe & Dostrovsky, 1971; Hafting et al., 2005). Abstraction is the distillation of generalizable knowledge from particular experiences, Aristotle’s katholou extracted by no ˆ us, Kant’s transcendental ideas generated by Vernunft, and we use cases like Raven’s Matrices to test (Aristotle, 1984b; Kant, 1781; Carey, 2009; Badre & Nee, 2018). Knowledge, as Aristotle has referred, is the telos of human life. Through the faculties of our mind, humans accumulate and refine knowledge over time. This knowledge may be intrinsic, that is, foundational or core knowledge we are born with, or learned (Aristotle, 1984c;b;d;e; Spelke, 2000; Li et al., 2025). To operationalize these faculties in a video-based reasoning setting, VBVR implements each category as a family of parameterized task generators. Representative task instances for each faculty are illustrated in Fig. 2. Full philosophical grounding and neuroscientific evidence for each faculty appear in Sec. A. 4 VIDEO-REASON.COMA Very Big Video Reasoning Suite Figure 2 Sample task instances generated from the VBVR parameterized task suite, organized by five cognitive faculties. Each sequence illustrates the structured reasoning process required to reach a valid solution. Tasks are implemented as deterministic generators supporting scalable instance variation while preserving visual clarity and video dependency. Each row corresponds to a faculty defined in Section 3.1: abstract cognitive constructs are instantiated as executable, verifiable video-based reasoning tasks. 3.2. Data Statistics Table 1 compares VBVR-Dataset with existing video reasoning benchmarks. VBVR-Dataset surpasses prior work by multiple orders of magnitude across all key dimensions. Notably, most existing video reasoning benchmarks provide few or no video samples (often lacking training data altogether), which has been a major bottleneck for studying scaling. In total, VBVR-Dataset comprises 200 tasks: 150 tasks will be publicly released, while the remaining 50 tasks are reserved as a hidden set for future leaderboard evaluation to preserve benchmark integrity. 5 VIDEO-REASON.COMA Very Big Video Reasoning Suite Figure 3Task designs grounded in cognitive architecture are implemented as parameterized generators, then executed at scale via distributed Lambda workers writing to centralized S3 storage. Foundations Cognitive Architecture Abstraction Perception Spatiality Transformation Knowledge Generator Template Tasks Task Design Task Review Generator Generator Implementation Generator Review 200+ Generators Cloud-based Message FIFO Queue .... λ Generator Worker 1 Generator Worker 2 Generator Worker 3 Generator Worker 4 Generator Worker n S3 Storage 3.3. Data Curation The curation process follows a three-stage pipeline: (1) task design and approval, (2) task-specific generator implementation, and (3) large-scale distributed generation with quality control. Each stage produces well- defined outputs for downstream processing. 3.3.1. Task Design and Approval Each task is designed to probe a specific video-based reasoning capability from the taxonomy defined in Sec. 3.1. Rather than relying on implicit or post-hoc task definitions, VBVR explicitly constrains the task space through a unified set of quality standards. All task proposals are evaluated against six criteria: (1) Information sufficiency, requiring all necessary reasoning cues to be present in the first frame and the prompt; (2) Deterministic solvability, ensuring a unique and verifiable success criterion; (3) Video dependency, such that the task cannot be solved from a single static image but through a process; (4) Visual clarity, ensuring all visual elements are distinguishable with unambiguous layouts; (5) Parametric diversity, supporting the generation of at least 10,000 non- trivial instances; (6) Technical feasibility, avoiding unsolvable or pathological configurations under standard rendering pipelines. Task proposals are submitted by internal contributors and the open-source community. Each proposal specifies the reasoning objective, expected input–output structure, and parameter space. Proposals undergo a design review process conducted by designated reviewers, who assess cognitive validity, verifiability, and scalability. Through this rigorous review process, slightly more than 200 task designs have already been approved from over 500 initial proposals. Only approved tasks proceed to implementation. 3.3.2. Task-Specific Generator Implementation The second stage implements approved designs as executable generators. Each approved task design is implemented as a task-specific parameterized generator. For example, a grid navigation generator produces diverse task instances by specifying different grid sizes, obstacle placements, and start/end positions. For each configuration, it algorithmically computes the solution and generates both the task and ground-truth outputs. To ensure consistency across tasks, VBVR offers a standardized generator template that defines interfaces, output formats, and validation hooks. Each generator deterministically produces a four-component output: (1)firstframe.png(initial state), 6 VIDEO-REASON.COMA Very Big Video Reasoning Suite Table 3Benchmarking results on VBVR-Bench. Overall In-Domain (ID) and Out-of-Domain (OOD) scores are reported alongside category-wise performance. Higher is better. Bold: best in group;underline: second best. In-Domain by CategoryOut-of-Domain by Category ModelsOverallAvg.Abst. Know. Perc. Spat. Trans.Avg.Abst. Know. Perc. Spat. Trans. Human0.9740.9600.9190.9561.000.951.000.9881.001.000.9901.000.970 Open-source Models CogVideoX1.5-5B-I2V (Yang et al., 2024)0.2730.2830.2410.3280.257 0.3280.3050.2620.2810.2350.250 0.2540.282 HunyuanVideo-I2V (Kong et al., 2024)0.2730.2800.2070.3570.293 0.2800.3160.2650.1750.3690.290 0.2530.250 Wan2.2-I2V-A14B (WanTeam, 2025)0.3710.4120.4300.3820.415 0.4040.4190.3290.4050.3080.343 0.2360.307 LTX-2 (HaCohen et al., 2026)0.3130.3290.3160.3620.3260.3400.3060.2970.2440.3370.3170.2310.311 Proprietary Models Runway Gen-4 Turbo (Runway Research, 2025)0.4030.3920.3960.4090.429 0.3410.3630.4140.5150.4290.419 0.3270.373 Sora 2 (OpenAI, 2025)0.5460.5690.6020.4770.581 0.5720.5970.5230.5460.4720.525 0.4620.546 Kling 2.6 (Kuaishou Technology, 2025)0.3690.4080.4650.3230.375 0.3470.5190.3300.5280.1350.272 0.3560.359 Veo 3.1 (Google DeepMind, 2026)0.4800.5310.6110.5030.5200.4440.5100.4290.5770.2770.4200.4410.404 Data Scaling Strong Baseline VBVR-Wan2.20.6850.7600.7240.7500.782 0.7450.8330.6100.7680.5720.547 0.6180.615 (2)prompt.txt(task instruction), (3)finalframe.png(target state), and (4)groundtruth.mp4 (complete solution trajectory). Components (1) and (2) constitute model inputs; components (3) and (4) provide verifiable supervision—complete reasoning paths that enable learning “how” to reason, not just “what” the answer is. Task diversity is achieved through structured parameter spaces defined per generator. Parameters vary across task-relevant dimensions, including object count, spatial configuration, structural complexity, and difficulty level. Generators employ stratified sampling to ensure balanced coverage within each task’s parameter space. Before deployment, all generators undergo code review to verify scalability, visual quality, edge-case handling, and reproducibility under fixed random seeds. Only generators that satisfy these requirements are admitted to large-scale production. 3.3.3. Large-Scale Generation and Control The final stage executes validated generators at scale within a distributed generation framework. VBVR generates one million training samples across 100 training tasks (10,000 per task) and 7500 test samples across 150 test tasks (50 per task). Training and test splits are constructed using disjoint random seed ranges to prevent data leakage. Quality control is fully automated during generation. Each sample is validated for the existence of a solution, visual compliance, and boundary constraints. Failed generations trigger automatic retries; persistent failures are logged for further generator refinement. System-level monitoring tracks generation statistics and validation failure rates across tasks. From the 150 approved task designs, we organize 100 tasks for training and 100 tasks for testing, with a carefully designed dual-split strategy to assess both in-distribution robustness and out-of-distribution generalization in Sec. 4.1. The parameterized infrastructure supports continuous expansion: standardized generator templates enable community contributors to develop new tasks while automated validation ensures consistent quality. This positions VBVR as a living benchmark that evolves with the field’s understanding of video reasoning. Implementation details and generator specifications are provided in Sec. B. 4. Benchmark In this section, we introduce the evaluation toolkit (Sec. 4.1) and assess its validity through alignment with human preferences (Sec. 4.2). We subsequently report the performance of leading video generation models (Sec. 4.3) and investigate the correlations among their reasoning capabilities (Sec. 4.4). 7 VIDEO-REASON.COMA Very Big Video Reasoning Suite 4.1. Evaluation Kit To systematically assess model reasoning capabilities, VBVR-Bench employs a dual-split evaluation strategy across 100 diverse tasks. The first split contains 50 tasks that overlap with the training categories but differ in unseen parameter configurations and sample instances, providing a test of in-domain generalization. The second split includes the remaining 50 tasks, which are entirely novel and are designed to measure out-of- domain generalization. It tests whether models can solve reasoning challenges without prior exposure to similar structures, and thus whether they acquire transferable reasoning primitives rather than relying on task-specific memorization. Each task consists of five test samples, enabling statistically robust evaluation across diverse reasoning scenarios. A key feature of VBVR-Bench is its fully rule-based evaluation framework, which is feasible because most test tasks have a unique, verifiable correct answer, allowing interpretable evaluation based on spatial position, color, object identity, path, or logical outcome. Moreover, geometric, physical, or deductive constraints are also considered in the scoring rubrics. Each of the 100 test tasks is paired with a dedicated evaluation rule, with scores on multiple aspects to compute a weighted, comprehensive score. Sub-criteria include spatial accuracy, trajectory correctness, temporal consistency, and logical validity. For example, in the Task G-45: Key Door Matching (More examples are included in Sec. D), a green dot agent must first locate a color-specified key and then navigate to the matching door within a grid maze. Performance is scored across four weighted dimensions: target identification accuracy (30%), path validity (30%), path efficiency (20%), and animation quality (20%). Target identification verifies that the agent selects the correct key and door without confusing colors, path validity ensures the agent follows allowed paths without wall collisions, path efficiency compares the actual trajectory to the optimal BFS path, and animation quality checks smooth frame-by-frame movement and precise object alignment. A full score indicates perfection in all four dimensions (correct key and door selection, near-optimal pathing, and precise spatial and temporal alignment). Overall, VBVR-Bench provides: •Reproducibility and Determinism. The evaluation is fully deterministic and avoids the stochastic variability or hallucinations associated with LLM-based judgments. •Granular Verifiability. Each task is decomposed into interpretable vectors, allowing precise measure- ment of spatial, temporal, and logical correctness, even at the pixel or object-property level. • Transparent Diagnosis. By explicitly encoding reasoning constraints, the benchmark ranks models and reveals systematic trade-offs, capability gaps, trade-offs, and cross-domain performance trends. 4.2. Human Preference Alignment Analysis To assess alignment between VBVR-Bench and human perception, we conduct a large-scale human preference study and compare model win ratios derived from human judgments with those computed from VBVR-Bench’s automatic metrics. Specifically, human win ratios are obtained from pairwise preference annotations, where a model is considered to win if it is preferred over another model for the same prompt. In contrast, VBVR-Bench win ratios are computed by ranking models by their per-sample automatic scores and counting how often each model outperforms others. As shown in Fig. 4, the two sets of win ratios exhibit strong positive correlations across models, indicating that VBVR-Bench provides reliable and human-aligned performance estimates. Details of the study are provided in the Sec. C.2. 4.3. Leading Model Performances Table 3 shows performance across model families. Most open-source baselines cluster between 0.27 and 0.31 overall, indicating limited capability in complex video reasoning, while Wan2.2-I2V-A14B is the strongest open-source baseline at 0.371. Proprietary models perform better overall, led by Sora 2 (0.546) and Veo 3.1 (0.480), particularly in Abstraction and Transformation categories. Fine-tuning Wan2.2-I2V-A14B on VBVR-Dataset yields VBVR-Wan2.2, which achieves a new state of the art with an overall score of 0.685, representing an 84.6% relative improvement over its base model. VBVR-Wan2.2 8 VIDEO-REASON.COMA Very Big Video Reasoning Suite Figure 4Human alignment analysis for VBVR-Bench. Our experiments show that VBVR-Bench evaluations in all splits closely match human perceptions. In each plot, a dot represents the human preference win ratio (horizontal axis) and VBVR-Bench evaluation win ratio (vertical axis) for a particular video generation model. We linearly fit a straight line to visualize the correlation, and calculate the Spearman’s correlation coefficient (ρ) for each dimension. attains the best performance across all evaluated categories, with especially strong results in Spatiality and Perception, suggesting that large-scale reasoning-oriented data substantially enhances integrated world-model reasoning capabilities. Notably, despite these gains, a considerable gap to human performance remains. This highlights the persistent challenges of long-horizon temporal reasoning and robust symbolic manipulation in video generation. We further analyze how performance evolves with increasing training data under a fixed architecture, and how in-domain and out-of-domain generalization behaviors differ, in Sec. 5. We also analyze the stability and consistency of model behavior using domain-wise score distributions that reveal performance variability and rating noise across domains (see Sec. C.3.2). 4.4. Capability Correlation Figure 5 Residualized capability correlation among five faculties across 9 models (Pearsonρ). We regress out a model-level general factor (overall strength) to highlight structural dependencies and inter-relations. We further study capability dependency among the five cognitive faculties, asking whether strengths in one capability tend to co-occur with strengths in another across models. A naive correlation is often dominated by 9 VIDEO-REASON.COMA Very Big Video Reasoning Suite overall model strength (i.e., stronger models score higher on all categories). To isolate structural dependencies, we compute category-level mean scores per model and regress out a model-level general factor (overall mean score) before measuring Pearson correlations on the residuals. See Sec. C.3.1 for more implementation details. Figure 5 reveals non-trivial structures underlying the cognitive faculties. We observe a strong positive coupling between Knowledge and Spatiality (ρ = 0.461). This result is particularly interesting given the past neuroscience studies which suggest human brains use hippocampal place cells and grid cells to support concept learning. One seminal study suggests patients with bilateral hippocampal damage were impaired at learning both spatial and non-spatial configural associations in a deterministic feedback task, showing only partial, inflexible residual learning, supporting a domain-general hippocampal role in binding and configural learning (Kumaran et al., 2007). The follow-up study shows a hippocampus–ventromedial prefrontal cortex circuit supports the emergence of conceptual knowledge that guides choices, and the hippocampus is uniquely required to transfer that knowledge to a perceptually novel setting (Kumaran et al., 2009). Edward Tolman in last century hypothesizes that our spatial map of the physical world could be transferred to be used as a cognitive map, which is a mental representation, not only used for spatial navigation, but conceptual space construction, and general knowledge learning (Tolman, 1948; Yang et al., 2025b). Converging evidence from neuroscience and AI are coming into places to suggest the deep inter-relatedness between cognitive spatial intrinsics and knowledge acquisition (Baram et al., 2024; Whittington et al., 2025; Xiao et al., 2025). In contrast, Knowledge correlates strongly negatively with Perception (ρ =−0.757). One very interesting debate in cognitive science is whether we should count core knowledge, which are knowledge that we are born with, as actually perception (Bai et al., 2025). Namely, instead of understanding intuitive physics or object permanence as a kind of knowledge, we should actually consider them as a kind of perception, which are actually supported by our perceptual neural circuits, such as medial temporal lobe, rather than our learning and memory circuits, such as hippocampus (Hassabis & Maguire, 2009; Martin & Barense, 2023) Abstraction shows a strong negative correlation with Transformation (ρ =−0.641) and a moderate one with Spatiality (ρ =−0.481), and do not show any signs of positive correlations with any other cognitive faculties. This result is consistent with our understanding with the modularity of abstraction faculty in our brain, namely prefrontal cortex (Vaidya & Badre, 2022; Passingham & Lau, 2023; Bein & Niv, 2025; Li et al., 2026) Moreover, Perception trades off with Spatiality (ρ = −0.565) but is nearly uncorrelated with both Ab- straction (ρ = −0.043) and Transformation (ρ = 0.057). Transformation is also nearly uncorrelated with Spatiality (ρ =−0.050). Overall, VBVR-Bench not only ranks models but also enables interpretable diagnosis of how capabilities co-develop or decouple across systems. 5. VBVR-Wan2.2 Analysis We further investigate VBVR-Wan2.2 to gain insights into scaling video reasoning. We first describe the experimental settings (Sec. 5.1), followed by an analysis of the scaling behavior (Sec. 5.2), comprehensive qualitative evaluations (Sec. 5.3), and performance on general video benchmarks (Sec. 5.4). 5.1. Experiment settings We conduct all experiments on Wan2.2-I2V-A14B without architectural modifications, as the goal of VBVR- Wan2.2 is to investigate data scaling behavior and provide a strong baseline model for the video reasoning research community. Leveraging the VBVR-Dataset, which to our knowledge constitutes one of the largest video reasoning datasets to date, enables a systematic investigation of scaling behaviors in video-based reasoning under a fixed model architecture. For training, we adopt a learning rate of 1e-4 and train for one epoch in each experiment. We employ LoRA adaptation on the DiT backbone, and applying LoRA to the modules q, k, v, o, ffn.0, ffn.2 with a lora rank of 32. 10 VIDEO-REASON.COMA Very Big Video Reasoning Suite Table 4 Performance by Data Scale (0K–500K). Bold: best per column; underline: second best. In-Domain by CategoryOut-of-Domain by Category ModelsOverallAvg.Abst.Know.Perc.Spat.Trans.Avg.Abst.Know.Perc.Spat.Trans. 0K0.3710.4120.4300.3820.4150.4040.4190.3290.4050.3080.3430.2360.307 50K 0.5490.5760.5270.5840.5370.6420.6540.5220.5960.5840.5070.4820.490 100K0.6230.7010.6220.6800.7770.7190.7590.5450.6220.5240.5330.5570.517 200K0.6890.7670.7390.7090.7910.7990.8250.6110.7480.6210.5450.6590.599 300K0.6820.7630.7330.7130.7950.7760.8270.6010.7320.5960.5420.6280.600 400K0.6820.7710.7440.7440.7930.7530.8480.5930.7420.5920.5320.6050.588 500K 0.6850.7600.7240.7500.7820.7450.8330.6100.7680.5720.5470.6180.615 5.2. Scaling Curve To systematically investigate data scaling behavior, we progressively increase the training data size from 0K samples (the original Wan2.2 base model) to 500K samples (VBVR-Wan2.2), and report the corresponding performance changes in Tab. 4. First, training VBVR-Wan2.2 shows clear but saturating gains from data scaling. In-domain (ID) perfor- mance improves substantially with increased training data, rising from 0.412 at initialization to about 0.771 at 400K samples, after which gains plateau and slightly fluctuate. The failure to approach perfect accuracy, even within familiar distributions, suggests that current video generation architectures exhibit fundamental represen- tational and optimization bottlenecks. In particular, these tasks require the simultaneous satisfaction of logical constraints and long-term temporal consistency, while the stochastic nature of video generation introduces cumulative rendering noise and temporal drift. Importantly, this saturation regime makes VBVR-Dataset a valuable testbed for researchers to investigate architectural advances, such as explicit state tracking, structured reasoning modules, or self-correction mechanisms, under controlled and scalable evaluation settings. Second, examining generalization to out-of-domain (OOD) tasks highlights critical insights. Both ID and OOD performance improve with more data, ID from 0.412 to 0.760, and OOD from 0.329 to 0.610. This indicates that scaling data enhances transferable reasoning capabilities beyond memorized patterns. Our qualitative analysis in Sec. 5.3 further illustrates how the model benefits from increased training data and generalizes to out-of-domain tasks, providing interpretable insights into improvements in temporal consistency, logical reasoning, and task transferability. However, a persistent 15% generalization gap remains, suggesting that increasing data within fixed task distributions is insufficient for robust systematic generalization. With our data factory, we plan to continuously introduce new task families and richer compositional regimes in future releases, enabling broader coverage of reasoning patterns and better closing the ID–OOD gap. 5.3. Qualitative Analysis We qualitatively compare Wan2.2-I2V-A14B (base model), VBVR-Wan2.2, and the strongest proprietary baseline in our study, Sora 2. A recurring pattern is that, after VBVR training, VBVR-Wan2.2 can match or even surpass Sora 2 on a broad set of tasks that require verifiable manipulations under stable scenes. This motivates our central takeaway: controllability before reasoning. If a model freely rewrites the scene (background/layout/object identity) during generation, intermediate states become unreliable and any “reasoning action” (delete/move/mark) is no longer verifiable. In practice, the base model(Wan2.2-I2V-A14B) often fails in precisely this way: it may not preserve target identity or stable layouts, thereby breaking the prerequisite for manipulation-based reasoning. To make qualitative comparisons direct and reproducible, we select examples as same-task, same-sample comparisons whenever multiple models are shown on the same case. Importantly, the representative cases in Fig. 6 are out-of-domain (OOD) task families held out from training, so improvements reflect transfer to novel task structures rather than memorization. Controllable execution under constraints (VBVR-Wan2.2 vs Sora 2). Panel A highlights that VBVR training primarily improves constraint-following, tool-like execution under stable scenes. OnO-5Task, Sora 2 introduces extra, unnecessary operations: after deleting the target symbol, it further merges/re-layouts the remaining symbols, violating the intended minimal-edit constraint from the task. In contrast, VBVR-Wan2.2 11 VIDEO-REASON.COMA Very Big Video Reasoning Suite Figure 6 Qualitative overview on held-out OOD task families. Panel A presents same-task, same-sample comparisons between VBVR-Wan2.2 and Sora 2 on three controllable-execution tasks:O-5(delete the marked symbol with minimal unintended changes),O-6(apply a 2D geometric rotation under the target cue), and O-30(rearrange a bookshelf by moving an object into the designated slot), with checkmarks/crosses indicating task success/failure. Panel B shows VBVR-Wan2.2-only emergent behaviors onO-49(complete a symmetric pattern with a consistent self-chosen policy) andO-11(“rationalizing”: modifying intermediate elements to fit an internally assumed transformation narrative). Panel C reports honest boundaries of VBVR-Wan2.2 on G-47(long-horizon key–door navigation, with possible agent duplication/flickering) andO-21(blueprint gap filling, where the video can be correct yet procedurally unfaithful). A. B. O-49: Emergent self-chosen completion policy O-11: “Rationalizing”: Making the scene align with expectations. C. G-47: Agent duplication/flickering O-21: Correct answer,wrong method ✗ ✗ VBVR-Wan 2.2 Sora 2 O-6: Rotate the shape about the marked center to match the target outline. ✗ ✔ O-30: Insert each blue book after the green cluster with the closest average height. ✗ ✔ O-5: Remove the red-bordered symbol while preserving the sequence. ⍻ ✔ exhibits an emergent “do exactly what is asked” capability, deleting the marked symbol without additional changes. OnO-6Task, Sora 2 may fail to maintain scene control and to distinguish the target region from the object to be manipulated, leading to a degenerate outcome where the box and object rotate together. In contrast, VBVR-Wan2.2 correctly separates the target cue from the manipulated object and performs a pivot-based rotation that aligns with the task requirement, suggesting emergent geometric manipulation skills beyond the training task families. OnO-30, VBVR-Wan2.2 successfully performs the required constrained relocation (moving a book into the designated slot). Notably, Sora 2 can fail by producing auxiliary markings/lines without executing the actual object manipulation, illustrating that even strong proprietary models may break down when success requires precise, constraint-following control rather than generic scene editing. Emergent strategies and multi-step behavior (VBVR-Wan2.2). Beyond controlled, tool-like execution, we observe emergent strategy-level regularities and multi-step behaviors on OOD task families. OnO-49, VBVR-Wan2.2 often produces a rule-consistent completion of the missing half while exhibiting a distinctive, self-chosen completion policy: across samples, the completion typically appears as a smooth, coherent “fade-in” fill rather than discrete cell-by-cell edits. This consistency suggests that the model is not merely matching static templates, but is transferring controllable execution primitives and organizing them into a stable policy under a 12 VIDEO-REASON.COMA Very Big Video Reasoning Suite new task structure. OnO-11, we sometimes observe behaviors resembling “understand→act→adjust”. In addition to applying the intended two-step transformation to the queried shape (first change color, then move it), VBVR-Wan2.2 may modify intermediate elements (e.g., shifting a misaligned reference shape toward the arrow cue; arrows are manually overlaid for visualization), effectively rationalizing an internally assumed transformation narrative. While such interventions may conflict with the ground-truth reasoning trace and still yield imperfect final answers, they provide a qualitative signal that the model is maintaining scene-level coherence and executing multi-step plans rather than producing one-shot, uncontrolled scene rewrites. Limitations and failure modes (VBVR-Wan2.2). Despite improved scene controllability, several challenging regimes remain. We observe process unfaithfulness in tasks with explicit procedural ground truth. OnO-21 (construction blueprint), the gold procedure scans candidate pieces one-by-one, previews each candidate at the gap, marks incorrect candidates with a cross, and stops when the correct candidate is found and placed. The generated video can mimic a plausible-looking trial-and-error process without faithfully reflecting the true decision mechanism (“correct answer, wrong method”), highlighting the need for stronger process-level supervision and evaluation. Finally, long-horizon control can break down in interactive tasks. OnG-47, compared to the base model that may move doors/keys directly, VBVR-Wan2.2 better distinguishes the agent from scene entities and exhibits the correct high-level subgoal structure (fetch key→reach door). However, it can still suffer from control failures such as agent duplication/flickering when traversing a coherent path, indicating that maintaining identity and stable dynamics over long horizons remains an open problem. In summary, these qualitative insights reinforce a fundamental shift in evaluating video intelligence: con- trollability is the bedrock of verifiable reasoning. Our results demonstrate that VBVR training moves beyond generic video synthesis, instilling a ’controllability-first’ execution logic that generalizes even to novel, out-of-domain task structures. While VBVR-Wan2.2 shows a nascent ability to coordinate multi-step strategies, the remaining gap in process faithfulness and long-term identity stability highlights the next frontier. Achieving true video reasoning will require not just larger scale, but a move toward models that can maintain rigorous causal and physical constraints over extended temporal horizons. 5.4. General Performance on VBench++ Table 5 Comprehensive evaluation results on VBench-I2V. Model Total Score I2V Score Quality Score Video-Text Camera Motion Video-Image Subject Consist. Video-Image Backgr. Consist. Subject Consistency Background Consistency Motion Smoothness Dynamic Degree Aesthetic Quality Imaging Quality Wan2.2-I2V-A14B 0.8816 0.9582 0.80500.54440.97520.99030.94680.96720.98320.52850.61530.7036 VBVR-Wan2.20.8835 0.9678 0.79920.65920.98040.99210.95470.97220.98520.41060.61530.7080 To evaluate the performance of VBVR-Wan2.2 in real-world video generation scenarios, we benchmarked it against the Wan2.2-I2V-A14B base model using the VBench-I2V suite. As shown in the Tab. 5, After LoRA training on VBVR-Dataset, the model maintains a high level of performance across all core metrics, demonstrating that VBVR-Wan2.2 does not undermine the fundamental generative capabilities of the base model. Notably, we observed a significant increase in Video-Text Camera Motion Consistency (rising from 0.5444 to 0.6592), accompanied by a decrease in Dynamic Degree. These quantitative results align closely with our qualitative findings. That is, the model exhibits a more precise understanding of motion dynamics, effectively discerning which regions require temporal change and which should remain preserved. This balance results in videos that are both more stable and better aligned with the provided motion prompts. 6. Conclusion In this work, we present VBVR-Dataset, the first large-scale and diverse training dataset designed for video reasoning, along with VBVR-Bench, a comprehensive evaluation toolkit for verifiable and reproducible assessment. Through systematic scaling studies, we demonstrate that increasing model scale leads to early signs of emergent generalization in video reasoning. 13 VIDEO-REASON.COMA Very Big Video Reasoning Suite References Allison, H. E. Kant’s Transcendental Idealism: An Interpretation and Defense. Yale University Press, New Haven, revised and enlarged edition, 2004. Andersen, R. A., Snyder, L. H., Bradley, D. C., and Xing, J. Multimodal representation of space in the posterior parietal cortex and its use in planning movements. Annual Review of Neuroscience, 20:303–330, 1997. doi: 10.1146/annurev.neuro.20.1.303. Anderson, J. R. Rules of the Mind. Lawrence Erlbaum Associates, Hillsdale, NJ, 1993. Anderson, J. R. How Can the Human Mind Occur in the Physical Universe? Oxford University Press, 2007. Anderson, J. R. Language, memory, and thought. Psychology Press, 2013. Anderson, J. R. and Lebiere, C. The Atomic Components of Thought. Lawrence Erlbaum Associates, Mahwah, NJ, 1998. Aristotle. On memory. In Barnes, J. (ed.), The Complete Works of Aristotle: The Revised Oxford Translation, volume 1, p. 714–720. Princeton University Press, 1984a. Aristotle. Metaphysics. In Barnes, J. (ed.), The Complete Works of Aristotle: The Revised Oxford Translation, volume 2, p. 1552–1728. Princeton University Press, 1984b. Aristotle. Physics, volume 1. Princeton University Press, 1984c. Aristotle. Posterior Analytics, volume 1. Princeton University Press, 1984d. Aristotle. Prior analytics. In Barnes, J. (ed.), The Complete Works of Aristotle: The Revised Oxford Translation, volume 1 of Bollingen Series. Princeton University Press, Princeton, NJ, 1984e. Aristotle. Sense and sensibilia. In Barnes, J. (ed.), The Complete Works of Aristotle: The Revised Oxford Translation, volume 1, p. 693–713. Princeton University Press, 1984f. Aristotle. On the soul. In Barnes, J. (ed.), The Complete Works of Aristotle: The Revised Oxford Translation, volume 1, p. 641–692. Princeton University Press, 1984g. Badre, D. and Nee, D. E. Frontal cortex and the hierarchical control of behavior. Trends in Cognitive Sciences, 22(2):170–188, 2018. Bai, D., Hafri, A., Izard, V., Firestone, C., and Strickland, B. “core perception”: Re-imagining precocious reasoning as sophisticated perceiving. Behavioral and Brain Sciences, p. 1–75, 2025. Baillargeon, R., Spelke, E. S., and Wasserman, S. Object permanence in five-month-old infants. Cognition, 20 (3):191–208, 1985. Baker, N., Lu, H., Erlikhman, G., and Kellman, P. J. Deep convolutional networks do not classify based on global object shape. PLoS Computational Biology, 14(12):e1006613, 2018. doi: 10.1371/journal.pcbi. 1006613. Banino, A., Barry, C., Uria, B., Blundell, C., Lillicrap, T., Mirowski, P., Pritzel, A., Chadwick, M. J., Degris, T., Modayil, J., Wayne, G., Soyer, H., Viola, F., Zhang, B., Goroshin, R., Rabinowitz, N., Pascanu, R., Beattie, C., Petersen, S., Sadik, A., Gaffney, S., King, H., Kavukcuoglu, K., Hassabis, D., Hadsell, R., and Kumaran, D. Vector-based navigation using grid-like representations in artificial agents. Nature, 557(7705): 429–433, 2018. doi: 10.1038/s41586-018-0102-6. Bao, P., She, L., McGill, M., Bhattacharyya, R., and Tsao, D. Y. A map of object space in primate inferotem- poral cortex. Nature, 583:103–108, 2020. doi: 10.1038/s41586-020-2350-5. 14 VIDEO-REASON.COMA Very Big Video Reasoning Suite Baram, A., Nili, H., Barreiros, I., Samborska, V., Behrens, T. E., and Garvert, M. M. An abstract relational map emerges in the human medial prefrontal cortex with consolidation. bioRxiv, p. 2024–10, 2024. Behrens, T. E. J., Muller, T. H., Whittington, J. C. R., Mark, S., Baram, A. B., Stachenfeld, K. L., and Kurth-Nelson, Z. What is a cognitive map? Organizing knowledge for flexible behavior. Neuron, 100(2): 490–509, 2018. doi: 10.1016/j.neuron.2018.10.002. Bein, O. and Niv, Y. Schemas, reinforcement learning and the medial prefrontal cortex. Nature Reviews Neuroscience, 26(3):141–157, 2025. Bengio, Y., Courville, A., and Vincent, P. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798–1828, 2013. Bertolero, M. A., Yeo, B. T. T., and D’Esposito, M. The modular and integrative functional architecture of the human brain. Proceedings of the National Academy of Sciences, 112(49):E6798–E6807, 2015. doi: 10.1073/pnas.1510619112. Bonhoeffer, T. and Grinvald, A. Iso-orientation domains in cat visual cortex are arranged in pinwheel-like patterns. Nature, 353:429–431, 1991. Bowman, C. R. and Zeithamova, D. Abstract memory representations in the ventromedial prefrontal cortex and hippocampus support concept generalization. Journal of Neuroscience, 38(10):2605–2614, 2018. Brincat, S. L., Siegel, M., von Nicolai, C., and Miller, E. K. Gradual progression from sensory to task-related processing in cerebral cortex. Proceedings of the National Academy of Sciences, 115(30):E7202–E7211, 2018. Burgess, P. W., Dumontheil, I., and Gilbert, S. J. The gateway hypothesis of rostral prefrontal cortex (area 10) function. Trends in Cognitive Sciences, 11(7):290–298, 2007. Cai, Z., Qiu, H., Ma, T., Zhao, H., Zhou, G., Huang, K.-H., Kordjamshidi, P., Zhang, M., Xiao, W., Gu, J., Peng, N., and Hu, J. MMGR: Multi-modal generative reasoning. arXiv preprint arXiv:2512.14691, 2025. URL https://arxiv.org/abs/2512.14691. Calvert, G. A. Crossmodal processing in the human brain: Insights from functional neuroimaging studies. Cerebral Cortex, 11(12):1110–1123, 2001. Carey, S. The Origin of Concepts. Oxford Series in Cognitive Development. Oxford University Press, New York, 2009. Carey, S. Pr ́ ecis of The Origin of Concepts. Behavioral and Brain Sciences, 34(3):113–124, 2011. Carey, S. and Spelke, E. Science and core knowledge. Philosophy of Science, 63(4):515–533, 1996. Chakladar, D. D. Cortex level connectivity between act-r modules during eeg-based n-back task. Cognitive Neurodynamics, 18(6):4033–4045, December 2024. doi: 10.1007/s11571-024-10177-y. Chen, H. H., Lan, D., Shu, W.-J., Liu, Q., Wang, Z., Chen, S., Cheng, W., Chen, K., Zhang, H., Zhang, Z., Guo, R., Cheng, Y., and Chen, Y.-C. Tivibench: Benchmarking think-in-video reasoning for video generative models. arXiv preprint arXiv:2511.13704, 2025. URL https://arxiv.org/abs/2511.13704. Christoff, K., Keramatian, K., Gordon, A. M., Smith, R., and M ̈ adler, B. Prefrontal organization of cognitive control according to levels of abstraction. Brain Research, 1286:94–105, 2009. Colby, C. L. and Goldberg, M. E. Space and attention in parietal cortex. Annual Review of Neuroscience, 22: 319–349, 1999. doi: 10.1146/annurev.neuro.22.1.319. d’anthropologie de Paris, S. Bulletins de la Soci ́ et ́ e d’anthropologie de Paris. Masson., 1898. 15 VIDEO-REASON.COMA Very Big Video Reasoning Suite DiCarlo, J. J., Zoccolan, D., and Rust, N. C. How does the brain solve visual object recognition? Neuron, 73 (3):415–434, 2012. Friston, K. The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11:127–138, 2010. doi: 10.1038/nrn2787. Geirhos, R., Narayanappa, K., Mitzkus, B., Thieringer, T., Bethge, M., Wichmann, F. A., and Brendel, W. Partial success in closing the gap between human and machine vision. Advances in Neural Information Processing Systems, 34:23885–23899, 2021. Gentner, D. Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2):155–170, 1983. Gentner, D. and Hoyos, C. Analogy and abstraction. Topics in Cognitive Science, 9(3):672–693, 2017. doi: 10.1111/tops.12278. Ghazanfar, A. A. and Schroeder, C. E. Is neocortex essentially multisensory? Trends in Cognitive Sciences, 10(6):278–285, 2006. doi: 10.1016/j.tics.2006.04.008. Gibson, J. The ecological approach to visual perception: classic edition, 2014. Gilboa, A. and Marlatte, H. Neurobiology of schemas and schema-mediated memory. Trends in Cognitive Sciences, 21(8):618–631, 2017. Goodale, M. A. and Milner, A. D. Separate visual pathways for perception and action. Trends in Neurosciences, 15(1):20–25, 1992. doi: 10.1016/0166-2236(92)90344-8. Google DeepMind. Veo 3.1. Technical report, Google DeepMind, 2026. URLhttps://blog.google/ innovation-and-ai/technology/ai/veo-3-1-ingredients-to-video/ .Released January 13, 2026. Greco, A., Moser, J., Preissl, H., and Siegel, M. Predictive learning shapes the representational geometry of the human brain. Nature communications, 15(1):9670, 2024. Guo, Z., Chen, X., Zhang, R., An, R., Qi, Y., Jiang, D., Li, X., Zhang, M., Li, H., and Heng, P.-A. Are video models ready as zero-shot reasoners? an empirical study with the MME-CoF benchmark. arXiv preprint arXiv:2510.26802, 2025. URL https://arxiv.org/abs/2510.26802. Guyer, P. Kant and the Claims of Knowledge. Cambridge University Press, Cambridge, 1987. HaCohen, Y., Brazowski, B., Chiprut, N., Bitterman, Y., Kvochko, A., Berkowitz, A., Shalem, D., Lifschitz, D., Moshe, D., Porat, E., Richardson, E., Shiran, G., Chachy, I., Chetboun, J., Finkelson, M., Kupchick, M., Zabari, N., Guetta, N., Kotler, N., Bibi, O., Gordon, O., Panet, P., Benita, R., Armon, S., Kulikov, V., Inger, Y., Shiftan, Y., Melumian, Z., and Farbman, Z. Ltx-2: Efficient joint audio-visual foundation model, 2026. URL https://arxiv.org/abs/2601.03233. Submitted 6 Jan 2026. Hafting, T., Fyhn, M., Molden, S., Moser, M.-B., and Moser, E. I. Microstructure of a spatial map in the entorhinal cortex. Nature, 436(7052):801–806, 2005. doi: 10.1038/nature03721. Hassabis, D. and Maguire, E. A. The construction system of the brain. Philosophical Transactions of the Royal Society B: Biological Sciences, 364(1521):1263–1271, 2009. He, X., Fan, Z., Li, H., Zhuo, F., Xu, H., Cheng, S., Weng, D., Liu, H., Ye, C., and Wu, B. RULER-bench: Prob- ing rule-based reasoning abilities of next-level video generation models for vision foundation intelligence. arXiv preprint arXiv:2512.02622, 2025. URL https://arxiv.org/abs/2512.02622. Hegarty, M., Montello, D. R., Richardson, A. E., Ishikawa, T., and Lovelace, K. Spatial abilities at different scales: Individual differences in aptitude-test performance and spatial-layout learning. Intelligence, 34(2): 151–176, 2006. doi: 10.1016/j.intell.2005.09.005. 16 VIDEO-REASON.COMA Very Big Video Reasoning Suite H ́ elie, S. and Sun, R. Creative problem solving: A CLARION theory. In 2010 International Joint Conference on Neural Networks (IJCNN), p. 1460–1466, Barcelona, Spain, 2010. IEEE. doi: 10.1109/IJCNN.2010. 5596891. URL https://doi.org/10.1109/IJCNN.2010.5596891. Helmholtz, H. v. Treatise on physiological optics, 3 vols. Optical Society of America, 1924. Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, p. 6840–6851, 2020. Holyoak, K. J. Analogy and relational reasoning. In Holyoak, K. J. and Morrison, R. G. (eds.), The Oxford Handbook of Thinking and Reasoning, p. 234–259. Oxford University Press, 2012. Holyoak, K. J., Gentner, D., and Kokinov, B. N. Introduction: The place of analogy in cognition. In Gentner, D., Holyoak, K. J., and Kokinov, B. N. (eds.), The Analogical Mind: Perspectives from Cognitive Science, p. 1–19. MIT Press, 2001. Hubel, D. H. and Wiesel, T. N. Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. The Journal of Physiology, 160:106–154, 1962. Hubel, D. H. and Wiesel, T. N. Receptive fields and functional architecture of monkey striate cortex. The Journal of Physiology, 195(1):215–243, 1968. Hubel, D. H. and Wiesel, T. N. Ferrier lecture: Functional architecture of macaque monkey visual cortex. Proceedings of the Royal Society of London. Series B, Biological Sciences, 198(1130):1–59, 1977. Kant, I. Critique of Pure Reason. Cambridge University Press, Cambridge, 1781. The Cambridge Edition of the Works of Immanuel Kant. Kanwisher, N. Functional specificity in the human brain: A window into the functional architecture of the mind. Proceedings of the National Academy of Sciences, 107(25):11163–11170, 2010. doi: 10.1073/pnas. 1005062107. Kanwisher, N., McDermott, J., and Chun, M. M. The fusiform face area: A module in human extrastriate cortex specialized for face perception. Journal of Neuroscience, 17(11):4302–4311, 1997. doi: 10.1523/ JNEUROSCI.17-11-04302.1997. Kazemian, A., Elmoznino, E., and Bonner, M. F. Convolutional architectures are cortex-aligned de novo. Nature Machine Intelligence, 2025. doi: 10.1038/s42256-025-01142-3. Keil, F. C. Concepts, kinds, and cognitive development. mit Press, 1992. Keller, G. B. and Mrsic-Flogel, T. D. Predictive processing: A canonical cortical computation. Neuron, 100 (2):424–435, 2018. doi: 10.1016/j.neuron.2018.10.003. Kersten, D., Mamassian, P., and Yuille, A. Object perception as bayesian inference. Annual Review of Psychology, 55:271–304, 2004. doi: 10.1146/annurev.psych.55.090902.142005. Kong, W., Tian, Q., Zhang, Z., Min, R., Dai, Z., Zhou, J., Xiong, J., Li, X., Wu, B., Zhang, J., et al. Hun- yuanVideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. Kosslyn, S. M. Image and Mind. Harvard University Press, Cambridge, MA, 1980. Kosslyn, S. M. Image and brain: The resolution of the imagery debate. MIT press, 1996. Kriegeskorte, N. Deep neural networks: A new framework for modeling biological vision and brain information processing. Annual Review of Vision Science, 1:417–446, 2015. 17 VIDEO-REASON.COMA Very Big Video Reasoning Suite Kuaishou Technology. Kling AI launches video 2.6 model with “simultaneous audio-visual generation” capability, redefining AI video creation workflow. Press Release, December 2025. Model released December 3, 2025. Press release published December 5, 2025. Kumaran, D., Hassabis, D., Spiers, H. J., Vann, S. D., Vargha-Khadem, F., and Maguire, E. A. Impaired spatial and non-spatial configural learning in patients with hippocampal pathology. Neuropsychologia, 45(12): 2699–2711, 2007. Kumaran, D., Summerfield, J. J., Hassabis, D., and Maguire, E. A. Tracking the emergence of conceptual knowledge during human decision making. Neuron, 63(6):889–901, 2009. K ̈ ording, K. P., Beierholm, U., Ma, W. J., Quartz, S., Tenenbaum, J. B., and Shams, L. Causal inference in multisensory perception. PLoS ONE, 2(9):e943, 2007. doi: 10.1371/journal.pone.0000943. Laird, J. E., Rosenbloom, P. S., and Newell, A. Soar: An architecture for general intelligence. Artificial Intelligence, 33(1):1–64, 1987. doi: 10.1016/0004-3702(87)90050-6. URLhttps://doi.org/10. 1016/0004-3702(87)90050-6. Langley, P., Laird, J. E., and Rogers, S. Cognitive architectures: Research issues and challenges. Cognitive Systems Research, 10(2):141–160, 2009. doi: 10.1016/j.cogsys.2006.07.004. URLhttps://doi.org/ 10.1016/j.cogsys.2006.07.004. LeCun, Y., Bengio, Y., and Hinton, G. Deep learning. Nature, 521(7553):436–444, 2015. Li, H., Chrysanthidis, N., Brincat, S. L., Rose, J., and Miller, E. K. Neural subspace reorganization reflects value-based decision making. bioRxiv, p. 2026–02, 2026. Li, Y., Gao, Q., Zhao, T., Wang, B., Sun, H., Lyu, H., Hawkins, R. D., Vasconcelos, N., Golan, T., Luo, D., and Deng, H. Core knowledge deficits in multi-modal language models. 2025. URLhttps: //arxiv.org/abs/2410.10855. Liu, X., Xu, Z., Li, M., Wang, K., Lee, Y. J., and Shang, Y. Can world simulators reason? Gen-ViRe: A generative visual reasoning benchmark. arXiv preprint arXiv:2511.13853, 2025. URLhttps://arxiv. org/abs/2511.13853. Longuenesse, B. Kant and the Capacity to Judge: Sensibility and Discursivity in the Transcendental Analytic of the Critique of Pure Reason. Princeton University Press, Princeton, 1998. Luo, D., Li, Y., and Deng, H. The philosophical foundations of growing ai like a child. arXiv preprint arXiv:2502.10742, 2025a. Luo, D., Wang, M., Wang, B., Zhao, T., Li, Y., and Deng, H. Machine psychophysics: Cognitive control in vision-language models. arXiv preprint arXiv:2505.18969, 2025b. Luo, Y., Zhao, X., Lin, B., Zhu, L., Tang, L., Liu, Y., Chen, Y.-C., Qian, S., Wang, X., and You, Y. V- reasonbench: Toward unified reasoning benchmark suite for video generation models. arXiv preprint arXiv:2511.16668, 2025c. URL https://arxiv.org/abs/2511.16668. Ma, W. J., Beck, J. M., Latham, P. E., and Pouget, A. Bayesian inference with probabilistic population codes. Nature Neuroscience, 9:1432–1438, 2006. doi: 10.1038/n1790. Marr, D. Vision: A Computational Investigation into the Human Representation and Processing of Visual Information. W.H. Freeman, San Francisco, 1982. Martin, C. B. and Barense, M. D. Perception and memory in the ventral visual stream and medial temporal lobe. Annual review of vision science, 9(1):409–434, 2023. 18 VIDEO-REASON.COMA Very Big Video Reasoning Suite McClelland, J. L., McNaughton, B. L., and Lampinen, A. K. Integration of new information in memory: New insights from a complementary learning systems perspective. Philosophical Transactions of the Royal Society B, 375(1799):20190637, 2020. McCorduck, P. Machines Who Think: A Personal Inquiry into the History and Prospects of Artificial Intelligence. A K Peters/CRC Press, 2 edition, 2004. ISBN 9781568812052. doi: 10.1201/9780429258985. Milner, A. D. and Goodale, M. A. Two visual systems re-viewed. Neuropsychologia, 46(3):774–785, 2008. Mitchell, M. Artificial intelligence learns to reason, 2025. Moser, E. I. and Moser, M.-B. Grid cells and cortical representation. Nature Reviews Neuroscience, 15: 466–481, 2014. doi: 10.1038/nrn3766. Newell, A. You can’t play 20 questions with nature and win: Projective comments on the papers of this symposium. In Chase, W. G. (ed.), Visual Information Processing, p. 283–308. Academic Press, New York, NY, 1973. ISBN 9780121701505. doi: 10.1016/B978-0-12-170150-5.50012-3. URLhttps:// linkinghub.elsevier.com/retrieve/pii/B9780121701505500123. Proceedings of the Eighth Annual Carnegie Symposium on Cognition, Carnegie-Mellon University, Pittsburgh, Pennsylvania, May 19, 1972. Newell, A. and Simon, H. A. The logic theory machine – a complex information processing system. IRE Transactions on Information Theory, 2(3):61–79, 1956. Newell, A. and Simon, H. A. Human Problem Solving. Prentice-Hall, Englewood Cliffs, NJ, 1972. ISBN 0134454030. Newell, A., Shaw, J. C., and Simon, H. A. Report on a general problem solving program. In IFIP congress, volume 256, p. 1959. Pittsburgh, PA, 1959. Nobel Prize Committee. The Nobel prize in physiology or medicine 2014.https://w.nobelprize. org/prizes/medicine/2014/summary/ , 2014. Awarded to John O’Keefe, May-Britt Moser and Edvard I. Moser. O’Keefe, J. and Dostrovsky, J. The hippocampus as a spatial map: Preliminary evidence from unit activity in the freely-moving rat. Brain Research, 34(1):171–175, 1971. doi: 10.1016/0006-8993(71)90358-1. Olshausen, B. A. and Field, D. J. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, 381:607–609, 1996. OpenAI. Sora: Openai’s text-to-video model.https://openai.com/index/sora-is-here, 2025. publicly released September 2025. Palmeri, T. J., Love, B. C., and Turner, B. M. Model-based cognitive neuroscience. Journal of Mathematical Psychology, 76(Pt B):59–64, February 2017. doi: 10.1016/j.jmp.2016.10.010. Papadopoulos, A., Sforazzini, F., Egan, G., and Jamadar, S. Functional subdivisions within the human intraparietal sulcus are involved in visuospatial transformation in a non-context-dependent manner. Human brain mapping, 39(1):354–368, 2018. Passingham, R. E. and Lau, H. Do we understand the prefrontal cortex? Brain Structure and Function, 228(5): 1095–1105, 2023. Peebles, W. and Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), p. 4195–4205, 2023. Peng, W., Wang, G., Yang, T., Li, C., Xu, X., He, H., and Zhang, K. SVBench: Evaluation of video generation models on social reasoning. arXiv preprint arXiv:2512.21507, 2025. 19 VIDEO-REASON.COMA Very Big Video Reasoning Suite Petersen, S. E. and Sporns, O. Brain networks and cognitive architectures. Neuron, 88(1):207–219, Oct 2015. doi: 10.1016/j.neuron.2015.09.027. Plato. Theaetetus, volume 12. Harvard University Press, Cambridge, MA, 1921. Loeb Classical Library. Original work published ca. 369 BCE. Polyak, A., Zohar, A., Brown, A., Tjandra, A., Sinha, A., Lee, A., Vyas, A., Shi, B., Ma, C., Chuang, C., et al. MovieGen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. Quian Quiroga, R., Boscaglia, M., Jonas, J., Rey, H. G., Yan, X., Maillard, L., Colnat-Coulbois, S., Koessler, L., and Rossion, B. Single neuron responses underlying face recognition in the human midfusiform face-selective cortex. Nature Communications, 14(1):5661, 2023. Ramnani, N. and Owen, A. M. Anterior prefrontal cortex: Insights into function from anatomy and neuroimag- ing. Nature Reviews Neuroscience, 5(3):184–194, 2004. Rao, R. P. N. and Ballard, D. H. Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1):79–87, 1999. doi: 10.1038/4580. Rapaport, W. J. Has ai succeeded? 2026. Renero, A. Nous and aisth ̄ esis: Two cognitive faculties in aristotle. M ́ ethexis, 26(1):103–120, 2013. doi: 10.1163/24680974-90000616. Runway Research. Introducing runway gen-4: Next-generation ai models for media generation and world consistency. Runway Research, March 2025. URLhttps://runwayml.com/research/ introducing-runway-gen-4. Accessed January 27, 2026. Russell, S. J. and Norvig, P. Artificial Intelligence: A Modern Approach. Pearson, Hoboken, 4th edition, 2021. ISBN 9780134610993. Shams, L. and Beierholm, U. R. Causal inference in perception. Trends in Cognitive Sciences, 14(9):425–432, 2010. Shepard, R. N. and Metzler, J. Mental rotation of three-dimensional objects. Science, 171(3972):701–703, 1971. doi: 10.1126/science.171.3972.701. Shields, C. Aristotle: De Anima. Oxford University Press, 2016. Spelke, E. S. Core knowledge. American Psychologist, 55(11):1233–1243, 2000. doi: 10.1037/0003-066X.55. 11.1233. Spelke, E. S. and Kinzler, K. D. Core knowledge. Developmental Science, 10(1):89–96, 2007. doi: 10.1111/j.1467-7687.2007.00569.x. Sporns, O., Tononi, G., and K ̈ otter, R. The human connectome: A structural description of the human brain. PLoS Computational Biology, 1(4):e42, 2005. doi: 10.1371/journal.pcbi.0010042. Stein, B. E. and Stanford, T. R. Multisensory integration: Current issues from the perspective of the single neuron. Nature Reviews Neuroscience, 9:255–266, 2008. doi: 10.1038/nrn2331. Swanson, L. R. The predictive processing paradigm has roots in kant. Frontiers in Systems Neuroscience, 10: 79, 2016. doi: 10.3389/fnsys.2016.00079. Tanaka, K. Inferotemporal cortex and object vision. Annual Review of Neuroscience, 19:109–139, 1996. Tolman, E. C. Cognitive maps in rats and men. Psychological Review, 55(4):189–208, 1948. doi: 10.1037/ h0061626. 20 VIDEO-REASON.COMA Very Big Video Reasoning Suite Tomov, M. S., Yagati, S., Kumar, A., Yang, W., and Gershman, S. J. Discovery of hierarchical representations for efficient planning. PLoS Computational Biology, 16(4):e1007594, 2020. Tong, J., Mou, Y., Li, H., Li, M., Yang, Y., Zhang, M., Chen, Q., et al. Thinking with video: Video generation as a promising multimodal reasoning paradigm. arXiv preprint arXiv:2511.04570, 2025. Treisman, A. M. and Gelade, G. A feature-integration theory of attention. Cognitive Psychology, 12(1): 97–136, 1980. Tsao, D. Y., Freiwald, W. A., Tootell, R. B. H., and Livingstone, M. S. A cortical region consisting entirely of face-selective cells. Science, 311(5761):670–674, 2006. doi: 10.1126/science.1119983. Tse, D., Takeuchi, T., Kakeyama, M., Kajii, Y., Okuno, H., Tohyama, C., Bito, H., and Morris, R. G. M. Schema-dependent gene activation and memory encoding in neocortex. Science, 333(6044):891–895, 2011. Vaidya, A. R. and Badre, D. Abstract task representations for inference and control. Trends in cognitive sciences, 26(6):484–498, 2022. von Helmholtz, H. Handbuch der physiologischen Optik. L. Voss, Leipzig, 1867. WanTeam. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. URL https://arxiv.org/abs/2503.20314. Wedin, M. V. Mind and Imagination in Aristotle. Yale University Press, New Haven, 1988. Wernicke, C. Der aphasische Symptomencomplex: Eine psychologische Studie auf anatomischer Basis. Max Cohn und Weigert, Breslau, 1874. English translation: The Symptom Complex of Aphasia. Whitehead, A. N. and Russell, B. Principia mathematica, volume 2. The University Press, 1927. Whittington, J. C., Dorrell, W., Behrens, T. E., Ganguli, S., and El-Gaby, M. A tale of two algorithms: Structured slots explain prefrontal sequence memory and are unified with hippocampal cognitive maps. Neuron, 113(2):321–333, 2025. Whittington, J. C. R., Muller, T. H., Mark, S., Chen, G., Barry, C., Burgess, N., and Behrens, T. E. J. The Tolman-Eichenbaum machine: Unifying space and relational memory through generalization in the hippocampal formation. Cell, 183(5):1249–1263, 2020. doi: 10.1016/j.cell.2020.10.024. Wiedemer, T., Li, Y., Vicol, P., Gu, S. S., Matarese, N., Swersky, K., Kim, B., Jaini, P., and Geirhos, R. Video models are zero-shot learners and reasoners. arXiv preprint arXiv:2509.20328, 2025. URL https://arxiv.org/abs/2509.20328. Xiao, Z., Wang, X., Zhang, J., Ou, J., He, L., Qu, Y., Hu, X., Behrens, T. E., and Liu, Y. Human hippocampal ripples align new experiences with a grid-like schema. Neuron, 113(21):3661–3672, 2025. Yamakawa, H. The whole brain architecture approach: Accelerating the development of artificial gen- eral intelligence by referring to the brain.Neural Networks, 144:478–495, 2021.doi: 10.1016/ j.neunet.2021.09.004.URLhttps://w.sciencedirect.com/science/article/pii/ S0893608021003543. Yamins, D. L. K., Hong, H., Cadieu, C. F., Solomon, E. A., Seibert, D., and DiCarlo, J. J. Performance- optimized hierarchical models predict neural responses in higher visual cortex. Proceedings of the National Academy of Sciences, 111(23):8619–8624, 2014. Yang, C., Wan, H., Peng, Y., Cheng, X., Yu, Z., Zhang, J., Yu, J., Yu, X., Zheng, X., Zhou, D., and Wu, C. Reasoning via video: The first evaluation of video models’ reasoning abilities through maze-solving tasks. arXiv preprint arXiv:2511.15065, 2025a. URL https://arxiv.org/abs/2511.15065. 21 VIDEO-REASON.COMA Very Big Video Reasoning Suite Yang, J., Yang, S., Gupta, A. W., Han, R., Fei-Fei, L., and Xie, S. Thinking in space: How multimodal large language models see, remember, and recall spaces. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 10632–10643, 2025b. Yang, Z., Teng, J., Zheng, W., Ding, M., Huang, S., Xu, J., Yang, Y., Hong, W., Zhang, X., Feng, G., et al. CogVideoX: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. Zacks, J. M. Neuroimaging studies of mental rotation: A meta-analysis and review. Journal of Cognitive Neuroscience, 20(1):1–19, 2008. Zheng, Z., Peng, X., Yang, T., Shen, C., Li, S., Liu, H., Zhou, Y., Li, T., and You, Y. Open-SORA: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024. 22 VIDEO-REASON.COMA Very Big Video Reasoning Suite A. Details of Cognitive Architecture Does intelligence have a structure? Such discussion could date back to Aristotle, where cognition is treated not as a flat “inner theater”, i.e. in his mentor (Plato 1921, Theaetetus 189e–190a), but as an organization of dunameis, in other words, cognitive faculties. Knowing begins with aisth ˆ esis, perception, where the senses receive the form of external things without their matter, i.e. seeing “redness,” not absorbing the red object (Aristotle 1984f;g, De Anima 2.1, 412a19–23; 2.12, 424a17–24). The perceived then gets filtered through koin ̄ e aist ˆ hesis, in other words, common sense, where our judgments are supplied by core knowledge born with us, such as spatiality, intuitive physics, and logic thoughts (Carey & Spelke 1996; Carey 2009; Luo et al. 2025a; Li et al. 2025; Aristotle 1984g, De Anima 3.1, 425a27–b11). Phantasia serves as the substrate that preserves, recombines, organizes and transforms the perceptual empeiria, experiences, and stores them into mn ̄ em ̄ e, memory (Aristotle 1984a;g, De Anima 3.3, 427b14–428a9, On Memory, 449b3–453b11). Above empeiria and mn ̄ em ̄ e lies the no ˆ us, the uniquely human faculty of understanding and abstraction, and from empeiria, no ˆ us extracts katholou, universals, an arc that Aritsole foregrounds as he traces from aisth ˆ esis, phantasia, empeiria, mn ̄ em ̄ e to no ˆ us. Essentially, the path to katholou, for Aristole, is the path to knowledge and intelligent human behaviors, and therefore such is the flow of Aristotle’s cognitive architecture (Aristotle, 1984b, Metaphysics 1.1, 980a21–27). German Idealism is another undertaking where human thinkers try to theorize the structure of our cognition and mind. Kant posits that the mind does not simply mirror the but actively structures the world that we perceive as it’s (Kant 1781, CPR, A19–A24/B33–B38; A30–A33/B45–B49). He holds that raw sensory input can only become meaningful when organized by the mind’s built-in frameworks – sinnlichkeit, a priori forms of intuition, i.e. space and time, and verstand, categories of understanding, which include understanding of numbers, of quality, modality, of causality, and of relations (Kant 1781, CPR, A50–A51/B74–B75; A80/B106). Between sinnlichkeit and verstand lies einbildungskraft, where structured perceptual mental content would be sent to for transformation and synthesis. Kant gives particular descriptions of such transformation of mental content: sinnlichkeit first organizes the spatiotemporal shape of representations, ordering raw data as a unified intuition in space and time, and einbildungskraft brings together intuitions so that they can be combined, and finally verstand recognizes the unified data by subsuming it under a category (Kant 1781, CPR, A98–A110). At the apex lies vernunft, which offers the ultimate abstraction of mental representations, and generate ideas such as the soul, the self as a whole, the world, the universe as a totality, and God (Allison 2004; Longuenesse 1998; Guyer 1987; Kant 1781, CPR, A327/B384; A642/B670; A249–A252). Modern discussions on cognitive architecture usually combine with research in artificial intelligence and cognitive neuroscience. At the Dartmouth Workshop in 1956, where the term ”artificial intelligence” is coined, Herbert Simon and Alan Newell has presented Logical Theorist, a computer program which they had claimed to be the “the first artificial intelligence program” (Newell & Simon, 1956; McCorduck, 2004). While Logical Theorist is narrow and could only prove 38 of the first 52 theorems in Russell’s philosophy textbook, it sparks the first AI boom where people try to build principled computer programs that would produce human-like intelligent behaviors (Whitehead & Russell, 1927; Russell & Norvig, 2021). The General Problem Solver has shown that human reasoning could be simulated using symbolic production systems, establishing a computational account of problem solving, while critiques argue such is only a subset of what human intelligence could do (Newell et al., 1959; Newell & Simon, 1972; Newell, 1973). Full-scale cognitive architecture models soon have flourished, including ACT-R, SOAR, and CLARION models (Anderson, 2013; Laird et al., 1987; Anderson & Lebiere, 1998; Anderson, 1993; 2007; Langley et al., 2009; H ́ elie & Sun, 2010). As techniques in neuroscience have developed, a dialogue has emerged between cognitive architectures and brain architectures (Petersen & Sporns, 2015). Human brain is interestingly modular, with cognitive functions specifically implemented in particular physical anatomical locations, yet these specialized regions communicate constantly to orchestrate cognition (d’anthropologie de Paris, 1898; Wernicke, 1874; Kanwisher, 2010). Recently, cognitive neuroscientists hope to map out the structural and functional connectivity between brain areas as to understand the architectural organization underlying cognition (Sporns et al., 2005; Bertolero et al., 2015). Explicit combinations of the cognitive architecture approaches and connectomics approaches have 23 VIDEO-REASON.COMA Very Big Video Reasoning Suite been carried out in research in recent years and yielded validations of some of the hypothesized architectural claims about human cognition in the human brain (Palmeri et al., 2017; Yamakawa, 2021; Chakladar, 2024). Sitting on the shoulder of philosophy giants and combining works from cognitive architecture modeling and neuroscience, we have consolidated a cognitive architecture made of perception, transformation, spatiality, abstraction, and knowledge. A.1. Perception Perception is the the extraction of form without matter (Aristotle, 1984g, De Anima 2.12, 424a17–24). Helmholtz calls this as ”unconscious inference” (von Helmholtz, 1867; Helmholtz, 1924). The visual system exemplifies perception’s hierarchical architecture. V1 neurons function as oriented edge detectors organized into columnar structures reflecting natural image statistics (Hubel & Wiesel, 1962; 1968; Bonhoeffer & Grinvald, 1991; Hubel & Wiesel, 1977; Olshausen & Field, 1996). Beyond V1, the ventral and dorsal streams achieve progressively invariant object recognition (Milner & Goodale, 2008; Goodale & Milner, 1992; DiCarlo et al., 2012; Tanaka, 1996), with category-selective regions like the fusiform face area emerging in inferotemporal cortex (Kanwisher et al., 1997; Tsao et al., 2006; Quian Quiroga et al., 2023), where cells collectively map a low-dimensional ”object space” (Bao et al., 2020). Just as Kant has hypothesized (Swanson, 2016), the brain generates predictions about sensory input and computes prediction errors, with representational geometry shaped by environmental regularities (Rao & Ballard, 1999; Friston, 2010; Keller & Mrsic-Flogel, 2018; Greco et al., 2024; Kersten et al., 2004; Ma et al., 2006). Multimodal integration binds signals across modalities through causal inference in regions like superior temporal cortex (Ghazanfar & Schroeder, 2006; Stein & Stanford, 2008; Calvert, 2001; K ̈ ording et al., 2007; Shams & Beierholm, 2010). Marr’s tri-level framework connects biological and artificial vision (Marr, 1982): CNNs exhibit cortex-aligned representations, though they lack robust higher-level processing found in biological systems (Yamins et al., 2014; Kriegeskorte, 2015; Kazemian et al., 2025; Baker et al., 2018; Geirhos et al., 2021). Gibson’s ecological approach reminds us that perception ultimately serves action through affordances (Gibson, 2014)—connecting to Aristotle’s view that aisth ˆ esis serves the organism’s engagement with its environment. Perception constitutes the entry point of human cognitive architecture, a thesis from Marr and Gibson to modern deep learning (Gibson, 2014; Marr, 1982; Yamins et al., 2014; Kriegeskorte, 2015; Kazemian et al., 2025; Baker et al., 2018; Geirhos et al., 2021). A.2. Transformation Transformation refers to the cognitive faculty that manipulates, recombines, and synthesizes mental representa- tions, corresponding to Aristotle’s phantasia and Kant’s Einbildungskraft. Aristotle insists that “whenever one contemplates, one necessarily at the same time contemplates in images,” positioning phantasia not as passive storage but as active transformation of empeiria into material suitable for no ˆ us (Aristotle 1984g, De Anima 3.7, 431a16–17; Wedin 1988). Kant systematizes this through his three syntheses, apprehension gathering the sensory manifold, reproduction in imagination holding representations together, and recognition subsuming the unified manifold under categories, with Einbildungskraft functioning productively to generate experiential structure by binding intuitions according to Verstand (Kant 1781, CPR, A98–A110; Longuenesse 1998; Allison 2004). Modern cognitive science operationalizes these insights: scientists have demonstrated that mental rotation operates analogically on quasi-spatial representations, proposing that mental images are constructed in a “visual buffer” for manipulation (Shepard & Metzler, 1971; Kosslyn, 1980; 1996). Another work has showed that focused attention binds preattentively registered features into coherent objects, a process that breaks down with parietal lesions, echoing Kant’s synthesis of apprehension (Kant 1781, CPR, A98–A110; Treisman & Gelade, 1980). Neuroscience suggests the posterior parietal cortex, particularly the intraparietal sulcus, as the substrate for transformation (Zacks, 2008; Papadopoulos et al., 2018), while cognitive architectures like ACT-R formalize this through an imaginal module with parietal correlates (Anderson, 2007). Transformation thus occupies the crucial intermediate position both Aristotle and Kant recognized: between raw aisth ˆ esis or Sinnlichkeit and the higher operations of no ˆ us or Vernunft. 24 VIDEO-REASON.COMA Very Big Video Reasoning Suite A.3. Spatiality For Kant, space constitutes an a priori form of sinnlichkeit—a transcendental precondition for perception itself, not something derived from experience (Kant 1781, CPR, A22–A24/B37–B40). This insight finds support in developmental psychology identifying spatiality as a core knowledge system: infants possess domain-specific machinery for representing places and their geometric relationships, with deep phylogenetic roots observable across species and cultures (Spelke & Kinzler, 2007; Spelke, 2000). Tolman first proposed that animals construct internal “cognitive maps” enabling flexible navigation beyond simple stimulus-response associations (Tolman, 1948). The neural substrates were later revealed: place cells in the hippocampus fire at specific locations (O’Keefe & Dostrovsky, 1971), while grid cells in the entorhinal cortex provide metric coordinates through hexagonal firing patterns (Hafting et al., 2005), which are discoveries recognized by the 2014 Nobel Prize as “a positioning system in the brain” (Moser & Moser, 2014; Nobel Prize Committee, 2014). The posterior parietal cortex complements this system by organizing sensory coordinates into motor-relevant reference frames (Colby & Goldberg, 1999; Andersen et al., 1997). Beyond representation, humans perform spatial operations such as mental rotation, where reaction times increase linearly with angular disparity (Shepard & Metzler, 1971; Hegarty et al., 2006). Computational models that are grid-cell-inspired have been proposed, but human-level performances remain challenging (Banino et al., 2018; Whittington et al., 2020; Behrens et al., 2018). Spatiality serves as a fundamental function, a prerequisite for perception, scaffold for memory, and substrate for reasoning in human cognitive architecture. A.4. Abstraction Abstraction represents the apex of our cognitive architecture, where embodied experiences are distilled into generalizable knowledge. For Aristotle, no ˆ us extracts katholou, universals, from accumulated empeiria through selective attention to essential features x(Aristotle, 1984b, Metaphysics 1.1, 980a21–27; Renero, 2013; Aristotle, 1984g, De Anima 3.4–8; Shields, 2016). Kant’s Vernunft similarly generates transcendental ideas extending beyond the categories of Verstand, though exceeding the boundaries of possible experience (Kant, 1781, CPR, A327/B384; A642/B670; Allison, 2004; Guyer, 1987). Contemporary accounts elaborate these mechanisms: one developmental psychology work demonstrates that abstract concepts emerge through Quinian Bootstrapping, producing representational systems with greater expressive power than core cognition, with concepts possessing causally deep “cores” (Carey, 2009; Carey, 2011; Carey, 2009; Keil, 1992). Gentner’s structure-mapping theory shows how analogical reasoning maps relational patterns across domains, with systematicity determining preferred mappings and enabling children to extract relational abstractions (Gentner, 1983; Gentner & Hoyos, 2017; Holyoak, 2012; Gentner, 1983; Gentner & Hoyos, 2017; Holyoak et al., 2001). Neuroscience reveals abstraction’s hierarchical implementation: the lateral PFC exhibits a rostral-to-caudal gradient whereby anterior regions process increasingly abstract information, with rostral PFC serving as a “gateway” for internally-generated thought (Badre & Nee, 2018; Christoff et al., 2009; Badre & Nee, 2018; Brincat et al., 2018; Burgess et al., 2007; Ramnani & Owen, 2004). Memory systems contribute through schema-guided consolidation, with hippocampus and vmPFC maintaining abstract prototype representations that support generalization (Bowman & Zeithamova, 2018; Gilboa & Marlatte, 2017; Tse et al., 2011; McClelland et al., 2020). Hierarchical abstraction might underly deep learning’s layered representations and cognitive architectures’ multiple processing levels (LeCun et al., 2015; Bengio et al., 2013; Anderson, 2007; Laird et al., 1987; Tomov et al., 2020; Luo et al., 2025b). Abstraction integrates the outputs of perception, transformation, and spatiality to achieve the katholou, knowledge, that Aristotle recognized as the telos of intelligent behavior (Aristotle, 1984d, Posterior Analytics 2.19, 100a3–100b5). A.5. Knowledge Aristotle opens the Metaphysics with the claim that “all human beings by nature desire to know” (Aristotle, 1984b, Metaphysics 980a21), and devotes the closing chapter of the Posterior Analytics to explaining how they do so. Knowledge begins in aisth ˆ esis, sense perception, which deposits mn ˆ em ˆ e, memory; repeated memories of the same kind consolidate into empeiria, experience; and from experience no ˆ us abstracts the universal first principles that ground epist ˆ em ˆ e, demonstrative knowledge (Aristotle, 1984d, Posterior Analytics I.19, 25 VIDEO-REASON.COMA Very Big Video Reasoning Suite 99b15–100b17). The progression is neither purely empirical nor purely rational: perception furnishes the material, but it is the active intellect that grasps the universal in the particular, “as when a rout has occurred in battle, first one soldier makes a stand, then another, until the original formation is restored” (Aristotle, 1984d, Posterior Analytics 100a12–13). For Aristotle, then, human cognition is structured from the outset by capacities that go beyond the sensory given, no ˆ us does not learn first principles from experience so much as recognize them through it. Kant radicalizes this insight. Whereas Aristotle allows that universals are latent in perception and extracted by intellect, Kant argues that the mind actively constitutes the very form of experience. “Thoughts without content are empty; intuitions without concepts are blind” (Kant, 1781, CPR A51/B75): knowledge requires both sensible Anschauungen delivered by receptivity and the Kategorien of Verstand imposed by spontaneity. The categories, substance, causality, unity, plurality, and the res, are not abstracted from experience but are its a priori conditions of possibility (Kant, 1781, CPR A80/B106; Longuenesse, 1998). The transcendental deduction establishes that these concepts must apply to any object of experience whatsoever, because it is only through their application that the manifold of intuition is synthesized into coherent representation (Kant, 1781, CPR B129–B169; Allison, 2004). Knowledge, for Kant, is thus never a passive reception but an active structuring: the mind brings to experience the very framework within which experience becomes intelligible. Contemporary developmental psychology has furnished striking empirical support for this rationalist inher- itance. Research on core knowledge demonstrates that human infants possess, from the earliest months of life, a set of domain-specific representational systems that structure cognition prior to and independently of explicit instruction (Spelke & Kinzler, 2007). These representations are abstract, often amodal, and operate according to principles that go well beyond what associative learning from perceptual statistics could deliver (Baillargeon et al., 1985). Carey’s account of conceptual change further shows that later, explicit knowledge is constructed not ex nihilo but through the enrichment, combination, and bootstrapping of these core systems, a process she terms Quinian bootstrapping (Carey, 2009). Figure 7 Distribution of 150 visual reasoning tasks across five cognitive faculties in the VBVR-Dataset. 30.0% 22.0% 19.3% 15.3% 13.3% 150 Tasks Cognitive Faculties Perception (45) Abstraction (33) Transformation (29) Knowledge (23) Spatiality (20) Table 6 Complete Cognitive Taxonomy of VBVR-Dataset (150 Visual Reasoning Tasks) Task IDCognitive CategoryDescription ABSTRACTION (33) G-7AbstractionReturn objects to correct bin by category G-26AbstractionMaintain object identity across different objects G-29AbstractionFind extreme values in chart (with data labels) G-37AbstractionIdentify random symmetry patterns G-38AbstractionIdentify shape symmetry patterns G-41AbstractionGrid highest cost path finding G-44AbstractionBreadth-first search traversal G-49AbstractionComplete missing contour segments G-51AbstractionPredict next color in sequence G-131AbstractionSelect next figure in increasing size sequence G-133AbstractionSelect next figure in decreasing size sequence G-134AbstractionSelect next figure in large-small alternating sequence G-135AbstractionSelect next figure in small-large alternating sequence G-193AbstractionDraw next sized shape in pattern O-7AbstractionShape color change operations O-8AbstractionShape rotation operations Continued on next page 26 VIDEO-REASON.COMA Very Big Video Reasoning Suite Table 6 — continued from previous page Task IDCognitive CategoryDescription O-9AbstractionShape scaling operations O-10AbstractionShape outline fill operations O-11AbstractionShape color then move (compound operation) O-12AbstractionShape color then scale (compound operation) O-13AbstractionShape outline then move (compound operation) O-14AbstractionShape scale then outline (compound operation) O-21AbstractionConstruction blueprint interpretation O-29AbstractionBall color clustering and merging O-30AbstractionBookshelf organization task O-37AbstractionLight sequence pattern recognition O-43AbstractionObject subtraction (quantity reasoning) O-45AbstractionSequence completion task O-47AbstractionSliding puzzle solving O-49AbstractionSymmetry completion task O-54AbstractionControl panel symbol manipulation O-56AbstractionRaven’s Progressive Matrices reasoning O-66AbstractionAnimal color sorting task KNOWLEDGE (23) G-27KnowledgeRead chart data and semantic comprehension G-30KnowledgeFind extreme values in chart (without data labels) G-35KnowledgeHit target after bounce G-48KnowledgeMultiple bounces prediction G-160KnowledgeCircle largest numerical value G-162KnowledgeLocate twelve o’clock arrows G-163KnowledgeIdentify digits 1 and 9 G-200KnowledgeCircle maximum value in set G-217KnowledgeCircle central dot G-247KnowledgeIdentify Chinese character G-273KnowledgeHigh density liquid behavior O-3KnowledgeSymbol reordering operations O-15KnowledgeBall bounces at given time O-18KnowledgeGlass refraction O-19KnowledgeMirror reflection O-23KnowledgeDomino chain branch path prediction O-24KnowledgeDomino chain gap analysis O-34KnowledgeDot-to-dot connection task O-52KnowledgeTraffic light state reasoning O-53KnowledgeClock reading and time reasoning O-62KnowledgeGravity physics simulation O-75KnowledgeCommunicating vessels (fluid dynamics) O-87KnowledgeFluid diffusion reasoning PERCEPTION (45) G-3PerceptionStable sort objects maintaining order G-4PerceptionIdentify and distinguish different objects G-5PerceptionMulti-object placement to specified positions G-9PerceptionIdentify objects in specified region G-19PerceptionSort objects by specified rule G-22PerceptionAttention shift to same object G-39PerceptionAttention shift to different object G-43PerceptionUnderstand scene structure and spatial layout G-54PerceptionConnecting matching colors G-132PerceptionFind fragment for gap filling G-136PerceptionLocate point in overlapping area G-137PerceptionIdentify figure in overlapping area G-138PerceptionSpot unique non-repeated color G-141PerceptionIdentify polygon with most sides G-143PerceptionSelect box with most dots G-146PerceptionCircle all squares from mixed shapes G-147PerceptionIdentify unique figure in uniform set G-158PerceptionIdentify all hollow points G-161PerceptionMark second largest shape G-165PerceptionMark tangent point after motion G-166PerceptionHighlight horizontal lines G-167PerceptionSelect longest polygon side G-168PerceptionIdentify rectangle nearest to square G-169PerceptionLocate intersection of line segments G-174PerceptionArrange circles by circumference G-189PerceptionDraw midpoint perpendicular line G-195PerceptionSelect nearest 2:1 rectangle G-198PerceptionMark right-angled triangles G-199PerceptionLocate line intersections G-202PerceptionMark wave peaks Continued on next page 27 VIDEO-REASON.COMA Very Big Video Reasoning Suite Table 6 — continued from previous page Task IDCognitive CategoryDescription G-206PerceptionIdentify pentagons G-212PerceptionFind incorrect arrow direction G-218PerceptionIdentify largest angle in triangle G-222PerceptionMark tangent point of circles G-223PerceptionHighlight horizontal lines G-248PerceptionMark asymmetrical shape G-250PerceptionColor triple intersection red O-1PerceptionColor mixing (additive) O-2PerceptionPigment color mixing (subtractive) O-16PerceptionColor addition operations O-17PerceptionColor subtraction operations O-31PerceptionBall eating mechanics O-33PerceptionCounting objects accurately O-38PerceptionIdentify majority color in set O-65PerceptionAnimal size sorting SPATIALITY (20) G-12SpatialityGrid navigation to obtain reward G-13SpatialityGrid number sequence navigation G-14SpatialityGrid color sequence navigation G-15SpatialityGrid navigation avoiding obstacles G-16SpatialityGrid navigation going through blocks G-17SpatialityGrid navigation avoiding red blocks G-18SpatialityGrid shortest path finding G-31SpatialityDirected graph navigation G-32SpatialityUndirected graph navigation G-33SpatialityVisual Jenga game G-45SpatialityKey-door matching puzzle G-46SpatialityFind keys and open doors G-47SpatialityMultiple keys for one door puzzle G-140SpatialityLocate topmost unobscured figure G-219SpatialitySelect leftmost shape G-221SpatialityOutline innermost square O-25SpatialityLEGO construction assembly O-39SpatialityMaze navigation and solving O-55SpatialityRotation operations O-83SpatialityPlanar warp verification TRANSFORMATION (29) G-1TransformationPredict object trajectory G-2TransformationReorder objects by rule G-6TransformationResize object G-8TransformationTrack object movement G-11TransformationHandle object reappearance after disappearance G-21TransformationMultiple occlusions (vertical) G-24TransformationSeparate objects (no rotation) G-25TransformationSeparate rotating objects G-34TransformationObject packing optimization G-36TransformationMultiple occlusions (horizontal) G-40TransformationCombined rotating objects G-50TransformationSuppress spurious edges G-194TransformationConstruct concentric ring G-240TransformationAdd borders to unbordered shapes O-4TransformationSymbol substitution operations O-5TransformationSymbol deletion operations O-6Transformation2D geometric transformation O-22TransformationConstruction stack (gravity balance) O-27TransformationMove 2 objects to 2 targets O-32TransformationRolling ball physics O-36TransformationGrid shift operations O-44TransformationRotation puzzle O-46TransformationShape sorter classification O-58TransformationSymbol delete operations O-59TransformationSymbol insert operations O-60TransformationSymbol substitute operations O-61TransformationSymbol edit operations O-64TransformationAnimal matching task O-85Transformation2D object rotation 28 VIDEO-REASON.COMA Very Big Video Reasoning Suite B. Data Curation This appendix provides a comprehensive account of the data curation pipeline that underlies the VBVR-Bench. We detail the five-stage process from task design to benchmark construction, the quality assurance mechanisms, and the infrastructure that enables large-scale data generation. B.1. Overview of Data Curation Pipeline The VBVR-Bench is built through a systematic five-stage pipeline that transforms cognitive task concepts into a standardized evaluation framework. Pipeline Flow. The data curation pipeline consists of five sequential stages. Stage 1: Task Design We begin by designing cognitive reasoning tasks grounded in cognitive science principles. Starting from over 300 task candidates, we employ a dual-review process to select tasks that meet six quality criteria. This stage produces two distinct task sets: 100 training tasks and 100 testing tasks with an overlapping of 50 tasks. Stage 2: Task Implementation Each training task is implemented as a parameterized generator capable of producing diverse samples. These generators form the our data spring, VBVR-DataFactory, a set of modular synthesis engines. All generators inherit from a standardizedBaseGeneratortemplate and are managed as independent repositories within a GitHub Organization. Implementation undergoes rigorous quality control by dedicated reviewers who verify scalability, code quality, and adherence to file specifications. Stage 3: Large-Scale Distributed Generation. The VBVR-DataFactory infrastructure orchestrates parallel generation of one million training samples (10,000 per task) stored privately on AWS S3. Simultaneously, we generate 500 fresh test cases (5 per task) from the Test task set. Each sample comprises an initial frame, task prompt, ground-truth solution, and reference video. Stage 4: Model Evaluation. We employ VBVR-EvalKit to evaluate eight state-of-the-art image-to-video (I2V) models on the 500 test cases, producing 4,000 generated videos. Human annotators assess each video across three reasoning-specific dimensions: Task Completion, Reasoning Logic, and Visual Quality, each scored on a 1-5 scale. Stage 5: Benchmark Construction and Model Training. The evaluation data is integrated into a standardized benchmark dataset. We train the VBVR-Wan on the one million training samples, demonstrating that explicit reasoning supervision significantly improves performance on reasoning-intensive tasks. B.1.1. Key Statistics Table 7 summarizes the scale and coverage of the VBVR benchmark: Table 7 Data Construction Statistics ComponentScaleDescription Task Candidates300+Initial task pool Training Tasks100Generate training data Test Tasks10050 In-Domain + 50 Out-of-Distribution Training Samples1,000,000100 tasks × 10,000 samples per task Test Cases7,500150 tasks × 50 samples per task Cognitive Categories5Systematic coverage Evaluated Models8State-of-the-art I2V models, 4 open-source, 4 commercial Generated Videos4,0008 models × 500 cases Evaluation Records4,000Task Completion/Reasoning Logic/Visual Quality annotations Development Team7553 OSS coders + 7 full-time employees + 5 QC reviewers Code Repositories300+Github Organization B.1.2. Design Principles The pipeline is guided by the following core design principles: 1. Dual Generalization Testing. The In- Domain/OOD split enables evaluation of both in-task generalization (seen task types with unseen samples) and cross-task generalization (unseen task types). 2. Reasoning-First Paradigm. Unlike benchmarks that primarily assess generation quality, VBVR emphasizes reasoning correctness through TC, RL, and VQ metrics that capture task completion, logical consistency, and visual fidelity. 3. Industrial-Scale Quality Assurance We employ a dual-review mechanism: peer review for task design (by five task designers) and dedicated 29 VIDEO-REASON.COMA Very Big Video Reasoning Suite quality control for implementation (by more than two specialized reviewers). Six quality criteria ensure tasks are well-specified, visually clear, and scalable. 4. Extensible Infrastructure. The modular generator architecture and comprehensive evaluation toolkit facilitate easy extension with new tasks and models. 5. Cognitive Inspiration. Our task design began with a loose cognitive taxonomy inspired by cognitive science literature. However, rather than rigidly adhering to a fixed theoretical framework, we adopted a task-driven design philosophy: we prioritized designing intellectually meaningful reasoning tasks first, then organized them into our categories. 6. Diversity Focus. This approach contrasts with top-down frameworks that first define strict categories, then fill them with tasks. Our bottom-up approach ensured that each task possesses genuine reasoning value, and maximizing our diversity, rather than forcing tasks to fit narrow pre-determined boxes. B.1.3. Task Selection and Filtering Initial Pool We designed over 300 task candidates through iterative brainstorming sessions. Each designer independently proposed tasks, which were then collectively discussed and refined. Selection Criteria (Six Standards) Tasks were evaluated against six criteria during peer review: 1.Information Sufficiency: The first frame must contain all information necessary for reasoning, without requiring external context. ✓Pass Example: G-15 (Grid Avoid Obstacles) — the first frame shows start point, end point, and all obstacles. × Fail Example: Tasks where the goal object is initially occluded or requires additional verbal explanation. 2.Deterministic Solution: Tasks should have a clear, unique solution (or explicitly defined criteria for multiple valid solutions, e.g., “shortest path”). ✓ Pass Example: G-18 (Grid Shortest Path) — the optimal path is unambiguous. × Fail Example: Open-ended tasks with ambiguous success conditions. 3.Video-Based Reasoning: Tasks must be suitable for video generation models, encompassing diverse reasoning types (temporal dynamics, static recognition, logical inference). ✓ Pass Example: O-23 (Domino Chain) — requires temporal causality reasoning. !Note: Unlike some benchmarks, we do NOT exclude static recognition tasks; we include both dynamic and static reasoning. 4. Visual Clarity: Objects must be distinguishable, text/numbers legible, and layouts uncluttered to avoid perceptual ambiguity. 5.Parametric Diversity (Scalability): The parameter space must be large enough to generate 10,000 non-trivial, distinct samples per task. 6.Technical Feasibility: Tasks must be implementable with our rendering pipeline (PIL-based), output standard file formats, and avoid edge cases (e.g., initial occlusions, boundary overflow). Rejection Cases Here are some of the tasks were excluded: • A. Multi-Step Logic Chains – Example: Logic Gate Circuits – Reason: Reasoning chains too long to clearly visualize in video format. • B. Complex Physics Requiring Precise Numerical Solutions 30 VIDEO-REASON.COMA Very Big Video Reasoning Suite – Example: Elastic Collisions with friction coefficients –Reason: Continuous parameters yield no deterministic solution (tiny parameter variations lead to drastically different trajectories); I2V models cannot precisely simulate physics. • C. Excessively Difficult Tasks – Example: Sudoku – Reason: High difficulty even for humans; unsuitable for video format (better suited for static reasoning); evaluation criteria too strict (any single cell error constitutes failure). B.1.4. Peer Review Process All task designs underwent a structured peer review by the five task designers, with each designer evaluating others’ proposals using a standardized checklist derived from the six quality criteria; tasks typically progressed through 2–3 rounds of iterative refinement based on this feedback, during which common issues were identified, including ambiguous success conditions, insufficient visual clarity, parameter spaces too small to support 10k samples, and unintentional edge cases (e.g., objects overlapping at initialization). A task advanced to implementation (Stage 2) only after receiving consensus approval from at least three reviewers. This process ensured that every task in VBVR meets rigorous cognitive and technical standards, forming a solid foundation for the subsequent implementation and evaluation stages. B.2. Details of Task Implementation This section details the implementation of the 200 tasks as parameterized generators. We describe the standardized architecture, code management infrastructure, parameterization strategies, visual rendering pipeline, and dedicated quality control procedures. B.2.1. Generator Architecture and Standardization BaseGenerator TemplateAll task generators inherit from a standardizedBaseGeneratorabstract base class, ensuring consistent interfaces and output formats. The template defines: 1. Required Methods: •init(self, seed, params): Initialize generator with random seed and task-specific parameters. • generatesample(self) -> Sample: Generate a single task instance. • validatesample(self, sample) -> bool: Verify sample meets quality criteria. • save sample(self, sample, outputdir): Save sample to standardized file structure. 2. Standard Output Format: Each sample is saved as a directory containing: • first frame.png: Initial state image (required) • prompt.txt: Task instruction in natural language (required) • finalframe.png or goal.txt: Target state or goal description • ground truth.mp4: Reference solution video demonstrating the correct reasoning process 3.Configuration Management: Generators expose configurable parameters through aconfig.yaml file specifying: • Parameter ranges (e.g., grid size: 5-10, number of objects: 3-8) • Difficulty levels (easy/medium/hard) • Visual settings (resolution: 512×512, frame rate: 24 fps) 31 VIDEO-REASON.COMA Very Big Video Reasoning Suite Design Rationale. TheBaseGeneratortemplate serves three core purposes: Consistency, by providing uniform interfaces that simplify integration with VBVR-DataFactory for batch generation; Quality Assurance, through thevalidatesample()method, which enforces critical checks such as ensuring no object occlusions and that valid solutions exist; and Extensibility, enabling new tasks to be added by implementing the abstract methods while leveraging shared utility functions. B.2.2. Code Management Infrastructure All task generators are managed as independent repositories within a GitHub Organization, an archi- tecture that provides clear organizational and operational advantages: each repository follows a strict Naming Convention ofType-IDtasknamedata-generator(e.g.,G-3stablesort- data-generatorandO-15ballbouncesgiventimedata-generator), where Type denotes G, which are contributed by commercial enterprises, or O, which are contributed by OSS developers, and ID is a unique numeric identifier; this structure enables Independent Versioning, as each task maintains its own commit history, tags, and release cycles, and supports Modular Updates, ensuring that bug fixes or parameter adjustments to one task do not affect others. All generators share thecore/directory structure through Git submodules or package dependencies, ensuring consistent utility functions while allowing each task to evolve independently. This design yields several version control benefits: 1. Traceability, as each sample records the exact generator version (git commit hash) used for generation; 2. Reproducibility, enabling researchers to check out specific generator versions to reproduce sample generation; and 3. Collaboration, allowing multiple team members to work on different tasks simultaneously without merge conflicts. Figure 8 This is a typical example of data samples in our dataset. The model receives a prompt and a first image, and is asked to generate a video that solves the prompt. Final image and ground truth videos are provided as references in data samples. Prompt First ImageFinal Image QuestionAnswer Inference Move the green dot from its starting position through the maze paths to the red flag. B.2.3. Parameterization Strategy for Diversity To generate 10,000 distinct, non-trivial samples per task, we employ systematic parameterization strategies: Example 1: G-15 (Grid Avoid Obstacles) Parameters: - grid_size: [5, 6, 7, 8, 9, 10] # 6 options - num_obstacles: [3, 4, 5, ..., 15] # 13 options - start_position: grid-dependent # ̃grid_size² options - end_position: grid-dependent # ̃grid_size² options - obstacle_layout: random_valid # ̃10ˆ6 variations Estimated unique samples: > 10ˆ9 Sampling strategy: Constrained random sampling ensuring solvability Example 2: O-56 (Raven’s Progressive Matrices) Parameters: - pattern_type: [size_progression, rotation, color_change, shape_replacement, combination] # 5 types - num_objects: [1, 2, 3, ..., n] # n options - object_shapes: [circle, square, triangle, ...] # 8 shapes 32 VIDEO-REASON.COMA Very Big Video Reasoning Suite - progression_complexity: [simple, compound] # 2 levels - visual_style: [minimal, decorated] # 2 styles Estimated unique samples: > 10ˆ5 Sampling strategy: Balanced distribution across pattern types Example 3: G-3 (Stable Sort) Parameters: - num_shape_types: [2, 3, 4] # 3 options - shapes_per_type: [2, 3, 4] # 3 options - color_palette: [vibrant, pastel, ...] # 5 palettes - initial_layout: random_permutation # n! permutations - shape_geometry: [geometric, organic] # 2 styles Estimated unique samples: > 10ˆ6 Sampling strategy: Uniform sampling across difficulty levels Diversity is enforced through four complementary mechanisms: stratified sampling, which divides the parameter space into strata (e.g., difficulty levels) and samples proportionally from each; constraint satisfaction, which ensures that all generated samples are valid (e.g., mazes admit solutions and puzzles are solvable); duplicate detection, which tracks parameter combinations via hash functions to prevent exact duplicates; and visual variety, which randomizes visual attributes such as colors, sizes, and positions beyond the core reasoning parameters. Each generated sample undergoes automatic validation, including a solvability check to verify that a ground- truth solution exists (e.g., via A* search for navigation tasks), a visual clarity check to ensure that objects do not overlap excessively and that text remains legible with a minimum font size, and a parameter bounds check to confirm that all parameters fall within their specified ranges. B.2.4. Visual Rendering Pipeline All visuals adhere to fixed specifications: a resolution of 512×512 pixels, 24-bit RGB color depth (8 bits per channel), a frame rate of 24 fps for ground-truth videos, and H.264 as the video codec. B.3. Large-Scale Distributed Generation of Training Data The VBVR-DataFactory component orchestrates the parallel generation of one million training samples using the 300+ generators in our data spring. It provides a complete, production-grade serverless system for cloud infrastructure, generation workflow, data organization, and quality assurance, enabling efficient, reliable, and cost-effective large-scale data production. All generated samples are stored on Amazon Web Services (AWS) Simple Storage Service (S3). S3 is chosen for its ability to scale to petabyte-level data with automatic capacity management, its high durability through redundant storage across multiple availability zones, and its fine-grained access control through IAM policies. Training data are kept in private buckets with server-side encryption to prevent public exposure and test-set contamination. In addition, S3’s tiered storage model makes long-term archival cost-effective. To efficiently generate one million samples, VBVR-DataFactory employs distributed serverless processing via AWS Lambda. The system distributes tasks across up to 990 concurrent Lambda function instances, each configured with 3 GB memory and a 15-minute execution timeout. Generation tasks are queued in Amazon SQS and automatically trigger Lambda invocations, eliminating the need for manual cluster management. Within each task, samples are produced in configurable batches of 25–100 to balance memory usage and processing efficiency. A typical one-million-sample generation completes in approximately 2–4 hours with the full fleet of 990 concurrent workers, at a total cost of roughly $800–1200 per run (primarily Lambda compute 33 VIDEO-REASON.COMA Very Big Video Reasoning Suite and S3 storage). Fault tolerance is built in through SQS’s automatic retry mechanism with configurable visibility timeouts, a Dead Letter Queue (DLQ) for failed messages after one retry attempt, and CloudWatch metrics for real-time monitoring of task success rates and processing durations. The end-to-end generation pipeline begins with configuration and planning, where generators, sample counts, random seeds, and output formats are defined. Tasks are submitted to the SQS queue as Pydantic-validated JSON messages, each specifying the generator type, number of samples, starting index for global sample numbering, random seed, and output format (individual files or compressed tar archives). Lambda functions receive these messages, execute the corresponding generator subprocess, validate and rename outputs with zero- padded global indices, upload results to S3 with structured prefixes, and emit success metrics to CloudWatch. Failed tasks automatically move to the DLQ for manual inspection and resubmission after debugging. On a per-task basis (100 samples, typical batch size), sample generation requires approximately 15 minutes, with generation speed varying significantly by task complexity—simple geometric tasks average 1–3 seconds per sample, while complex graph or physics tasks require 9–15 seconds per sample. File organization and S3 upload add 1–2 minutes depending on file sizes, and Lambda cold starts contribute an additional 5–10 seconds for the first invocation. Each task therefore completes in roughly 15–20 minutes on a Lambda instance, enabling the entire one-million-sample corpus to be produced in 2–4 hours with 990 parallel workers. For efficient access during training, VBVR-DataFactory maintains a hierarchical S3 structure organized by generator and task, with each sample directory containing standardized files (firstframe.png, prompt.txt, and optionally finalframe.png and groundtruth.mp4). This organization allows training pipelines to locate, filter, and stream data efficiently at scale through S3’s prefix-based listing and parallel download capabilities. Quality assurance is integrated throughout generation. Every sample is validated by the generator’s internal checks before being saved, ensuring solvability, visual clarity, and file integrity. The system continuously monitors task success rates, processing durations, and samples uploaded through CloudWatch metrics, with alarms configured to detect anomalies such as high failure rates or DLQ accumulation. Sample diversity is enforced through unique per-task random seeds that increment deterministically across batches, preventing duplicate generation. The DLQ captures and preserves failed tasks for post-mortem analysis, enabling systematic debugging of generator issues and infrastructure failures. Across typical production runs, task validation failures occur in fewer than 1 percent of cases, primarily due to edge-case parameter combinations that are caught by Pydantic schema validation before reaching the generator. Infrastructure-related failures (timeouts, out-of-memory errors) occur in approximately 0.1–0.5 percent of tasks and are resolved through DLQ resubmission after adjusting batch sizes or addressing generator bugs. All training samples are stored in private S3 buckets with server-side encryption (SSE-S3), versioning disabled to reduce costs, and IAM role-based access control restricted to authorized Lambda functions and training pipelines. Full S3 access logging is enabled for audit trails. The data contain no personally identifiable information, as all samples are fully synthetic and procedurally generated. Together, these design choices demonstrate that, with proper serverless engineering, million-scale, high-quality data generation is achievable, reliable, and cost-effective. The resulting one million samples form the training foundation for VBVR-Wan and future video reasoning systems. 34 VIDEO-REASON.COMA Very Big Video Reasoning Suite C. VBVR-Bench Details C.1. Model Inference Infrastructure VBVR-Bench is an end-to-end evaluation framework that integrates large-scale model inference (or API-based invocation for proprietary models) with a rule-based evaluation engine over a standardized 100-task video reasoning benchmark. At the model level, VBVR-Bench provides a unified VideoGenerator abstraction that encapsulates 29 video generation systems, including closed-source models such as Google Veo, OpenAI Sora, and Runway Gen-4, as well as open-source and research models such as CogVideoX, Wan2.2, LTX-2, and HunyuanVideo. This abstraction ensures that all models are executed through an identical pipeline, eliminating confounding factors introduced by model-specific inference logic. It supports batch execution at scale, incorporates caching mechanisms to avoid redundant generations, and is fully reusable, allowing interrupted evaluations to continue without loss of progress. C.1.1. Model Selection and Configuration For the VBVR benchmark, we evaluate eight state-of-the-art image-to-video (I2V) models representing diverse architectures and capabilities: Table 8 Evaluated Models ModelDeveloperArchitectureParametersKey Features CogVideoX1.5-5B-I2VTsinghua University / Zhipu AITransformer-based diffusion model5 billionStrong text-image alignment Wan2.2-I2V-A14BWanx AILatent diffusion with attention mechanisms14 billionHigh-fidelity temporal consistency LTX-2LightricksEfficient transformer architecture∼3 billionFast inference HunyuanVideo-I2VTencent HunyuanMulti-stage diffusion with spatial-temporal attention∼7 billionStrong on complex scenes VeoGoogle DeepMindProprietary (closed-source)Unknown (large-scale)Strong physics understanding SoraOpenAIProprietary (closed-source, rumored diffusion transformer)Unknown (likely >10B)Temporal coherence, realistic physics Runway Gen-3Runway MLProprietary (closed-source)UnknownCreative generation, text adherence Kling 2.6Kuaishou TechnologyProprietary (closed-source, likely diffusion-based)UnknownFast inference The models included in the benchmark are chosen to reflect the breadth and maturity of the current image- to-video landscape. The set intentionally mixes open-source and closed-source systems, ensuring that both academic and industrial approaches are represented. It spans a wide range of architectural paradigms, capturing fundamentally different design philosophies in video generation. The selected models cover a broad scale spectrum, from approximately 3B parameters to 14B+ and beyond, enabling analysis of how reasoning capability correlates with model size. All included systems represent state-of-the-art performance as of late 2025 and early 2026, ensuring that the benchmark reflects the current frontier of the field. To ensure fair and controlled comparison, all generated videos follow the resolution of their corresponding ground-truth videos, which include both square and rectangular aspect ratios. For open-source models that support custom input resolutions, we specify the resolution to match the ground truth. For closed-source models with fixed resolution constraints (e.g., Sora 2 only supports 1280×720 or 720×1280), we pad the first-frame image with a background color and automatically detect and remove the padding during evaluation. Frame rates range from 15–24 fps depending on model requirements. For API-based models, we adopt the default or provider-recommended generation settings, while for open-source models we use the configurations reported in their original papers. This standardization removes confounding factors arising from resolution or tuning differences, ensuring that observed performance differences reflect reasoning ability rather than generation format. C.2. Details of Human Preference Analysis C.2.1. Data Preparation To mitigate potential biases in pairwise comparisons, we adopt two complementar scoring schemes: relative scoring based on pairwise preference judgments, and absolute scoring based on independent per-video ratings. In cases where the absolute scores clearly indicate that one video is superior but the corresponding relative annotation contradicts this assessment, the final decision is revised in favor of the absolute judgment. For each sample consisting of a text prompt and an initial image pairp i , we generate videos using nine 35 VIDEO-REASON.COMA Very Big Video Reasoning Suite video generation modelsM =M 1 ,M 2 ,M 3 ,M 4 ,M 5 ,M 6 ,M 7 ,M 8 ,M 9 , producing a set of outputsG i = V i,1 ,V i,2 ,V i,3 ,V i,4 ,V i,5 ,V i,6 ,V i,7 ,V i,8 ,V i,9 . For relative evaluation, all pairwise combinations are constructed within each group, resulting in 9 2 = 36 unique video pairs per sample. For absolute evaluation, the dataset contains 500 distinct prompt–image pairs p i . Each pair is processed by all nine models, yielding a total of 500 × 9 = 4,500 single-video annotation instances. Each video is presented together with its corresponding text prompt and task-specific evaluation documents. To reduce annotation noise, every video is independently annotated five times. All samples are randomly shuffled before assigned to annotators. C.2.2. Human annotation Annotators are between 20 and 35 years old and possess basic domain knowledge relevant to the tasks. All annotators undergo standardized training to ensure consistent interpretation of evaluation criteria. For relative scoring, annotators are shown two videos generated from the same prompt and are asked to select which one better completes the task and aligns with the given prompt overall, with ties allowed. For absolute scoring, annotators rate each video along three dimensions using a 5-point Likert scale, where higher scores indicate better performance. Task Completion (TC) assessing whether the task goal is achieved; Reasoning Logic (RL), assessing the correctness of the reasoning process; and Visual Quality (VQ), assessing visual clarity, temporal coherence, and rendering fidelity. C.2.3. Win Ratio Calculation From relative human annotations, we compute a win ratio for each model. In each comparison, the preferred model receives a score of 1 and the other 0; in the case of a tie, both receive 0.5. Win ratios are aggregated across all pairwise comparisons for each evaluation split. From absolute annotations, we average the scores of the TC, RL and VQ dimensions to obtain a final score per sample. These absolute scores are used for cross-verification of pairwise results and for resolving contradictory annotations between the two scoring schemes. Finally, we compare the win ratios derived from relative human annotations (with cross-verification using absolute scores) with the win ratios computed from VBVR-Bench ’s automatic evaluation metrics, and measure the correlation between the two across models. C.3. Detailed Analysis Protocols and Additional Results C.3.1. Residualized Capability Correlation This section details the computation behind Fig. 5 in the main paper. Our goal is to quantify capability dependency between cognitive categories while avoiding trivial correlations induced by overall model strength. Category scores.Letmindex models andcindex categories (five total). For each modelmand categoryc, we compute a category-level score by averaging the per-sample Overall ratings: S m,c = mean Overall over all evaluated samples with Category = c for model m. General factor. We define a model-level general factor as the overall mean score across all samples: G m = mean Overall over all evaluated samples for model m. Residualization. For each category c, we regress S m,c on G m across models: S m,c = a c + b c G m + ε m,c , and retain the residuals ε m,c as the strength beyond overall model quality. 36 VIDEO-REASON.COMA Very Big Video Reasoning Suite Capability dependency matrix.For each pair of categories(c 1 ,c 2 ), we compute Pearson correlation across models using residuals: R c 1 ,c 2 = corr (ε m,c 1 m ,ε m,c 2 m ). The resulting 5× 5 matrix R is visualized as a heatmap (main paper, Fig. 5). Implementation notes. We use the benchmark table and compute (i)S m,c , (i)G m , (i) residuals by ordinary least squares, and (iv) Pearson correlations on residuals. We also report Spearmanρin auxiliary analysis to confirm robustness to monotonic transformations. C.3.2. Domain-wise Score Distributions (Boxplots) To complement mean performance summaries, we report domain-wise score distributions across models using boxplots: task-level distributions, where each point corresponds to a task mean score within a domain (capturing cross-task variability). Figure 9Domain-wise score distributions across 9 models (red dashed line separates baselines and VBVR- Wan2.2). CogVHunyuan LTXWanKling Runway VeoSoraVBVR Model 1 2 3 4 5 Overall Score Abstraction (23 tasks) CogVHunyuan LTXWanKling Runway VeoSoraVBVR Model Knowledge (14 tasks) CogVHunyuan LTXWanKling Runway VeoSoraVBVR Model Perception (30 tasks) CogVHunyuan LTXWanKling Runway VeoSoraVBVR Model Spatiality (13 tasks) CogVHunyuan LTXWanKling Runway VeoSoraVBVR Model Transformation (20 tasks) Task-Level Boxplot: Model Performance by Domain (9 Models: Open-source | Closed-source | Ours) 37 VIDEO-REASON.COMA Very Big Video Reasoning Suite D. Selected Tasks and Rubrics This appendix includes a curated set of tasks that require multi-step planning and/or multiple interacting constraints. For each task, we show the initial and final frames, followed by the evaluation rubric used to score model outputs. Stable Sort (Task G-3) Rearrange objects by grouping them by shape type and sorting each group by size while preserving all attributes, testing multi-constraint spatial reasoning, attribute fidelity, and rule following. (a) First(b) 1st Progression(c) 2nd Progression(d) Final Figure 10 Stable Sort Example Prompt “The scene contains two types of shapes, each type has three shapes of different sizes arranged randomly. Keep all shapes unchanged in appearance (type, size, and color). Only rearrange their positions: first group the shapes by type, then within each group, sort the shapes from smallest to largest (left to right), and arrange all shapes in a single horizontal line from left to right.” Human Annotation Scoring (1–5): • 5 (Perfect). Correct grouping; correct within-group ascending size order (left to right); horizontal alignment; attributes preserved (type/size/color); reasonable spacing, no overlaps. •4 (Near-perfect). Correct grouping and ordering with one minor imperfection (e.g., small color deviation (hue shift within±10%), small size deviation (within±5%), slight vertical misalignment, or mildly uneven spacing) while the intended layout remains clear. • 3 (Partially correct). Grouping is correct but ordering is wrong in at least one group; or grouping and ordering are correct but attribute changes are noticeable (e.g., color shift > 10%, size change > 5%, or mild deformation while still recognizable). • 2 (Multiple errors). Incorrect grouping and/or incorrect ordering, and/or missing/extra objects. • 1 (Failure). Objective not achieved (no intended grouping/ordering) and/or severe object modification/loss. Evaluation dimensions (suggested weights): •Classification accuracy (30%). Correctly identify the two shape types; group identical shapes; include all 6 objects. • Ordering correctness (30%). Ascending size order within each group; coherent left-to-right organization. • Object fidelity (30%). Preserve shape type, size, and color; maintain clear contours/edges. • Layout accuracy (10%). Horizontal alignment and reasonable spacing. 38 VIDEO-REASON.COMA Very Big Video Reasoning Suite Grid Avoid Obstacles (Task G-15)Navigate a10× 10grid from start to goal using only 4-neighbor moves without entering obstacle cells, testing shortest-path planning under hard constraints. (a) First(b) 1st Progression(c) 2nd Progression(d) Final Figure 11 Grid Avoid Obstacles Example Prompt: “The scene shows a 10x10 grid with a blue start square (containing a yellow circular agent), a red end square, and multiple black X marks indicating obstacles. Starting from the blue start square, the agent can move to adjacent cells (up, down, left, right) each step. The goal is to move the agent to the red end square along the shortest path without entering any cells marked with black X obstacles.” Human Annotation Scoring (1–5): • 5 (Perfect). Reaches the goal; avoids all obstacles; uses a shortest (or tied-shortest) obstacle-avoiding path; movement is legal (4-neighbor, within-grid) and agent appearance is preserved. • 4 (Near-perfect). Reaches the goal and avoids all obstacles; only a minor imperfection (e.g.,≤ 2 extra steps vs. optimal, slight appearance drift, or minor legal-motion jitter). • 3 (Partially correct). Clear attempt but with a notable issue (e.g., enters exactly one obstacle cell once, > 2 extra steps, stops short of the goal, or an occasional illegal/diagonal tendency). • 2 (Mostly incorrect). Multiple violations (e.g., enters 2–3 obstacle cells) and/or out-of-grid motion and/or highly inefficient/random routing and/or fails to reach the goal. • 1 (Failure). No meaningful progress or unrelated output; frequent obstacle crossings (≥ 4) or the grid is corrupted. Evaluation dimensions (suggested weights): • Obstacle avoidance (40%). Zero obstacle entries; respects non-traversable constraint. • Path optimality (30%). Shortest path length under obstacle constraints. • Motion-rule compliance (20%). 4-neighbor steps, within-grid, continuous trajectory. • Task completion (10%). Reaches the goal with a coherent start→goal trajectory. 39 VIDEO-REASON.COMA Very Big Video Reasoning Suite Grid Go Through Block (Task G-16) Plan a near-shortest 4-neighbor route that visits all marked target cells (in blue) before reaching the goal (in red) in a10× 10grid, testing multi-goal route optimization under movement constraints. (a) First(b) 1st Progression(c) 2nd Progression(d) Final Figure 12 Grid Go Through Block Example Prompt :“The scene shows a 10x10 grid with a green start square (containing an orange circular agent), a red end square, and multiple blue rectangular blocks. Starting from the green start square, the agent can move to adjacent cells (up, down, left, right) each step. The goal is to move the agent to the red end square along the shortest path that passes through all blue blocks (the agent must visit every blue block before reaching the red end square).” Human Annotation Scoring (1–5): • 5 (Perfect). Starts at the start cell and ends at the goal; visits all targets; uses a globally shortest route under the visit-all constraint; motion is legal (4-neighbor, within-grid) and agent appearance is preserved. •4 (Near-perfect). Reaches the goal and visits all targets; only small imperfections (e.g.,≤ 3extra steps, minor appearance drift, or slight presentation issues while the realized path remains grid-orthogonal). • 3 (Partially correct). Clear attempt but with a notable issue (e.g., misses exactly one target, > 3 extra steps, stops before the goal, or an occasional illegal move). • 2 (Mostly incorrect). Misses two or more targets and/or fails to reach the goal, and/or frequent illegal/out-of-grid motion, and/or highly inefficient/random routing. • 1 (Failure). No meaningful progress; ignores targets or output is unrelated / grid is broken. Evaluation dimensions (suggested weights): • Target coverage (40%). All targets visited; none missing; targets remain visible. •Route optimality (30%). Near-global shortest route under the visit-all constraint (target order matters). • Task completion (20%). Full start→(all targets)→end sequence completed. • Motion legality (10%). 4-neighbor steps only; stays within grid; coherent step-by-step movement. 40 VIDEO-REASON.COMA Very Big Video Reasoning Suite Directed Graph Navigation (Task G-31) Navigate from the start node to the goal node by traversing only along directed edges (respecting arrow direction) with a minimum-hop path, testing goal-directed graph planning under directionality constraints. (a) First(b) 1st Progression(c) 2nd Progression(d) Final Figure 13 Directed Graph Navigation Prompt :“The scene shows a network of nodes connected by directed edges (edges with arrows indicating direction) with a green starting node, a red ending node, and a blue triangular agent positioned at the green starting node. The agent can only move along edges in the direction they point (from the source node to the target node, cannot move backwards), moving from one node to an adjacent node each step. Move the blue triangular agent from the green starting node to the red ending node along the path with the minimum number of steps.” Human Annotation Scoring (1–5): • 5 (Perfect). Reaches the goal; uses a minimum-hop path; strictly follows arrow directions and moves only along existing edges; motion is continuous and the graph is preserved. • 4 (Near-perfect). Reaches the goal with correct direction/edge-following behavior; only minor presentation issues while path length remains shortest. • 3 (Partially correct). Reaches the goal but with a notable issue (e.g., +1–2 extra hops, a single direction violation, or slight deviation from drawn edges). • 2 (Mostly incorrect). Severe issues: ≥ 3 extra hops, multiple direction violations, “flying”/cutting across space, reaches the wrong node, or alters graph structure. • 1 (Failure). Does not reach the goal, behavior is random/invalid (ignores arrows), or the graph is corrupted. Evaluation dimensions (suggested weights): • Shortest-path correctness (40%). Minimum-hop path length. • Direction compliance (35%). No reverse traversals; respects arrow orientation. • Motion legality (15%). Moves along existing edges, one edge per step; no teleportation. • Graph fidelity (10%). Nodes/edges/arrows remain unchanged. 41 VIDEO-REASON.COMA Very Big Video Reasoning Suite Key Door Matching (Task G-45) In a maze, first collect the prompt-specified colored key and then reach the matching door via legal corridor moves near shortest, testing instruction grounding, sequencing, and constrained navigation. (a) First(b) 1st Progression(c) 2nd Progression(d) Final Figure 14 Key Door Matching Prompt :“The scene shows a maze with a green circular agent, colored diamond-shaped keys, and colored hollow rectangular doors. Find the Blue key and then navigate to the matching Blue door, showing the complete movement process step by step.” Human Annotation Scoring (1–5): • 5 (Perfect). Reaches the correct target-colored key, then reaches the matching door (correct order); path is legal and continuous; path length is near-optimal (e.g.,≤ 120%of BFS shortest for start→key→door); key disappears when reached; full trajectory is shown. • 4 (Near-perfect). Correct key and door in correct order; minor issues only (e.g., slightly longer path ∼ 120%–150%, small jitter/backtracking, or a 1–2 frame delay in key disappearance). • 3 (Partially correct). Finds the correct key but makes a major mistake (e.g., goes to wrong door, completes only one stage, noticeably inefficient path∼ 150%–200%, or 1–2 minor wall violations). • 2 (Mostly incorrect). Wrong key color and/or wrong order (door before key), and/or multiple wall crossings, and/or extremely inefficient path (> 200%), or stops at irrelevant locations. • 1 (Failure). Agent barely moves or moves randomly; ignores maze constraints; incorrect key/door colors; or agent behavior is abnormal. Evaluation dimensions (suggested weights): • Target identification (30%). Correctly identify the prompt-specified key and matching door; do not confuse with other colors. • Path validity (30%). Follow allowed corridors only, avoid wall collisions, and maintain step-by-step movement. • Path efficiency (20%). Actual path length relative to the BFS-optimal path;≤ 110% = excellent, 110–130% = acceptable, > 200% = poor. • Animation quality (20%). Smooth frame-by-frame movement; agent centers align with key and door; key pickup effect visible. 42