Paper deep dive
OpenCoF: Learning to Reason Through Video Generation
Xinyan Chen, Ziyu Guo, Renrui Zhang, Dongzhi Jiang, Hongsheng Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 7/18/2026, 8:21:13 AM
Summary
The paper introduces OpenCoF, a framework for enhancing video reasoning through Chain-of-Frame (CoF) reasoning. It presents the OpenCoF-17K dataset, comprising 17,312 videos across 11 task families, and the Wan-CoF model, a fine-tuned version of Wan2.2-I2V-A14B. The study explores visual and textual reasoning tokens to improve spatial and temporal reasoning capabilities, demonstrating significant gains over baselines on video reasoning benchmarks.
Entities (14)
Relation Signals (13)
OpenCoF → contains → OpenCoF-17K
confidence 95% · we introduce OpenCoF, a framework comprising the OpenCoF-17K dataset
Wan-CoF → isfinetunedfrom → Wan2.2-I2V-A14B
confidence 95% · we fine-tune it on OpenCoF-17K... yielding Wan-CoF
OpenCoF-17K → usedfortraining → Wan-CoF
confidence 95% · we fine-tune it on OpenCoF-17K... yielding Wan-CoF
OpenCoF-17K → coverstasks → Chess
confidence 90% · Chess... This task aims to bolster the reasoning proficiency
OpenCoF-17K → coverstasks → Sudoku
confidence 90% · Sudoku... To inject algorithmic and computational reasoning supervision
OpenCoF-17K → coverstasks → Physics Motion
confidence 90% · Physics motion... Physics-based motion reasoning serves as a capability
Wan-CoF → evaluatedon → MME-CoF
confidence 90% · Across four video reasoning benchmarks... MME-CoF
Wan-CoF → evaluatedon → VIPER
confidence 90% · Across four video reasoning benchmarks... VIPER
Wan-CoF → evaluatedon →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Reasoning has become a core capability for large models, especially when reliable decisions require understanding logical consequences. Recent video generation models offer a reasoning path distinct from previous Chain-of-Thought (CoT): reasoning can unfold through temporally connected frames, known as Chain-of-Frame (CoF) reasoning. However, existing video generators are primarily trained on general video corpora, still lacking diverse supervision and dedicated designs for CoF reasoning. To address this gap, we introduce OpenCoF, a framework comprising the OpenCoF-17K dataset, a reasoning video dataset spanning 11 task families, and Wan-CoF, a fine-tuned video model for studying whether diverse temporal supervision improves CoF behavior. Across four video reasoning benchmarks, Wan-CoF achieves considerable gains over the Wan2.2-I2V-A14B baseline. Building on this, we empirically explore more advanced designs for CoF capabilities, i.e., equipping the model with visual and textual reasoning tokens. This mechanism respectively captures low-level visual cues and high-level semantic priors for spatial and temporal reasoning. Through performance comparisons and attention analysis, we examine how these tokens contribute across model depth, denoising steps, space, and time. Our results suggest that stronger video reasoning requires both broad temporal supervision and explicit mechanisms for organizing intermediate reasoning state. We open-source the dataset, model, and code to facilitate future research on reasoning-oriented video generation.
Tags
Links
- Source: https://arxiv.org/abs/2607.08763v1
- Canonical: https://arxiv.org/abs/2607.08763v1
Trouble viewing inline? Open PDF directly →
Full Text
59,463 characters extracted from source content.
Expand or collapse full text
OpenCoF: Learning to Reason Through Video Generation Xinyan Chen 1,2,∗ , Ziyu Guo 3,∗ , Renrui Zhang 1,† , Dongzhi Jiang 1,2 , Hongsheng Li 2 1 ByteDance Seed, 2 CUHK MMLab, 3 CUHK IMIXR ∗ Equal contribution, † Corresponding author Abstract Reasoning has become a core capability for large models, especially when reliable decisions require understanding logical consequences. Recent video generation models offer a reasoning path distinct from previous Chain-of-Thought (CoT): reasoning can unfold through temporally connected frames, known as Chain-of-Frame (CoF) reasoning. However, existing video generators are primarily trained on general video corpora, still lacking diverse supervision and dedicated designs for CoF reasoning. To address this gap, we introduce OpenCoF, a framework comprising the OpenCoF-17K dataset, a reasoning video dataset spanning 11 task families, and Wan-CoF, a fine-tuned video model for studying whether diverse temporal supervision improves CoF behavior. Across four video reasoning benchmarks, Wan-CoF achieves considerable gains over the Wan2.2-I2V-A14B baseline. Building on this, we empirically explore more advanced designs for CoF capabilities, i.e., equipping the model with visual and textual reasoning tokens. This mechanism respectively captures low-level visual cues and high-level semantic priors for spatial and temporal reasoning. Through performance comparisons and attention analysis, we examine how these tokens contribute across model depth, denoising steps, space, and time. Our results suggest that stronger video reasoning requires both broad temporal supervision and explicit mechanisms for organizing intermediate reasoning state. We open-source the dataset, model, and code to facilitate future research on reasoning-oriented video generation. Date: July 10, 2026 Correspondence: Renrui Zhang at renruizhang@bytedance.com Project Page: https://opencof.github.io/ 1 Introduction Enhancing reasoning ability [16,24,48,49] has emerged as a central objective in the development of large language (LLM) and multimodal (LMM) models. Prior works demonstrate that robust reasoning capabilities are indispensable for reliable decision-making in vision-language contexts. However, mainstream visual Chain-of-Thought (CoT) [17,27,43] pipelines remain largely anchored to static visual observations. These approaches typically rely on extracting localized evidence [4,34], invoking external tools [11,13,15,50], or employing auxiliary image-generation steps [20,46], which limits their ability to intrinsically model dynamic transitions and multi-step visual consequences. The rapid advancement of video generation models offers a promising alternative: reasoning via video [10, 23,30,44]. Rather than relying solely on textual steps or static visual content, models can reason through 1 arXiv:2607.08763v1 [cs.CV] 9 Jul 2026 Chess Sudoku Sudoku Physics Motion Cubefolding Tangram Puzzle Dot-to-dot 2D Geometry Embodied Manipulation Maze VBVR Instance-based Rendering Expert-guided Rendering Procedural Scene Synthesis External Video Repurposing 4CurationPipelines,11Tasks OpenCoF-17K Figure 1 Overview of OpenCoF. We construct the OpenCoF-17K dataset, comprising 17,312 videos across 11 tasks, via four complementary pipelines, to provide diverse temporal supervision for Chain-of-Frame (CoF) reasoning. the temporal evolution of frames, a process formalized as Chain-of-Frame (CoF) reasoning. Recent studies suggest that video models already exhibit nontrivial world knowledge and early signs of physical and causal understanding, indicating significant potential for CoF [44]. Nevertheless, this potential remains far from fully realized. In reasoning-intensive tasks, current video models still struggle with long-term temporal coherence, physical and spatial consistency, and logical continuity [10]. Ultimately, high visual realism does not guarantee reliable reasoning. Starting from MME-CoF [10], recent works have predominantly focused on establishing benchmarks to delineate and evaluate the reasoning scope of video generation models, including RULER-Bench [12], Gen- ViRe [23], VIPER [22], and others [25,38]. While these efforts are instrumental in defining the problem space, research dedicated to actively enhancing these reasoning capabilities remains noticeably sparse. A few preliminary studies have explored tailored training and test-time scaling strategies, though they often concentrate on specific task domains, such as maze solving in VR-Bench [47] and Thinking in Frames [21], or custom cognitive-task suites like VBVR [42]. Despite these initial advancements, current approaches exhibit two limitations: (1) Their training and evaluation protocols are localized to specific domains or intrinsically coupled to their own curated suites, leaving true generalizability to independent, externally curated reasoning benchmarks underexplored; and (2) their primary emphasis has been on data scaling and inference strategies, leaving internal video model architectures explicitly tailored for reasoning largely uninvestigated. Consequently, there remains a fundamental gap in understanding how to structurally enhance video reasoning capabilities. To bridge this gap, we introduce OpenCoF, a framework dedicated to advancing video reasoning. As the foundation of this project, we construct the OpenCoF-17K dataset, a diverse collection comprising 17,312 videos across 11 task families. Its curation pipeline integrates four distinct approaches: instance-based rendering, expert-guided rendering, procedural scene synthesis, and the repurposing of existing videos, as illustrated in Figure 1. Leveraging this dataset, we investigate the enhancement of video reasoning capabilities within a robust open-source video generation model, Wan2.2-I2V-A14B [41]. First, we fine-tune it on OpenCoF-17K without incorporating any reasoning-specific techniques, yielding Wan-CoF. This data- centric phase aims to verify whether diverse temporal supervision alone can effectively elicit CoF behavior. 2 OpenCoF-17K 4CurationPipelines,11Tasks Exploration of CoFTechniques Wan2.2 Baseline Wan-CoF Wan-CoFvt Wan-CoFtt OpenCoF-17K Wan2.2Baseline Wan-CoF Wan-CoFvtWan-CoFtt ExperimentalResults Textualreasoningtokens ... ... Visual reasoningtokens Exploration of CoFTechniques Wan2.2 Baseline Wan-CoF Wan-CoFvt Wan-CoFtt OpenCoF-17K Wan2.2Baseline Wan-CoF Wan-CoFvtWan-CoFtt ExperimentalResults Textual reasoning tokens ... ... Visual reasoning tokens Figure 2 Exploration of CoF techniques. Fine-tuning Wan2.2-I2V-A14B on OpenCoF-17K yields Wan-CoF, which achieves substantial gains over the Wan2.2 baseline across four external video-reasoning benchmarks (MME-CoF, Gen-ViRe, VIPER, RULER-Bench). We further explore Visual Reasoning Tokens (vt) and Textual Reasoning Tokens (t) as complementary reasoning-token designs, yielding Wan-CoF vt and Wan-CoF t , which improve further while exhibiting different strengths across benchmarks. The subsequent evaluations across four external video-reasoning benchmarks demonstrate substantial gains over the base model. Beyond this data-centric stage, we further explore how dedicated reasoning techniques can help a video generator organize the intermediate reasoning state required by CoF. In current DiT-based video generators, such reasoning remains implicit in generation-oriented visual latents and text conditions. We therefore explore two complementary reasoning-token designs that operate at different levels of the model: Visual Reasoning Tokens (vt), inserted into the visual latent sequence to capture low-level visual cues, and Textual Reasoning Tokens (t), appended to the text-conditioning sequence to provide high-level semantic priors. These two variants further improve Wan-CoF on external benchmarks while exhibiting different strengths across task dimensions, as shown in Figure 2. Through empirical performance evaluations and attention analysis, we examine how these token types facilitate reasoning across model depth, denoising steps, spatial dimensions, and time. Our contributions are summarized as follows: • We introduce the OpenCoF-17K dataset, comprising 17K videos across 11 task families with a scalable four-source curation pipeline. •We develop Wan-CoF by fine-tuning Wan2.2-I2V-A14B on OpenCoF-17K, demonstrating substantial gains over the base model across four external video-reasoning benchmarks. •We propose complementary reasoning-token mechanisms (vt and t) and study their role in CoF reasoning through performance and attention analyses. 2 OpenCoF-17K Current video generators lack dedicated temporal supervision for reasoning tasks, making diverse reasoning- centric data essential for robust CoF capabilities. To meet this need, we construct OpenCoF-17K, a video reasoning dataset spanning 11 task families. We first summarize its scale, format, and statistics in Section 2.1, then detail the four-part curation pipeline in Section 2.2. 2.1 Dataset Statistics The OpenCoF-17K dataset comprises 17,312 samples across 11 task families. Formulated for conditional generation, each data instance pairs an initial conditioning image and a textual prompt with a target reasoning video. To ensure optimization consistency across diverse data sources, all videos are standardized to a resolution of 480p, a frame rate of 15 fps, and a temporal length of 81 frames. Table 1 reports the per-task statistics of OpenCoF-17K. In total the dataset comprises 17,312 videos across 11 task families, with an average prompt length of 53.2 words. The VBVR family aggregates 30 subtasks: 29 of them contribute 250 demonstration videos each and the remaining one contributes 500, summing to 7,750 videos. Figures 5 and 6 3 Figure 3 Composition of the OpenCoF-17K dataset. Chess 1,000 (5.8%) Sudoku 1,000 (5.8%) 2D geometry 1,162 (6.7%) Dot-to-dot 1,000 (5.8%) Tangram 800 (4.6%) Cube folding 800 (4.6%) 3D polycube 800 (4.6%) Physics motion 1,000 (5.8%) Maze 1,000 (5.8%) Embodied manip. 1,000 (5.8%) VBVR 30 subtasks 7,750 (44.8%) Table 1 Per-task statistics of OpenCoF-17K.Avg prompt denotes the mean prompt length in words. TaskNumAvg prompt Chess1,00042.0 Sudoku1,00052.0 2D geometry1,16223.4 Dot-to-dot1,00031.0 Tangram puzzle80013.5 Cube folding80017.7 3D polycube rotation80025.3 Physics motion1,00045.9 Maze1,000179.4 Embodied manipulation1,00010.2 VBVR (30 subtasks)250 × 29 + 50062.9 Total17,31253.2 show representative cases for each of the 11 task families in the OpenCoF-17K dataset. For every case we display three uniformly sampled frames from the demonstration video together with the original text prompt. We next describe how these data are collected and rendered into a unified video-reasoning format. 2.2 Dataset Curation To obtain the coverage summarized above, our curation protocol comprises four pipelines: (1) Instance-based rendering, (2) Expert-guided rendering, (3) Procedural scene synthesis, and (4) External video repurposing. As illustrated in Figure 4, the first three pipelines focus on generative construction. To complement these efforts, the fourth pipeline strategically repurposes high-quality demonstrations from existing external datasets. Collectively, these approaches integrate multifaceted data into a unified video-reasoning format, establishing a scalable foundation for future expansion. In contrast to prior efforts like VBVR and VR-Bench that predomi- nantly rely on procedural scene synthesis, our dataset curation pipeline combines multiple complementary construction strategies. 2.2.1 Instance-based Rendering Within this pipeline, we synthesize video data from pre-existing structured assets, such as puzzle states or board configurations. Individual instances are sampled from these assets and sequentially translated into video formats via deterministic code rendering. Chess. This task aims to bolster the reasoning proficiency of video models within highly structured, rule-based environments. We sample instances from the Mate in One (Chess) 1 dataset and procedurally render the corresponding video sequences. The models are tasked with identifying and visually executing the singular legal move that results in a checkmate. Sudoku. To inject algorithmic and computational reasoning supervision, we leverage the training split of the Sudoku-Easy 2 dataset, augmenting it with procedurally generated puzzles of equivalent difficulty. Here, video models are required to resolve 4× 4 Sudoku puzzles through a rigorous, step-by-step deductive process. 2.2.2 Expert-guided Rendering Under this protocol, we leverage human expertise, LLM-generated concepts, or outputs from text-to-image (T2I) models to establish structures and annotations. Guided by these representations, we extract or render intermediate structures and convert them into sequential videos with code. 1 https://w.kaggle.com/datasets/ancientaxe/mate-in-one-chess 2 https://huggingface.co/datasets/LangAGI-Lab/Sudoku-Easy 4 Instance-based rendering Expert-guided rendering Procedural scene synthesis Code Rendering +-------+-------+ | 4 . | . . | | . 3 | . . | +-------+-------+ | . 1 | 3 . | | . . | . . | +-------+-------+ Categories: Animals, Buildings, Food, Sports .... Code Rendering Chess Sudoku 2D geometry Dot-to-dot Human Experts Physics engines or Graphics pipelines Random Seed Code Rendering 2D geometry Dot-to-dot Chess Sudoku Tangram puzzle Physics motion 3D polycuberotation Cube folding a b c d e f g h 8 . k . . . . . . 7 . . . . . R . . 6 . . . R . . . . 5 . . . . . . . . 4 . . . . . . . . 3 . . . . . . . . 2 . . . . . . . . 1 . . . . . K . . Coordinates: A:[x_1, y_1], D:[x_2,y_2] LLM Expert T2I Model Expert Concepts: Animals: Elephant, Whale ... Buildings: House, Castle ... Food: Apple, Banana ... Sports: Cricket Bat, Oar ... ... Edge detection Procedural Scene Synthesis 2D 3D Figure 4 Overview of the first three curation pipelines in OpenCoF-17K. Instance-based rendering converts structured assets into videos through code rendering. Expert-guided rendering uses human, LLM, or T2I-model guidance to define structures before rendering. Procedural scene synthesis uses physics engines or graphics pipelines to instantiate scenes. 2D geometry. Targeting geometric reasoning, we sample 2D geometry images from Geo170K [7] that require auxiliary-line construction. Expert annotators define ground-truth coordinates for these segments, after which we render videos showing the step-by-step drawing process. Dot-to-dot. The dot-to-dot task can strengthen the model’s understanding of abstract graphics and its ability to follow strict sequential rules. We predefine various categories and prompt LLMs [6] to generate specific concepts. Subsequently, we synthesize minimalist line-drawing images using T2I models [35], which are then converted into sequential dot-to-dot animation videos via classical computer vision routines. 2.2.3 Procedural Scene Synthesis This pipeline synthesizes video data entirely from scratch. We utilize random seeds to instantiate foundational scene parameters, while physics engines or graphics pipelines subsequently render the unfolding dynamics into continuous video streams. Tangram puzzle. Designed to cultivate spatial awareness and shape manipulation capabilities, this task requires models to systematically rearrange tangram pieces into a target configuration. We procedurally generate puzzle pieces with randomized shapes and initial placements, directly rendering the sequential assembly process into visual formats. Cube folding. Targeting 2D-to-3D spatial reasoning, we procedurally generate cube-folding sequences. Given a flat 2D cube net, the models must predict the correct folding sequence to form a complete 3D cube. The initial nets are randomized across varying topologies, viewing angles, and surface patterns, with the folding dynamics programmatically animated using Matplotlib 3D. 3D polycube rotation. To inject geometric motion control supervision across diverse perspectives, we formulate a 3D polycube rotation task. Here, polycubes are instantiated with randomized complex geometries and 5 Chess Please play as White or Black (determine your side by looking at the board and piece positions) and deliver checkmate on your next move. Maintain a fixed camera angle and lighting throughout the game. Static scene, no zoom, no pan, no dolly. Sudoku Create a static, smooth, animation that solves the given 4x4 sudoku. Enter the missing numbers one by one. Do not change anything else in the picture. Only fill the numbers in the empty cells so the sudoku is solved properly. A cursor moves and fills the correct number in the empty boxes. 2D Geometry Smoothly connecting point B and point D with a line. The video ends once the connection process is complete. Static view, no zoom or pan. Dot-to-dot Animate the dots connecting sequentially from 1 to 12, each straight line appearing smoothly until the full outline emerges. Keep the background with the smiling sun and plants unchanged. Static shot. Tangram Puzzle Place piece A at (x, y) = (2, 2). Cube Folding Fold the net upward into a cube, keeping the camera fixed. 3D Polycube Rotation A structure composed of 10 cubes viewed from 318° azimuth, 32° elevation. The camera executes a smooth horizontal orbit around the structure, transitioning to a 95° azimuth which is a 137-degree rotation while maintaining a constant distance from the object. Physics Motion A ball is placed at the top of an inclined plane (inclination angle θ = 21.2 deg). The surface has a friction coefficient μ = 0.15. The ball is released and rolls down the slope onto the horizontal ground. Show the complete motion. Static shot. Maze Create a 2D animation based on the provided image of a grid puzzle. The purple wizard moves into position behind the purple wooden crate and smoothly pushes it toward the blue magic circle. The purple wooden crate only slides when pushed from behind by the purple wizard and moves in a straight line along the gray stone bricks tiles. ... Embodied Manipulation fold the cloth from right to left VBVR A system of 3 communicating vessels with equal-diameter vertical tubes is filled with water (water-like (low viscosity)), which appears blue in color. As shown in the initial frame, the liquid levels in the tubes are [14, 36, 36] cm respectively. Due to pressure differences between the tubes, the liquid begins to flow through the connecting channels at the bottom. The flow is governed by hydrostatic pressure equalization and damped by viscous resistance with coefficient k=3.92. As the liquid redistributes, the height differences gradually decrease, and the system evolves toward equilibrium. Eventually, through conservation of volume, all tubes reach the same final liquid level, which equals the average of the initial heights. Simulate this settling process from the initial unbalanced state to the final stable equilibrium. Figure 5 Representative examples from OpenCoF-17K. 6 Chess Please play as White or Black (determine your side by looking at the board and piece positions) and deliver checkmate on your next move. Maintain a fixed camera angle and lighting throughout the game. Static scene, no zoom, no pan, no dolly. Sudoku Create a static, smooth, animation that solves the given 4x4 sudoku. Enter the missing numbers one by one. Do not change anything else in the picture. Only fill the numbers in the empty cells so the sudoku is solved properly. A cursor moves and fills the correct number in the empty boxes. 2D Geometry Smoothly connecting point O and point E with a line. The video ends once the connection process is complete. Static view, no zoom or pan. Dot-to-dot Animate the dots connecting sequentially from 1 to 9, each straight line appearing smoothly until the full outline emerges. Keep the background with the smiling sun and plants unchanged. Static shot. Tangram Puzzle Rotate piece A by 60 degrees clockwise, and place it at (x, y) = (2, 0). Cube Folding The net is folded to form a single cube, with folding edges clearly shown. Static camera perspective, no zoom or pan. 3D Polycube Rotation Rotating around the 3D shape to suggest spatial structure, keeping lighting and geometry fixed. Physics Motion A block is released from the top edge of a U-shaped semicircular track (radius R = 174). The surface has a friction coefficient μ = 0.27. The block slides down along the curved surface, passes through the lowest point, and climbs up the opposite side. Show the complete motion. Static shot. Maze Create a 2D animation based on the provided image of a maze. The blue circle slides smoothly along the white square path, stopping perfectly on the green circle. The blue circle never slides into or crosses the red x (trap areas). ... Embodied Manipulation move the fork above cloth VBVR A black ball starts with the shown initial direction arrow. It moves in straight lines and reflects off the two vertical walls following the law of reflection (the angle of reflection equals the angle of incidence). Follow this unique trajectory until the ball hits one of the red hollow circle targets in the lower-right area. Identify which target circle is hit and show the full trajectory step by step. Figure 6 Representative examples from OpenCoF-17K. 7 Table 2 Results on MME-CoF [10]. The scores are reported on a 0–4 scale. ModelOverall Instruction TemporalVisual ContentFocus Alignment Consistency Stability Fidelity Relevance Closed-source Models Kling-v1 [19] 0.710.090.212.460.300.48 Seedance-1.0-Pro [8]1.480.391.712.051.222.04 Veo-3.0-Fast [9] 1.520.731.411.831.252.36 Veo-3.0-Preview [9]1.530.701.461.911.242.36 Sora-2 [28] 1.800.741.602.361.752.53 Open-source Models Wan2.2-TI2V-5B [41]0.720.080.961.160.480.90 HunyuanVideo-I2V [18] 0.980.231.381.520.681.08 Data-only Fine-tuning Wan2.2-I2V-A14B (Baseline) [41] 1.000.220.961.520.681.62 Wan-CoF1.300.311.291.991.001.89 ∆ over the Baseline +0.30+0.10+0.33+0.47+0.32+0.28 Exploration of Reasoning Techniques Wan-CoF vt 1.340.261.382.330.991.76 Wan-CoF t 1.350.341.272.181.171.77 viewpoints, and are subsequently rotated across varying axes and angles. Physics motion. Physics-based motion reasoning serves as a capability for grounding video models in physical reality. We simulate scenarios featuring balls and blocks traversing flat or curved tracks under varying friction coefficients (e.g., smooth versus rough surfaces). Models are tasked with forecasting the plausible dynamic trajectories of these objects strictly from their initial states. 2.2.4 External Video Repurposing Complementing the three active-construction pipelines visualized in Figure 4, we strategically leverage high- quality demonstrations from existing video datasets. These resources are repurposed by standardizing their spatial resolution, frame rate, and temporal length to seamlessly align with our unified training suite. Maze. We sample maze navigation data from the training split of the VR-Bench [47] dataset. Since some instances in this dataset follow similar distributions, we sample a subset while preserving coverage across subtasks and difficulty levels. VBVR. Because some VBVR subtasks are overly simple or target similar capabilities, we select 30 representative subtasks from the broader VBVR [42] dataset and randomly sample demonstration videos from each. Embodied manipulation. The data for the embodied manipulation task are sampled from BridgeData V2 [40]. This task requires video models to comprehend spatial relationships and perform controlled manipulations. 3 Training of Wan-CoF Before introducing explicit reasoning techniques in the models, we first conduct the most direct study: how far can a strong open video generator go when it is trained only with diverse CoF supervision? To answer this, we use Wan2.2-I2V-A14B [41], a general-domain pretrained I2V model, and directly fine-tune it on OpenCoF-17K with LoRA [14]. The resulting model, Wan-CoF, marks the data-only stage of our study and serves as a clean reference point for attributing performance gains to the dataset itself. We evaluate Wan-CoF on four video reasoning benchmarks, which allows us to separate the effect of OpenCoF-17K from the reasoning-token designs explored later in Section 4. Please refer to Section A for experimental details. 8 Table 3 Results on Gen-ViRe [23]. Following the Gen-ViRe protocol, the scores are on a 0–1 scale. ModelOverallAbstract Algo. & Logi. Analogy Perceptual Planning Spatio-Temporal Closed-source Models Kling-v1 [19]0.1980.0710.0570.1170.1400.4430.359 Seedance-1.0-Pro [8]0.3010.1540.1640.0830.1710.6090.621 Veo-3.1 [9]0.4860.4400.4510.3670.3860.7220.550 Sora-2 [28]0.5600.6040.4720.4830.4960.7680.537 Data-only Fine-tuning Wan2.2-I2V-A14B (Baseline) [41]0.3040.2200.3590.0830.1900.5190.452 Wan-CoF 0.3910.3280.5060.1330.3470.5780.451 ∆ over the Baseline+0.087+0.107+0.147+0.050+0.157+0.059-0.001 Exploration of Reasoning Techniques Wan-CoF vt 0.4410.4750.5070.2330.2980.6280.507 Wan-CoF t 0.4060.4240.4740.1670.3060.5650.499 Table 4 Results on VIPER [22]. We follow the original VIPER protocol and report POC@1.0. ModelOverallTemporal Structural Symbolic Spatial Physics Planning Closed-source Models Seedance-1.5 [32]9.55.65.60.05.36.533.8 Veo-3.1 [9]20.322.220.013.310.714.741.0 Sora-2 [28]23.310.442.558.34.09.415.1 Open-source Models Hunyuan-1.5 [45]8.18.45.00.02.711.121.1 Data-only Fine-tuning Wan2.2-I2V-A14B (Baseline) [41] 3.35.66.20.00.02.85.3 Wan-CoF7.518.810.00.05.38.35.3 ∆ over the Baseline+4.2+13.2+3.80.0+5.3+5.50.0 Exploration of Reasoning Techniques Wan-CoF vt 8.212.811.30.04.08.315.8 Wan-CoF t 8.814.815.00.08.02.810.5 3.1 Main Results Comparison with the baseline. As shown in the Data-only Fine-tuning rows of Tables 2 to 5, relative to our baseline, Wan2.2-I2V-A14B, Wan-CoF lifts the headline metric on every benchmark: MME-CoF Overall from 1.00 to 1.30 (+0.30), Gen-ViRe average from 0.304 to 0.391 (+0.087), VIPER POC from 3.3 to 7.5 (+4.2), and RULER-Bench overall average from 55.8 to 56.8 (+1.0). The gains are broad-based—every MME-CoF aspect improves, and on Gen-ViRe and VIPER almost every category also gains. The severe non-positive cells sit within evaluation noise or correspond to dimensions where open-source models still score near zero. Moreover, the largest jumps cluster on reasoning-oriented sub-dimensions like Temporal Consistency, Physics, Spatial, Perceptual, and Algorithmic & Logical, rather than on visual fidelity alone, suggesting that our training strengthens frame-level reasoning rather than acting as a visual-quality regularizer. Comparison with other models. In the Data-only Fine-tuning rows of Tables 2 to 5, Wan-CoF surpasses open-source competitors and narrows the gap to the closed-source frontier. On MME-CoF, Wan-CoF exceeds HunyuanVideo-I2V and Wan2.2-TI2V-5B, approaching Seedance-1.0-Pro. On Gen-ViRe, Wan-CoF reaches 0.391 average, outperforming Kling-v1 and Seedance-1.0-Pro and moving closer to Veo-3.1. On VIPER, Wan-CoF attains POC of 7.5, on par with open-source Hunyuan-1.5 and closed-source Seedance-1.5. Under a LoRA fine-tuning budget on a single open backbone, this places Wan-CoF among the strongest open-source I2V models on CoF reasoning. 9 Table 5 Results on RULER-Bench [12]. We report the filtered image-to-video setting on a 0–100 scale, since our baseline does not support text-to-video generation. IF: Instruction Following, VC: Visual Consistency, VF: Visual Fidelity, RC: Rule Coherence. ModelOverall VisionScienceGameHumanity IF VC VFRC Avg.IFVC VF RC Avg.IFVCVFRCAvg.IFVC VFRC Avg. Data-only Fine-tuning Wan2.2-I2V-A14B (Baseline) [41] 55.80.0 69.1 65.8 28.0 54.344.2 73.5 83.6 18.4 54.921.6 65.6 83.4 12.0 45.661.1 91.7 91.7 40.2 71.2 Wan-CoF56.80.0 65.1 67.0 31.7 54.645.6 70.8 83.1 18.9 54.625.0 74.4 87.0 23.9 52.647.2 86.1 94.4 49.4 69.3 ∆ over the Baseline +1.00.0 -4.0 +1.3 +3.6 +0.3+1.4 -2.7 -0.5 +0.5 -0.3+3.4 +8.8 +3.6 +11.9 +6.9-13.9 -5.6 +2.8 +9.3 -1.9 Exploration of Reasoning Techniques Wan-CoF vt 59.60.0 72.7 71.0 35.6 59.848.5 68.5 81.4 15.5 53.530.7 73.4 86.4 22.1 53.258.3 88.9 98.2 56.9 75.6 Wan-CoF t 57.70.0 71.2 68.1 34.2 57.845.9 69.8 84.6 18.9 54.828.6 76.5 87.2 23.1 53.852.8 86.1 94.9 49.4 70.8 3.2 Out-of-distribution Analysis A key challenge in evaluating CoF reasoning is ensuring that models generalize beyond their training distributions. While recent works [42,47] evaluate distribution shifts internally within their respective task suites, we explore generalization by transferring our model directly to independent, externally curated benchmarks. Because these external benchmarks encompass distributions distinct from our training set, testing on them measures performance under a distribution shift. Under this setting, Wan-CoF demonstrates consistent improvements across all four benchmarks. Notably, the per-category gains appear to align with the types of supervision provided by OpenCoF-17K: training on Physics motion transfers to VIPER Physics (+5.5); structured-grid tasks (e.g., Chess, Sudoku, Maze) correlate with Gen-ViRe Algorithmic & Logical (+0.147) and the RULER-Bench Game split (+6.9); and spatial-shape tasks map to Gen-ViRe Perceptual (+0.157) and VIPER Spatial (+5.3). Together, these transfer signals suggest that OpenCoF-17K helps foster transferable CoF reasoning skills across suite boundaries, rather than overfitting to specific training domains. 4 Exploration of CoF Techniques The data-only experiment in Section 3 establishes an encouraging starting point: training on OpenCoF-17K transfers to external reasoning benchmarks even when reasoning techniques are not explicitly added. This result also sharpens the empirical question: if CoF gains can emerge from supervision alone, can dedicated reasoning techniques give the model a more explicit place to organize the intermediate rules, plans, and abstract goals behind those gains? The underlying generator still lacks a dedicated channel for maintaining reasoning state, and must express such information implicitly through pixel-level latents. Since existing video generation models are not originally designed for stepwise reasoning, they offer no specialized structures to host either the low-level reasoning cues operating inside the visual latent space or the high-level reasoning semantics operating at the prompt level. We therefore explore two complementary designs to target these distinct layers: Visual Reasoning Tokens (vt), which reside within the latent space of the DiT to capture low-level reasoning cues, and Textual Reasoning Tokens (t), which augment the text condition to capture high-level reasoning semantics. We subsequently analyze when, where, and how such tokens help. 4.1 Two Strategies of Reasoning Tokens We now introduce our two reasoning-token strategies, as illustrated in Figure 7. The formulation is agnostic to the specific video generator and applies to any DiT-based architecture. For concreteness we instantiate it on Wan2.2-I2V-A14B, the same backbone used for Wan-CoF. Visual reasoning tokens (vt). A DiT video generator takes as input a sequence of visual tokensx∈R T×H×W×d that represent the compressed latent of the video, flattens them into a 1D sequence, and processes them through a stack of self-attention blocks. To host reasoning state directly in this visual space, we introduceN r learnable Visual Reasoning Tokens, denotedr v ∈R N r ×d , initialized randomly and prepended to the visual token sequence before the first DiT block: z 0 = [r v 1 , r v 2 , . . . , r v N r , x 1 , x 2 , . . . , x M ],(1) whereM=T × H × Wis the total number of visual tokens. Inside every DiT block, self-attention lets each visual tokenx i attend to every reasoning tokenr v j , and the reasoning tokens in turn attend to the full visual 10 + Text Encoder Animatethedotsconnecting sequentiallyfrom1to8,each straightlineappearingsmoothly untilthefulloutlineemerges. Keepthebackgroundwiththe smilingsunandplantsunchanged. Staticshot. ... ... DiT Block × N ... Vision Encoder ... Diffusion Process ... Vision Decoder Self-attention VisualTokens Visual Reasoning Tokens Textual Sequence LoRA Text Encoder Animatethedotsconnecting sequentiallyfrom1to8,each straightlineappearingsmoothly untilthefulloutlineemerges. Keepthebackgroundwiththe smilingsunandplantsunchanged. Staticshot. ... DiT Block × N Vision Encoder ... Diffusion Process ... Vision Decoder Cross-attention VisualTokens Textual Sequence LoRA ... + Textual Reasoning Tokens (a) Visual Reasoning Tokens (b) TextualReasoning Tokens Figure 7 Overview of reasoning-token designs. (a) Visual Reasoning Tokens are inserted into the visual token sequence and interact with visual tokens through self-attention. (b) Textual Reasoning Tokens are appended to the text-conditioning sequence and guide generation through cross-attention. sequence and to each other. This bidirectional flow enables vt to aggregate global context and distribute the resulting features back to the visual tokens. After the final DiT block, the reasoning tokens are discarded and the prediction is read from the visual portion of the output. Textual reasoning tokens (t). A complementary place to host reasoning state is the text-conditioning side. The DiT consumes a text-conditioning sequencec∈R L×d produced by a text encoder, whereLis the length of that sequence. We introduceN t learnable Textual Reasoning Tokens, denotedr t ∈R N t ×d , and prepend them to this sequence before it enters the cross-attention of each DiT block: c 0 = [r t 1 , r t 2 , . . . , r t N t , c 1 , c 2 , . . . , c L ].(2) The cross-attention in each block then lets every visual tokenx i attend to both the original text tokens and the newr t j , sor t supplies task-level information that supplements the per-prompt content ofc. Unlike vt,r t enters only as additional key/value context and never appears on the query side, so it is not refreshed by the visual sequence and is shared, unchanged, across every spatial patch and every temporal frame. This also meansr t does not need to be removed before the final output readout. t therefore acts as a unidirectional, prompt-independent textual prior, complementary to the spatially grounded, bidirectional reasoning channel that vt places inside the visual stream. 4.2 Main Results We fine-tune Wan2.2-I2V-A14B with LoRA using vt and t separately, yielding Wan-CoF vt and Wan-CoF t . As shown in the Exploration of Reasoning Techniques rows of Tables 2 to 5, both variants improve further 11 0510152025303540 DiT block id 0.000 0.001 0.002 0.003 0.004 VT mean self-attention score VT high-noise model VT low-noise model T high-noise model T low-noise model 0.000 0.005 0.010 0.015 0.020 0.025 0.030 0.035 T mean cross-attention score (a) Depth-wise attention 01020304050 denoise step 0.0008 0.0010 0.0012 0.0014 0.0016 0.0018 VT mean self-attention score high-noise modellow-noise model VT high-noise model VT low-noise model T high-noise model T low-noise model 0.0120 0.0125 0.0130 0.0135 0.0140 0.0145 0.0150 T mean cross-attention score (b) Denoising-step attention Figure 8 Attention patterns of reasoning tokens. We average attention over 120 MME-CoF cases during inference. For vt, we report visual-token attention to vt in self-attention; for t, we report visual-token attention to t in cross-attention. Panel (a) shows depth-wise attention, and panel (b) shows denoising-step attention. over Wan-CoF on the headline metric of every benchmark. Wan-CoF vt reaches the highest aggregate score on Gen-ViRe and RULER-Bench, while Wan-CoF t leads on MME-CoF and VIPER. At the sub-dimension level, the two variants specialize in opposite directions, in a way consistent with their design: vt dominates dimensions that require maintaining a global, persistent plan inside the visual sequence, such as Planning in VIPER, Abstract, Analogy and Planning in Gen-ViRe, and Visual Stability in MME-CoF. In contrast, t dominates dimensions that benefit from a learnable, prompt-independent textual prior, such as Instruction Alignment in MME-CoF and Structural in VIPER. 4.3 Attention Analysis To understand how the tokens operate, we extract attention during inference on 120 MME-CoF cases and average two signals: visual tokens’ self-attention [39] scores assigned to vt and cross-attention [39] scores assigned to t. We then inspect these attention patterns along four axes: DiT block depth, denoising step, spatial position within frames, and temporal position across frames. Depth-wise attention. Panel (a) of Figure 8 shows that both vt and t are used non-uniformly across DiT depth, suggesting that reasoning tokens mainly participate in specific computation stages rather than acting as constant global biases. vt shows stronger layer selectivity: in the low-noise model it peaks around shallow-to-middle blocks, while in the high-noise model it remains active into later blocks and exhibits stronger mid-to-late spikes, consistent with visual reasoning tokens helping organize latent spatial-temporal structure when denoising is harder. t follows a smoother early-to-middle emphasis and then decays with depth, matching its role as a prompt-level prior supplied through cross-attention. Denoising-step attention. Panel (b) of Figure 8 shows a similar pattern along the denoising trajectory. Both token types receive stronger attention early in denoising, especially in the high-noise model, where global semantics and motion structure are still underdetermined. vt attention starts high and drops quickly in the high-noise model, indicating an early role in visual latent planning; in the low-noise model, it becomes flatter and lower, consistent with later refinement. t is smoother across denoising steps, but its high-noise curve stays above the low-noise curve, suggesting that task-level textual priors are most useful before the generation process settles. Spatial attention. Within a single frame, RT attention reveals different spatial roles for vt and t, as visualized in the top row of Figure 9. vt attention is relatively sparse and forms pronounced horizontal and vertical bands, with stronger responses near the gripper and the beans. This pattern suggests that vt captures localized visual cues while retaining axis-aligned structure from the visual token grid, possibly influenced by Wan’s separable 3D RoPE [37]. In contrast, t attention is more diffuse and covers broader task-relevant regions, 12 Single continuous take with no cuts. Begin with a static full-frame shot (1- 2s). Then animate the robot arm performing a smooth, constant-speed motion to wipe the beans on the table from left to right. Only the robot arm should move; the scene is rigid and static with no appearance or geometric changes throughout the shot. (a)VTSpatial Self-attention(b)TTSpatialCross-attention SpatialAttention Visualization Prompt Show the rough semicircular track with height label h and a block at P; release it, add faint friction streaks as it slides down and up the right side, stopping below the rim. Show the move quickly and completely. Static shot. (c)VTTemporal Self-attention (d)TTTemporal Cross-attention TemporalAttention Visualization Prompt Figure 9 Spatial and temporal attention visualizations. Top: representative high-noise spatial attention maps for vt self-attention and t cross-attention on an embodied manipulation prompt. Bottom: temporal attention maps for vt and t on a physics-reasoning prompt, where frames are uniformly sampled over time and ordered from left to right. including the robot arm, table surface, and target objects, consistent with t acting as a high-level semantic prior rather than a localized spatial tracker. Temporal attention. As shown in the bottom row of Figure 9, vt and t exhibit different temporal emphases rather than a simple object-tracking pattern. vt attention is more visible at the beginning and end of the sequence, suggesting that visual reasoning tokens are involved when the model anchors the initial state and resolves the final consequence of the motion. t attention is less pronounced in the earliest frame, but becomes comparably salient across the later sampled frames, consistent with a prompt-level prior that remains active once the temporal context has been established. This contrast suggests a complementary temporal division: vt emphasizes boundary states in the visual evolution, while t provides a steadier semantic constraint over the subsequent reasoning process. 5 Related Work 5.1 Modern Video Generation Models Video generation models have advanced rapidly in visual quality, temporal consistency, and instruction following. Architecturally, the shift from U-Net-based diffusion [1,31] to DiT [29] and flow-matching frameworks [26] has enabled scalable training and higher-resolution synthesis. Open systems such as HunyuanVideo [18], HunyuanVideo-1.5 [45], and Wan [41] provide strong backbones for image-to-video and text-to-video generation, while proprietary systems including Kling [19], Veo [9], Seedance [33], and Sora [28] continue to push generation quality and controllability. These advances make video generation a promising substrate for visual reasoning, but high-fidelity generation alone does not guarantee coherent multi-step reasoning. 5.2 Video Reasoning Benchmarks Reasoning has become a central capability for multimodal models [4, 16], but conventional chain-of-thought pipelines [43] often rely on text, static visual observations, external tools [15], or auxiliary image generation [20]. Video generation provides another path, where reasoning unfolds through temporally connected frames [44]. 13 Recent benchmarks, including MME-CoF [10], VideoThinkBench [38], V-ReasonBench [25], Gen-ViRe [23], RULER-Bench [12], MMGR [2], TiViBench [3], and VIPER [22], formalize and evaluate this capability from different perspectives. OpenCoF builds on these benchmarks but focuses on improving CoF reasoning through the OpenCoF-17K dataset and reasoning-token exploration. 5.3 Video Generation Models for Reasoning Training video generation models for reasoning remains less explored than benchmarking. VR-Bench [47] studies maze-solving and test-time scaling, Thinking in Frames [21] focuses on domains such as mazes and tangram puzzles, and VBVR [42] scales video-reasoning data within a cognitive task suite. These efforts suggest that targeted supervision improves video reasoning, but they are often domain-specific or internally coupled between training and evaluation. OpenCoF-17K instead combines four data sources and supports transfer evaluation on independent external benchmarks, while OpenCoF further explores how dedicated reasoning techniques organize reasoning state through Visual and Textual Reasoning Tokens. 6 Conclusion and Limitations We present the OpenCoF-17K Dataset, a diverse video reasoning dataset for improving Chain-of-Frame reasoning in video generation models. Fine-tuning Wan2.2-I2V-A14B on the OpenCoF-17K Dataset yields Wan-CoF, showing that diverse temporal supervision improves CoF behavior across four external benchmarks. We further explore Visual and Textual Reasoning Tokens and analyze how they organize reasoning state through benchmark results and attention patterns. One limitation is that vt and t are investigated separately. How to effectively combine them remains future work. 14 References [1]Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A space-time diffusion model for video generation. InSIGGRAPHAsia 2024ConferencePapers, pages 1–11, 2024. [2]Zefan Cai, Haoyi Qiu, Tianyi Ma, Haozhe Zhao, Gengze Zhou, Kung-Hsiang Huang, Parisa Kordjamshidi, Minjia Zhang, Wen Xiao, Jiuxiang Gu, et al. Mmgr: Multi-modal generative reasoning.arXivpreprintarXiv:2512.14691, 2025. [3]Harold Haodong Chen, Disen Lan, Wen-Jie Shu, Qingyang Liu, Zihan Wang, Sirui Chen, Wenkai Cheng, Kanghao Chen, Hongfei Zhang, Zixin Zhang, et al. Tivibench: Benchmarking think-in-video reasoning for video generative models.arXivpreprintarXiv:2511.13704, 2025. [4]Xinyan Chen, Renrui Zhang, Dongzhi Jiang, Aojun Zhou, Shilin Yan, Weifeng Lin, and Hongsheng Li. Mint-cot: Enabling interleaved visual tokens in mathematical chain-of-thought reasoning.arXivpreprintarXiv:2506.05331, 2025. [5]Hyung Won Chung, Noah Constant, Xavier Garcia, Adam Roberts, Yi Tay, Sharan Narang, and Orhan Firat. Unimax: Fairer and more effective language sampling for large-scale multilingual pretraining.arXivpreprint arXiv:2304.09151, 2023. [6]Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXivpreprintarXiv:2507.06261, 2025. [7]Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model.arXiv preprintarXiv:2312.11370, 2023. [8] Yu Gao, Haoyuan Guo, Tuyen Hoang, Weilin Huang, Lu Jiang, Fangyuan Kong, Huixia Li, Jiashi Li, Liang Li, Xiaojie Li, et al. Seedance 1.0: Exploring the boundaries of video generation models.arXivpreprint arXiv:2506.09113, 2025. [9] Google DeepMind. Veo-3 technical report. Technical report, Google DeepMind, May 2025. URLhttps: //storage.googleapis.com/deepmind-media/veo/Veo-3-Tech-Report.pdf. [10]Ziyu Guo, Xinyan Chen, Renrui Zhang, Ruichuan An, Yu Qi, Dongzhi Jiang, Xiangtai Li, Manyuan Zhang, Hongsheng Li, and Pheng-Ann Heng. Are video models ready as zero-shot reasoners? an empirical study with the mme-cof benchmark.arXivpreprintarXiv:2510.26802, 2025. [11] Ziyu Guo, Rain Liu, Xinyan Chen, and Pheng-Ann Heng. Atlas: Agentic or latent visual reasoning? one word is enough for both.arXivpreprintarXiv:2605.15198, 2026. [12]Xuming He, Zehao Fan, Hengjia Li, Fan Zhuo, Hankun Xu, Senlin Cheng, Di Weng, Haifeng Liu, Can Ye, and Boxi Wu. Ruler-bench: Probing rule-based reasoning abilities of next-level video generation models for vision foundation intelligence.arXivpreprintarXiv:2512.02622, 2025. [13]Jack Hong, Chenxiao Zhao, ChengLin Zhu, Weiheng Lu, Guohai Xu, and Xing Yu. Deepeyesv2: Toward agentic multimodal model.arXivpreprintarXiv:2511.05271, 2025. [14]Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022. [15]Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models.Advancesin NeuralInformationProcessingSystems, 37:139348–139379, 2024. [16] Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, et al. Mme-cot: Benchmarking chain-of-thought in large multimodal models for reasoning quality, robustness, and efficiency.arXivpreprintarXiv:2502.09621, 2025. [17] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners.Advancesinneuralinformationprocessingsystems, 35:22199–22213, 2022. 15 [18]Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXivpreprint arXiv:2412.03603, 2024. [19] Kuaishou Technology. Kling ai: Next-generation ai creative studio. https://klingai.com/, June 2024. [20]Chengzu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vulić, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought.arXivpreprintarXiv:2501.07542, 2025. [21]Chengzu Li, Zanyi Wang, Jiaang Li, Yi Xu, Han Zhou, Huanyu Zhang, Ruichuan An, Dengyang Jiang, Zhaochong An, Ivan Vulić, et al. Thinking in frames: How visual context and test-time scaling empower video reasoning. arXivpreprintarXiv:2601.21037, 2026. [22]Yifan Li, Yukai Gu, Yingqian Min, Zikang Liu, Yifan Du, Kun Zhou, Min Yang, Wayne Xin Zhao, and Minghui Qiu. Viper: Process-aware evaluation for generative video reasoning.arXivpreprintarXiv:2512.24952, 2025. [23]Xinxin Liu, Zhaopan Xu, Ming Li, Kai Wang, Yong Jae Lee, and Yuzhang Shang. Can world simulators reason? gen-vire: A generative visual reasoning benchmark.arXivpreprintarXiv:2511.13853, 2025. [24]Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chun yue Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models.ArXiv, abs/2310.02255, 2023. [25]Yang Luo, Xuanlei Zhao, Baijiong Lin, Lingting Zhu, Liyao Tang, Yuqi Liu, Ying-Cong Chen, Shengju Qian, Xin Wang, and Yang You. V-reasonbench: Toward unified reasoning benchmark suite for video generation models. arXivpreprintarXiv:2511.16668, 2025. [26]Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. InEuropean ConferenceonComputerVision, pages 23–40. Springer, 2024. [27]Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Tiancheng Han, Botian Shi, Wenhai Wang, Junjun He, et al. Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning.arXivpreprintarXiv:2503.07365, 2025. [28]OpenAI. Sora 2 system card. Technical report, OpenAI, September 2025. URLhttps://cdn.openai.com/pdf/ 50d5973c-c4f-4c2d-986f-c72b5d0f069/sora_2_system_card.pdf. [29] William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedingsoftheIEEE/CVF internationalconferenceoncomputervision, pages 4195–4205, 2023. [30]Yu Qi, Xinyi Xu, Ziyu Guo, Siyuan Ma, Renrui Zhang, Xinyan Chen, Ruichuan An, Ruofan Xing, Jiayi Zhang, Haojie Huang, et al. Mme-cof-pro: Evaluating reasoning coherence in video generative models with text and visual hints.arXivpreprintarXiv:2603.20194, 2026. [31]Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternationalConferenceonMedicalimagecomputingandcomputer-assistedintervention, pages 234–241. Springer, 2015. [32]Team Seedance, Heyi Chen, Siyan Chen, Xin Chen, Yanfei Chen, Ying Chen, Zhuo Chen, Feng Cheng, Tianheng Cheng, Xinqi Cheng, et al. Seedance 1.5 pro: A native audio-visual joint generation foundation model.arXiv preprintarXiv:2512.13507, 2025. [33]Team Seedance, De Chen, Liyang Chen, Xin Chen, Ying Chen, Zhuo Chen, Zhuowei Chen, Feng Cheng, Tianheng Cheng, Yufeng Cheng, et al. Seedance 2.0: Advancing video generation for world complexity.arXivpreprint arXiv:2604.14148, 2026. [34]Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning.AdvancesinNeuralInformationProcessingSystems, 37:8612–8642, 2024. [35]David Sharon and Nicole Brichtova. Image editing in gemini just got a major upgrade. The Keyword (Google Blog), August 26 2025. URLhttps://blog.google/products-and-platforms/products/gemini/ updated-image-editing-model/. 16 [36]Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXivpreprintarXiv:1701.06538, 2017. [37]Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024. [38] Jingqi Tong, Yurong Mou, Hangcheng Li, Mingzhe Li, Yongzhuo Yang, Ming Zhang, Qiguang Chen, Tianyi Liang, Xiaomeng Hu, Yining Zheng, et al. Thinking with video: Video generation as a promising multimodal reasoning paradigm.arXivpreprintarXiv:2511.04570, 2025. [39]Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advancesinneuralinformationprocessingsystems, 30, 2017. [40]Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, An- dre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. InConferenceonRobotLearning, pages 1723–1736. PMLR, 2023. [41]Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, Tianxing Wang, Tianyi Gui, Tingyu Weng, Tong Shen, Wei Lin, Wei Wang, Wei Wang, Wenmeng Zhou, Wente Wang, Wenting Shen, Wenyuan Yu, Xianzhong Shi, Xiaoming Huang, Xin Xu, Yan Kou, Yangyu Lv, Yifei Li, Yijing Liu, Yiming Wang, Yingya Zhang, Yitong Huang, Yong Li, You Wu, Yu Liu, Yulin Pan, Yun Zheng, Yuntao Hong, Yupeng Shi, Yutong Feng, Zeyinzi Jiang, Zhen Han, Zhi-Fan Wu, and Ziyu Liu. Wan: Open and advanced large-scale video generative models.arXivpreprint arXiv:2503.20314, 2025. [42]Maijunxian Wang, Ruisi Wang, Juyi Lin, Ran Ji, Thaddäus Wiedemer, Qingying Gao, Dezhi Luo, Yaoyao Qian, Lianyu Huang, Zelong Hong, et al. A very big video reasoning suite.arXivpreprintarXiv:2602.20159, 2026. [43]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advancesinneuralinformationprocessing systems, 35:24824–24837, 2022. [44]Thaddäus Wiedemer, Yuxuan Li, Paul Vicol, Shixiang Shane Gu, Nick Matarese, Kevin Swersky, Been Kim, Priyank Jaini, and Robert Geirhos. Video models are zero-shot learners and reasoners.arXivpreprintarXiv:2509.20328, 2025. [45]Bing Wu, Chang Zou, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Jack Peng, Jianbing Wu, Jiangfeng Xiong, Jie Jiang, et al. Hunyuanvideo 1.5 technical report.arXivpreprintarXiv:2511.18870, 2025. [46]Yi Xu, Chengzu Li, Han Zhou, Xingchen Wan, Caiqi Zhang, Anna Korhonen, and Ivan Vulić. Visual planning: Let’s think only with images.arXivpreprintarXiv:2505.11409, 2025. [47]Cheng Yang, Haiyuan Wan, Yiran Peng, Xin Cheng, Zhaoyang Yu, Jiayi Zhang, Junchi Yu, Xinlei Yu, Xiawu Zheng, Dongzhan Zhou, et al. Reasoning via video: The first evaluation of video models’ reasoning abilities through maze-solving tasks.arXivpreprintarXiv:2511.15065, 2025. [48]Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, et al. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization. InProceedingsoftheIEEE/CVFInternationalConferenceonComputerVision, pages 2376–2385, 2025. [49]Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? ECCV2024, 2024. [50]Ziwei Zheng, Michael Yang, Jack Hong, Chenxiao Zhao, Guohai Xu, Le Yang, Chao Shen, and Xing Yu. Deepeyes: Incentivizing" thinking with images" via reinforcement learning.arXivpreprintarXiv:2505.14362, 2025. 17 Appendix Overview We organize our supplementary material as follows. • Additional Implementation Details – Backbone and LoRA Setup – Model Variants – Compared Systems – Evaluation Protocol • Additional Ablation Study – Reasoning Token Count Ablation A Additional Implementation Details A.1 Backbone and LoRA setup. Wan2.2-I2V-A14B adopts a Mixture-of-Experts [36] denoising architecture with separate high-noise and low-noise experts across diffusion timesteps, an umT5 [5] text encoder, a Wan-VAE for visual encoding and decoding, and a mainstream DiT denoiser. We fine-tune all three models (Wan-CoF, Wan-CoF vt , and Wan-CoF t ) from the same Wan2.2-I2V-A14B backbone using LoRA [14], applied only to the DiT denoiser while the umT5 text encoder and the Wan-VAE are kept frozen. Within each DiT block, LoRA targets the attention projections (q,k,v,o) together with the two MLP linear layers (ffn.0,ffn.2), with rank set to 32. The learning rate is configured to 2× 10 −5 . A.2 Model variants. To attribute every gain to the dataset alone, we keep this architecture entirely unchanged and fine-tune it on OpenCoF-17K with LoRA, yielding Wan-CoF. For reasoning-token exploration, we run two independent fine-tuning experiments with LoRA on Wan2.2-I2V-A14B, one with only vt and one with only t, yielding the models Wan-CoF vt and Wan-CoF t , where we setN r =N t = 16. Wan-CoF is trained on OpenCoF-17K for 2 epochs, while Wan-CoF vt and Wan-CoF t are trained for 5 epochs to give the additional reasoning-token parameters more updates to converge. A.3 Compared systems. We compare Wan-CoF with both closed-source and open-source video generation models. The closed-source set covers state-of-the-art systems including Kling-v1 [19], the Seedance series [33] Seedance-1.0-Pro and Seedance-1.5, the Veo-3 [9] series Veo-3.0-Fast, Veo-3.0-Preview, Veo-3.0 and Veo-3.1, and Sora-2 [28]. The open-source set includes HunyuanVideo-I2V [18] and HunyuanVideo-1.5 [45]. A.4 Evaluation protocol. We evaluate our models on four video reasoning benchmarks: MME-CoF [10], RULER-Bench [12], VIPER [22], and Gen-ViRe [23]. Together, these benchmarks span reasoning categories well beyond the 11 task types in OpenCoF-17K. For RULER-Bench specifically, we restrict evaluation to the image-to-video cases, since our baseline does not support text-to-video generation. For every benchmark we follow its official judge model. B Additional Ablation Study Reasoning token count ablation. The exploration in Section 4 fixesN r =N t = 16. To check whether a larger token budget changes our conclusions, we re-train Wan-CoF vt and Wan-CoF t withn= 32 under otherwise 18 Table 6 Ablation on the number of reasoning tokens. Headline metrics on all four benchmarks for vt and t with n∈16,32; Wan-CoF is included as a reference. RULER Overall is the overall score on RULER-Bench, averaged across the four scoring dimensions (Instruction Following, Visual Consistency, Visual Fidelity, and Rule Coherence). Bold marks the better value within each n=16 / n=32 pair. Model MME-CoF Gen-ViRe VIPER RULER-Bench OverallAvg.POC↑Overall Wan-CoF1.300.3917.556.8 Wan-CoF vt (n=16)1.340.4418.259.6 Wan-CoF vt (n=32)1.280.4168.958.5 Wan-CoF t (n=16)1.350.4068.858.4 Wan-CoF t (n=32) 1.340.4126.857.7 identical settings and report the headline metric of every benchmark in Table 6. Doubling the token count does not produce a uniform improvement for either variant: each benchmark moves in a different direction, indicating that reasoning tokens do not act as generic capacity that scales withn. At the benchmark level, n= 16 outperformsn= 32 on three of the four benchmarks for both vt and t, withn= 32 leading only on a single benchmark in each case. Combined with the fact thatn= 32 also roughly doubles the reasoning-token parameters, we keep n = 16 as the default in the main text. 19