Paper deep dive
From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models
Meng Luo, Yicheng Liu, Jiahao Wang, Yuanxing Zhang, Xin Tao, Pengfei Wan, Kun Gai, Hao Fei
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/11/2026, 4:43:10 AM
Summary
The paper introduces VWG-Bench, a comprehensive benchmark for evaluating 'think-with-video' reasoning across 9 dimensions and 38 tasks, utilizing a three-level VLM-as-Judge protocol to assess video fluency, task rule adherence, and goal realization. It identifies a gap where video models excel in rendering but fail in logical reasoning. To address this, the authors propose Vid-PRE, a model-agnostic prompt rewriter trained via SFT and GRPO with text-based rewards, which significantly improves reasoning capabilities across various video generators without architectural changes.
Entities (8)
Relation Signals (9)
Jiahao Wang → affiliatedwith → Kuaishou Technology
confidence 95% · Jiahao Wang 3... 3 Kling Team, Kuaishou Technology
Yicheng Liu → affiliatedwith → Nanjing University
confidence 95% · Yicheng Liu 2... 2 Department of Computer Science and Technology, Nanjing University
Meng Luo → affiliatedwith → National University of Singapore
confidence 95% · Meng Luo 1... 1 School of Computing, National University of Singapore
Vid-PRE → enhances → Video Generative Models
confidence 95% · Vid-PRE yields substantial reasoning improvements across multiple generators without architectural modifications.
VWG-Bench → hasauthor → Meng Luo
confidence 95% · From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models Meng Luo 1,∗,†
VWG-Bench → proposessolutionfor → Think-with-Video Reasoning
confidence 95% · We introduce VWG-Bench... a comprehensive benchmark spanning 9 reasoning dimensions... to stress-test think-with-video reasoning
Vid-PRE → usesalgorithm → SFT
confidence 92% · Vid-PRE is trained via Supervised Fine-tuning (SFT) on rejection-sampled reasoning tuples
Vid-PRE → usesalgorithm → GRPO
confidence 92% · Vid-PRE is trained via Supervised Fine-tuning (SFT) on rejection-sampled reasoning tuples and further optimized with Group Relative Policy Optimization (GRPO)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Video generation has advanced to produce visually compelling and temporally coherent results. Yet, whether these models can genuinely think with video--executing symbolic rules, respecting physical laws, and pursuing intentional goals--remains an open question. Existing benchmarks only partially address this, often conflating visual quality with cognitive correctness. We introduce VWG-Bench (Video World Generalist Benchmark), a comprehensive benchmark spanning 9 reasoning dimensions and 38 fine-grained tasks. To enable precise diagnosis, we design a three-level VLM-as-Judge protocol that independently assesses video-level fluency, task-level rule adherence, and sample-level goal realization. Evaluations of leading models reveal a striking gap: while models achieve strong rendering scores, they consistently fail on logic-heavy and rule-constrained tasks. To address this, we propose Vid-PRE (Video Prompt Reasoner and Enhancer), a model-agnostic prompt rewriter that offloads the cognitive burden of reasoning to a dedicated VLM. Trained via reinforcement learning with purely text-based rewards, Vid-PRE produces concise, constraint-aware prompts without the instability of video-level reward signals. Experiments show that Vid-PRE yields substantial reasoning improvements across multiple generators without architectural modifications. Together, VWG-Bench and Vid-PRE offer a rigorous diagnostic lens and a scalable path toward true think-with-video capabilities. All data and code are publicly available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2609.11242v1
- Canonical: https://arxiv.org/abs/2609.11242v1
Trouble viewing inline? Open PDF directly →
Full Text
79,146 characters extracted from source content.
Expand or collapse full text
From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models Meng Luo 1,∗,† , Yicheng Liu 2,∗,† , Jiahao Wang 3,‡ , Yuanxing Zhang 3 , Xin Tao 3 , Pengfei Wan 3 , Kun Gai 3 , and Hao Fei 4,‡ 1 School of Computing, National University of Singapore, Singapore, Singapore 2 Department of Computer Science and Technology, Nanjing University, Nanjing, China 3 Kling Team, Kuaishou Technology, Beijing, China 4 Department of Computer Science, University of Oxford, Oxford, United Kingdom Abstract. Video generation has advanced to produce visually com- pelling and temporally coherent results. Yet, whether these models can genuinely think with video—executing symbolic rules, respecting phys- ical laws, and pursuing intentional goals—remains an open question. Existing benchmarks only partially address this, often conflating vi- sual quality with cognitive correctness. We introduce VWG-Bench (Video World Generalist Benchmark), a comprehensive benchmark spanning 9 reasoning dimensions and 38 fine-grained tasks. To enable precise di- agnosis, we design a three-level VLM-as-Judge protocol that indepen- dently assesses video-level fluency, task-level rule adherence, and sample- level goal realization. Evaluations of leading models reveal a striking gap: while models achieve strong rendering scores, they consistently fail on logic-heavy and rule-constrained tasks. To address this, we propose Vid-PRE (Video Prompt Reasoner and Enhancer), a model-agnostic prompt rewriter that offloads the cognitive burden of reasoning to a dedicated VLM. Trained via reinforcement learning with purely text- based rewards, Vid-PRE produces concise, constraint-aware prompts without the instability of video-level reward signals. Experiments show that Vid-PRE yields substantial reasoning improvements across multiple generators without architectural modifications. Together, VWG-Bench and Vid-PRE offer a rigorous diagnostic lens and a scalable path toward true think-with-video capabilities. All data and code are publicly available at https://huggingface.co/datasets/KlingTeam/VWG-Bench. Keywords: Video Generation· Evaluation Benchmark· Prompt En- hancement· Reinforcement Learning ∗ Equal contribution. † Work done during an internship at Kling Team, Kuaishou Technology. ‡ Corresponding authors. arXiv:2609.11242v1 [cs.CV] 10 Sep 2026 2M. Luo et al. 1 Introduction Reasoning through continuous spatio-temporal signals is a hallmark of intelli- gence that text and static images cannot fully replicate [7,22,28,45]. “Thinking with Text” established Chain-of-Thought (CoT) [27, 30, 37, 42] prompting as a foundational paradigm for language reasoning; “Thinking with Images” extended this to visual chains in Vision-Language Models [23,33,44]. A natural next fron- tier is Think with Video: reasoning expressed through continuous video frames that encode physical dynamics, object interactions, and long-horizon goal execu- tion in a unified temporal substrate. Recent advances, including Veo 3’s chain-of- frames, Sora-2’s visual problem solving [13,26,34], and emergent spatial planning in open-source models, suggest that video generation is crossing a threshold from photorealistic rendering toward genuine world simulation [6, 24, 36]. Yet a fun- damental question remains unresolved: do these models truly understand the rules they appear to follow, or are they exploiting surface statistics to produce plausible-looking sequences? Answering this question requires benchmarks that probe reasoning rather than aesthetics. Existing efforts have made important strides: TiViBench [4] in- troduces hierarchical evaluation across four structural and logical dimensions; V-ReasonBench [29] provides reproducible, answer-verifiable tasks grounded in spatial and physical cognition; VR-Bench [40] systematically measures spatial planning through maze-solving; RULER-Bench [16] targets cognitive-rule ad- herence across six categories; RISE-Video [25] probes implicit world-rule under- standing across eight knowledge domains; and MMGR [3] covers abstract and embodied reasoning. Despite this progress, the evaluation landscape remains fragmented. Each benchmark covers only a subset of the reasoning spectrum. Moreover, existing protocols typically collapse evaluation into a single binary or holistic score, masking where and why models fail, specifically whether at the level of visual fluency, task-level rule tracking, or precise goal realization. To close both gaps, we introduce VWG-Bench, a comprehensive and hierarchi- cal benchmark designed to stress-test think-with-video reasoning across the full spectrum of world simulation. As illustrated in Figure 1, VWG-Bench covers 9 core reasoning dimensions and 38 fine-grained tasks. Specifically, these dimensions systematically evaluate: (1) Symbolic Reasoning & Logic Puzzles, (2) Graph- ical Reasoning & Pattern Matching, (3) Embodied & Experiential Reasoning, (4) GUI & Digital Interaction, (5) Hypothetical & Counterfactual Reasoning, (6) Spatial Geometry & Logic, (7) Multiscale Temporal Reasoning, (8) Natu- ral Laws & World Simulation, and (9) Social Understanding & Commonsense. To construct evaluation samples at scale, we develop a highly automated data pipeline that orchestrates image generation, consistency filtering, and intent for- mulation, enabling scalable production of high-quality instances. Beyond scale and coverage, VWG-Bench introduces a three-level VLM-as- Judge evaluation framework with a fine-grained 1–5 scoring scale. Unlike existing holistic scoring that conflates rendering quality with cognitive correctness, our protocol explicitly decouples three orthogonal facets of performance: Video-level, assessing overall visual fluency and temporal coherence; Task-level, measuring Think-with-Video Reasoning for Video Generative Models3 1. Symbolic Reasoning & Logic Puzzles √ Arithmetic √ Tic-tac-toe √ Code √ ... 2. Graphical Reasoning & Pattern Matching √ Seq. Completion √ Box Matching √ Shape Fitting √ Visual Analogy 3. Embodied & Experiential Reasoning √ Implicit goal reasoning √ Embodied Trajectory Reasoning √ ... 4. GUI & Digital Interaction √ GUI Instruction Execution √ GUI Click Reasoning 5. Hypothetical & Counterfactual Reasoning √ Reverse Temporal Logic √ PhyicsInversion √ ... 6. Spatial Geometry & Logic √ Spatial Geometry √Maze √ Rotation Reasoning √ ... 7. Multiscale Temporal Reasoning √ Long-term Progression √ Short-term Dynamics√ ... 8. Natural Laws & World Simulation √ Physics-based Reasoning √ Biological Kinematics √ ... 9. Social Understanding & Commonsense √ Social Rules √ Cultural Custom √ Emotion Recognition √ ... Samples Task Level Constraints -Foreground/background -Implicit Rules e.g.no mesh pass (maze) physical rule (embodied) Sample Level Goals -Progress Goal e.g.changes after click a button (gui) -Last Frame Goal e.g. Right solution of Sudoku (sudoku) Video Model ... Videos Video Quality(1-5) Progress Cons.(1-5) Implicit Rule(1-5) Progress goal(1-5) Last frame goal(1-5) video sample task Evaluation Dimensions & Task Taxonomy (9 Core Dimensions) Fig. 1: Comprehensive overview of VWG-Bench’s taxonomy, automated data construc- tion, and structured evaluation framework. progress consistency and adherence to implicit task rules across frames; and Sample-level, evaluating the precise realization of intermediate milestones and final objectives. This decomposition enables fine-grained, interpretable diagnos- tics that identify failure modes at each level of the reasoning hierarchy. Extensive evaluations of leading commercial and open-source models on VWG-Bench ex- pose a consistent pattern: models achieve strong Video-level rendering scores but suffer severe degradation at the Task- and Sample-level on compositional, rule-constrained, and logic-heavy scenarios. This gap is not primarily a failure of visual generation; it is a failure of reasoning through prompts: given only a natural-language description, models lack the structured representation of task rules, success criteria, and constraint hierarchies needed to guide generation to- ward correct reasoning trajectories. To directly address this limitation, we propose Vid-PRE, a model-agnostic prompt rewriter that decouples the reasoning stage from the rendering stage. Vid-PRE offloads the cognitive burden to a dedicated VLM, which infers task rules and success criteria from the image context and distills them into a concise, 4M. Luo et al. constraint-aware prompt for the downstream video generator. Vid-PRE is trained via Supervised Fine-tuning (SFT) on rejection-sampled reasoning tuples and further optimized with Group Relative Policy Optimization (GRPO) [32] using a meticulously designed purely text-based rewards system, keeping it agnostic to any specific generator architecture. Extensive experiments confirm substantial improvements across multiple generators without any architectural modification. In summary, our contributions are threefold: – A Unified Think-with-Video Benchmark. VWG-Bench is the first bench- mark to jointly cover all nine core dimensions of think-with-video world sim- ulation across 38 tasks, constructed via a fully automated data pipeline for scalable and reproducible scenario generation. – A Fine-Grained, Hierarchical Evaluation Protocol. Our three-level VLM-as-Judge framework with a 1–5 scoring scale diagnoses Video-level flu- ency, Task-level rule adherence, and Sample-level goal realization, revealing failure modes that holistic or binary evaluation cannot. – A Universal and Plug-and-Play Prompt Reasoner and Enhancer. Vid-PRE is a model-agnostic rewriter trained via SFT and GRPO with text- based rewards that significantly boosts reasoning accuracy across multiple video generators, without any modification to their architectures. 2 VWG-Bench: Video World Generalist Benchmark 2.1 Design Principles VWG-Bench is designed to comprehensively evaluate the reasoning capabilities of video generation models. Compared with prior reasoning benchmarks for video generation, VWG-Bench aims to support both broader task coverage and a more systematic evaluation framework. For task coverage, as shown in Table 1, our benchmark spans nine major rea- soning dimensions and 38 task categories, fully covering symbolic and graphical reasoning, embodied and interactive tasks, as well as spatial–temporal reasoning, physical world simulation, and social commonsense understanding. To balance evaluation comprehensiveness with practical cost, we sample only 10 high-quality instances from each task category, significantly reducing the computational and annotation overhead required for evaluation. For the evaluation framework, as illustrated in Figure 1, we establish a three- level evaluation protocol. The video-level performs coarse-grained quality control to ensure the generated video is a valid visual medium. The task-level verifies whether the model respects task-specific constraints and logical rules throughout the video generation process. Finally, the sample-level assesses whether the final outcome satisfies the intended task objective. Compared to prior benchmarks that typically involve only one or two evaluation dimensions, our framework provides a more comprehensive and structured assessment of reasoning perfor- mance. Think-with-Video Reasoning for Video Generative Models5 2.2 Evaluation Dimensions and Task Taxonomy Here we introduce the nine evaluation dimensions and the 38 task categories covered by VWG-Bench. VWG-Bench organizes tasks into nine reasoning dimensions spanning sym- bolic logic, visual pattern induction, embodied interaction, digital interfaces, counterfactual reasoning, spatial geometry, temporal evolution, world simula- tion, and social understanding. Symbolic Reasoning & Logic Puzzles eval- uate the execution of formal rules and algorithms, such as “Arithmetic”, “Coding”, and “Sudoku”. Graphical Reasoning & Pattern Matching focuses on discov- ering visual correspondences and completing structured patterns, such as “Box Matching”, “Color Connecting”, and “Visual Analogy”. Embodied & Experien- tial Reasoning examines reasoning in physical interaction scenarios, including tasks like “Embodied Action Execution”, “Embodied Trajectory Reasoning”, and “Implicit Goal Reasoning”. GUI & Digital Interaction evaluates reasoning within graphical user interfaces, such as “GUI Instruction Execution” and “GUI Click Reasoning”. Hypothetical & Counterfactual Reasoning probes rea- soning under altered or fictional premises, for example “Reverse Temporal Logic”, “Physics Inversion”, and “Causal Intervention”. Spatial Geometry & Logic tests geometric consistency and spatial transformations, including tasks such as “Maze”, “Rotation Reasoning”, and “Spatial Geometry”. Multiscale Temporal Reasoning evaluates the ability to maintain coherent dynamics across different time horizons, such as “Short-term Temporal Dynamics” and “Long-term Tempo- ral Progression”. Natural Laws & World Simulation measures whether mod- els follow real-world physical or biological principles, for example “Physics-based Reasoning”, “Biological Kinematics”, and “Physical Commonsense Reasoning”. Social Understanding & Commonsense focuses on human-centric seman- tics and social behavior modeling, including tasks like “Social Rules”, “Cultural Customs”, and “Emotion Recognition”. 2.3 Data Construction Pipeline Image-Pair Initialization. To collect prompt-image pairs across such a diverse set of task categories, we adopt three complementary strategies: (1) identifying suitable open-source datasets and sampling high-quality categories from them (e.g., GUI and embodied tasks); (2) generating samples through programmatic algorithms (e.g., arithmetic and Sudoku tasks); or (3) automatically generating images via Text-to-Image (T2I) models using predefined prompt templates and dynamic variable pools to ensure broad scenario diversity. To guarantee visual integrity, we apply a strict visual filtering mechanism to T2I outputs. A VLM acts as an impartial judge to verify whether the generated image faithfully adheres to the T2I prompt without missing elements or visual hallucinations. If the image fails this check, the pipeline either regenerates the image or discards it entirely. Multi-Granularity Evaluation Annotation. To support our fine-grained, three-level evaluation framework, the pipeline automatically annotates diagnos- tic criteria that form the core ground truth for VWG-Bench. For task-level con- straints, we define rule dictionaries that delineate the expected behaviors of the 6M. Luo et al. Table 1: Comparison of VWG-Bench with representative video-generation reasoning benchmarks. Sym: Symbolic & Logic, Gph: Graphical & Pattern, Emb: Embodied & Experiential, GUI: GUI & Digital Interaction, Cft: Hypothetical & Counterfactual, Spa: Spatial Geometry, Tmp: Multiscale Temporal, Phy: Natural Laws & Physics, Soc: Social & Commonsense). Auto = whether an automated data-construction pipeline is provided.✓ = fully covered; ◦ = partially covered;✗ = not covered. ScaleReasoning CoverageEvaluation BenchmarkDims Tasks Samples Sym Gph Emb GUI Cft Spa Tmp Phy Soc Levels Auto MME-CoF [15]/1259✗ ◦✓✗✓ ◦✓✗1✗ VideoThinkBench [34]254,149 ◦✓✗ ◦✗ ◦✗1✗ TiViBench [4]424595✓ ◦✗✓✗1✗ VR-Bench [40]157,920✗✓✗1✗ V-ReasonBench [29]412327✓✗✓✗✓✗1✗ RULER-Bench [16]640622✓✗✓ ◦✓✗1✗ MMGR [3]3101,853✓✗✓✗✓✗✓✗1✗ RISE-Video [25]829467✓ ◦✓✗✓ ◦✓1✓ VWG-Bench (Ours)938380✓3✓ foreground (e.g., allowed state changes for manipulated objects), the background (e.g., strict static invariance for unmanipulated areas), and implicit rules (e.g., forbidding mesh interpenetration during object stacking). For sample-level goals, the VLM simulates the outcome of the action to generate linguistic descriptions of the expected final state (i.e., Last-Frame Goal, G last ). For tasks with com- plex temporal dynamics, intermediate milestones (i.e., Progress Goal, G prog ) are additionally annotated. 2.4 Evaluation Framework and Protocols Our evaluation framework is designed to systematically assess the reasoning capabilities of video generation models, prioritizing logical coherence and phys- ical adherence over mere visual fidelity. To mitigate evaluation hallucinations, we employ a VLM-as-Judge paradigm, where the judge strictly cross-references generated frames against structured annotations. All metrics across our three- level framework are scored on a 1–5 integer scale (according to the degree of consistency with the corresponding annotations). Video-Level Assessment (Coarse-Grained Quality Control). Before probing reasoning, we must ensure the generated video is a valid visual medium. This foundational filter evaluates Video Quality by assessing holistic visual fi- delity and temporal smoothness. The VLM judge explicitly penalizes technical flaws that obscure the video content—such as severe camera jitter, temporal in- coherence, unnatural flickering, or excessive static frames—ensuring that poor baseline rendering does not confound the subsequent reasoning evaluation. Task-Level Assessment (Fine-Grained Consistency & Constraints). This level leverages rule dictionaries to verify whether the model respects spatial boundaries and logical constraints throughout the generated video. Specifically, Progress Consistency scrutinizes the temporal coherence of dynamic elements by Think-with-Video Reasoning for Video Generative Models7 Fig. 2: An overview of the training framework for Vid-PRE. checking whether the foreground evolves logically while penalizing unprompted changes or morphological collapse in the background. Implicit Rule Following assesses adherence to physical laws and hidden boundaries inherent to the task, strictly penalizing violations such as objects interpenetrating solid meshes, re- versing gravity, or rigid bodies undergoing unreasonable soft deformations. Sample-Level Assessment (Goal Realization). The final level evaluates the semantic alignment between the temporal states of the video and the user’s explicit instructions. Progress Goal evaluates necessary intermediate milestones; for complex temporal progressions, the judge inspects intermediate keyframes to ensure the “Start → Intermediate → End” sequence is correctly chained with- out skipping or reversing logical steps. Last-Frame Goal assesses ultimate task completion by comparing the terminal state of the generated video against the annotated goal description to determine whether the objective was achieved. 3 Vid-PRE: A Universal Prompt Reasoner and Enhancer Current Image-to-Video (I2V) generation models excel at visual rendering but struggle with complex logical derivation and constraint satisfaction. To address this, we introduce Vid-PRE, a model-agnostic prompt rewriting module. The core philosophy of Vid-PRE is a strict division of labor: it offloads the cognitive burden of spatio-temporal reasoning to a dedicated VLM, allowing the down- stream video generator to focus exclusively on rendering based on highly explicit and executable instructions. 3.1 Problem Formulation We formulate video prompt rewriting as a conditional text generation problem. Let I denote the first-frame image and P user the user instruction. A naive for- mulation would learn a direct mapping (I,P user )→P opt , but this often leads to 8M. Luo et al. prompts that lack explicit logical grounding. To encourage structured reasoning during rewriting, we introduce an intermediate representation CoT that cap- tures the reasoning process linking the visual context and the final executable prompt. The objective of Vid-PRE is therefore to model the joint distribution P(CoT,P opt |I,P user ). Let x = (I,P user ) denote the input and y = [CoT;P opt ] the concatenated output sequence. The generation process follows a standard autoregressive factorization: P(y | x) = |y| Y t=1 P(y t | y <t ,x). This formulation encourages the model to first articulate the reasoning re- quired to interpret the visual scene and constraints, and then produce a concise executable prompt P opt . 3.2 Stage 1: Supervised Fine-Tuning with Reasoning Data The first stage initializes the VLM with the structural capability to follow I2V formatting conventions and execute rigorous reasoning. We construct a high- quality dataset of approximately 20K tuples I,P user , CoT,P opt through an automated rejection sampling paradigm using advanced proprietary APIs. As illustrated in the upper part of Figure 2, the data construction process follows four steps: Step1: Data Initialization. For each given pair (I,P user ), we computa- tionally define the rigorous ground truth (G truth , e.g., the expected last-frame state or progress goal). Step2: CoT and Prompt Generation. The API analyzes the initial im- age, clarifies the user’s intent, and identifies critical constraints to produce a natural-language CoT. It then generates a concise promptP opt suitable for video generation. Step3: Rejection Sampling. To ensure the reliability of generated samples, we introduce an independent evaluator to assess the quality of each candidate P opt . The evaluator has access to the input pair (I,P user ) as well as the corre- sponding ground truth G truth , and evaluates the output from two perspectives. First, it verifies visual consistency, checking whether the prompt introduces hal- lucinated objects, incorrect spatial relations, or directional misunderstandings with respect to the image. Second, it assesses whether the optimized prompt logically leads to the target state defined byG truth . The evaluator then assigns a comprehensive score on a 1–5 scale. Only samples with scores ≥ 4 are retained together with their corresponding CoT. This rejection sampling strategy ensures that the reasoning process and the resulting prompt remain logically coherent and consistent. Step4: Difficulty Patching. For exceptionally hard task categories where the API fails to deduce a valid logical chain under zero-shot conditions, we relax the constraint by providing G truth as a hint, allowing it to generate critical Think-with-Video Reasoning for Video Generative Models9 logical patches for the CoT. The patched outputs are then re-evaluated through the same rejection sampling pipeline in Step (3) to maintain quality control. These curated reasoning pairs are mixed with general SFT data—to pre- serve the base conversational capabilities of the VLM—and used to fine-tune a pre-trained Qwen3-VL-8B [1] model. The optimization minimizes the standard negative log-likelihood loss: L SFT (θ) =−E D SFT [logP θ (CoT,P opt |I,P user )].(1) 3.3 Stage 2: Policy Alignment via GRPO While SFT provides a strong logical initialization, performance on complex, long- horizon reasoning tasks remains suboptimal. As illustrated in the bottom part of Figure 2, to further unlock reasoning potential, we apply GRPO using 1.5K specialized prompts collected across 10 challenging tasks that exhibit the highest error rates after the SFT stage. A key design decision is our adoption of a purely text-based reward system in lieu of video-generation-based rewards. This choice is motivated by three con- siderations: (1) Stability and Efficiency: Current video models exhibit sub- stantial aesthetic variance, where poor rendering can easily suppress the reward signal of a logically correct prompt. Text-based evaluation avoids this instability and the considerable computational cost of rendering videos during RL training. (2) Generator Agnosticism: By decoupling from any specific video generator, Vid-PRE acquires no model-specific priors, functioning as a universal plug-and- play module. (3) Task Nature: I2V reasoning tasks are tightly bounded by the initial frame and deterministic physical/geometric constraints; thus, optimizing the textual prompt’s semantic correctness directly addresses the root cause of reasoning failures. During training, for each query x = (I,P user ,G truth ), the policy π θ samples a group of G outputso 1 ,...,o G , where o k = (CoT k ,P opt,k ). Our reward function evaluates each rollout along three dimensions on a 1–5 scale. Intent Preser- vation (r intent ) verifies thatP opt,k reasonably extends and refinesP user without altering the core objective. It ensures key objects and goals are preserved, while minor wording differences (e.g., camera terminology) are not overly penalized. Visual-Spatial Consistency (r image ) checks whether the spatial relations de- scribed in P opt,k (e.g., “bottom right”) accurately reflect the actual content of the image I, strictly penalizing hallucinated visual elements or invalid spatial relationships. Solution Correctness (r correct ) evaluates whether the explicit steps inferred in the prompt correctly solve the task and reach G truth , empha- sizing critical problem-solving details (e.g., specific paths or action steps) over generic generation constraints. In addition, a format compliance score (r format ) ensures outputs adhere to the prescribed structural template. The final reward is a weighted combination: r k = w 1 r intent + w 2 r image + w 3 r correct + w 4 r format . GRPO optimizes the policy by maximizing the advantage of outputs with higher relative rewards within the group [9,10,39]. The advantage for each roll- out is computed as the group-normalized reward: ˆ A k = (r k − μ G )/σ G , where 10M. Luo et al. μ G and σ G are the mean and standard deviation of rewards within the group r 1 ,...,r G . Let π ref denote the SFT-initialized policy serving as the fixed ref- erence for KL regularization: J (θ) =E x∼D RL , o∼π θ old " 1 G G X k=1 π θ (o k | x) π θ old (o k | x) ˆ A k − βD KL (π θ ∥π ref ) # . (2) Through this paradigm, Vid-PRE learns to produce logically precise and physically grounded reasoning chains that translate into robust instructions for downstream video generation. 4 Experiments 4.1 Settings We evaluate the general reasoning capability of six video generation models on VWG-Bench, including one open-weight model (Wan2.2) and five commercial systems (Wan2.5, Wan2.6-Flash, Kling-2.5-Turbo-Pro, Sora2, and Veo3.1). For each (I,P user ) pair, we sample three generations and report the averaged score for each evaluation dimension. During evaluation, we extract the last frame of each generated video to assess G last , while the remaining metrics are evaluated using a frame sampling configuration of 2 fps with a maximum of 16 frames. We initialize Vid-PRE with the pre-trained Qwen3-VL-8B [1] model, and use Gemini-3.0-Pro as the data generator and the unified evaluator. During the SFT stage, the model is trained on approximately 20K high-quality samples for 3 epochs with a batch size of 128. The maximum image token length is constrained to 1,024 to balance fine-grained visual detail and computational efficiency. In the subsequent GRPO stage, the policy is further optimized using 1.5K high-difficulty samples for approximately 3K optimization steps under the text-based reward system described earlier. All training is conducted on 8 H100 GPUs using the DeepSpeed ZeRO-3 optimization framework. During evaluation, we apply our prompt rewriting method only to models that are either open-weight (Wan2.2) or allow disabling their built-in Prompt Extension module (Wan2.5 and Wan2.6-Flash). For the latter two models, we ad- ditionally report results with their native Prompt Extension (PE) enabled to pro- vide a direct comparison. Beyond VWG-Bench, we further evaluate Vid-PRE on two additional high-quality open benchmarks. The first is MME-CoF, which con- tains 59 high-quality (Image, Prompt) pairs and evaluates generated videos from five aspects: instruction alignment, temporal consistency, visual stability, content fidelity, and focus relevance. The second benchmark is V-ReasonBench, which consists of 12 task categories with a total of 327 (Image, Prompt, Ground Truth) triplets. Each category is designed to be last-frame verifiable, meaning that task success can be determined solely from the final frame (e.g., Sudoku or arithmetic reasoning tasks). The benchmark groups the 12 categories into four reasoning dimensions: Structured Problem-Solving, Spatial Cognition, Pattern-based In- ference, and Physical Dynamics. We report the mean score over five random seeds for each dimension. Think-with-Video Reasoning for Video Generative Models11 Table 2: Baseline diagnostic on VWG-Bench. All models are evaluated without prompt extension. Commercial models are included to contextualize task difficulty. ModelVideo Quality Prog. Consist. Rule Follow. Prog. Goal Last Goal Overall Wan2.22.322.331.562.451.982.13 Wan2.52.702.863.142.482.562.75 Wan2.6-Flash2.732.923.232.502.622.80 Kling-2.5-Turbo-Pro2.953.123.382.622.782.97 Sora23.453.683.793.322.983.44 Veo3.13.543.783.953.453.053.55 4.2 Main Results Benchmarking Current Models on VWG-Bench. We first establish a baseline diagnostic using VWG-Bench (Table 2). Even the strongest commercial model (Veo3.1) achieves only 3.55 overall on a 1–5 scale, underscoring the challenge of our benchmark. A consistent pattern emerges across all models: video quality scores are substantially higher than reasoning-intensive metrics. For example, Wan2.2 attains 2.32 in video quality but only 1.56 in implicit rule following, and even Sora2 drops from 3.45 to 2.98 between video quality and last-frame goal accuracy. This confirms our central hypothesis: current models excel at visual rendering but struggle to maintain physical constraints, logical rules, and temporal causality without explicit textual guidance. Enhancing Reasoning with Vid-PRE. We evaluate Vid-PRE on two exter- nal benchmarks—MME-CoF and V-ReasonBench—as well as our VWG-Bench (Table 3). Integrating Vid-PRE yields substantial improvements across all mod- els without any architectural modifications to the video generators. On MME- CoF (Table 3a), Vid-PRE boosts Wan2.2’s overall score from 1.72 to 2.58, a 50% gain, with instruction alignment and content fidelity showing the largest abso- lute gains. It also outperforms Wan PE, raising Wan2.5 from 2.84 to 3.07 and Wan2.6-Flash from 2.82 to 3.09. On V-ReasonBench, the advantage is even more pronounced (Table 3b), whose tasks demand fine-grained physical and spatial reasoning. Vid-PRE improves Wan2.2’s average accuracy from 26.65 to 44.48, a 67% gain, with spatial cognition increasing from 1.32 to 36.25. Against Wan PE, Vid-PRE delivers an additional +25% gain on Wan2.5 and +33% on Wan2.6- Flash, confirming that generic prompt rewriting is no substitute for structured, constraint-aware reasoning. On VWG-Bench (Table 3c), Vid-PRE raises Wan2.2 from 2.13 to 2.51, an 18% gain, and improves its weakest baseline dimension, implicit rule following, from 1.56 to 2.12. For Wan2.6-Flash, Vid-PRE achieves 3.26, surpassing the Wan PE baseline of 3.04 and narrowing the gap to commer- cial systems like Sora2 (3.46) and Veo3.1 (3.55). 4.3 Analysis and Discussion Evaluator Reliability. We validate our VLM-as-Judge protocol by comparing automatic scores with human judgments. Specifically, we sample 100 prompts across all nine VWG-Bench dimensions and evaluate 200 videos generated by 12M. Luo et al. Table 3: Performance of Vid-PRE across reasoning benchmarks. Wan PE denotes the model’s built-in prompt extension. Vid-PRE consistently outperforms both the base models and their native PE across all metrics. (a) MME-CoF. ModelPrompt Extension Instr. Align. Temp. Consist. Vis. Stability Content Fidelity Focus Relevance Overall Score Wan2.2 None0.692.362.631.161.771.72 Vid-PRE1.652.953.112.382.822.58 Wan2.5 None1.572.853.342.452.872.62 Wan PE1.982.883.412.593.352.84 Vid-PRE2.343.283.592.693.473.07 Wan2.6-Flash None1.673.053.392.492.852.69 Wan PE2.002.853.342.613.312.82 Vid-PRE2.253.263.622.863.433.09 (b) V-ReasonBench. ModelPrompt Extension Struct. Solving Spatial Cognition Pattern Inference Physical Dynamics Average Wan2.2 None29.271.3245.1230.9026.65 Vid-PRE33.9036.2560.4947.2944.48 Wan2.5 None25.9520.0546.3033.0531.34 Wan PE35.5538.2552.3038.1541.06 Vid-PRE44.0542.3566.1552.2051.19 Wan2.6-Flash None26.8520.7147.1033.7432.10 Wan PE36.5729.1853.2339.0239.50 Vid-PRE45.4643.7867.3553.6852.57 (c) VWG-Bench. ModelPrompt Extension Video Quality Prog. Consist. Rule Follow. Prog. Goal Last Goal Overall Wan2.2 None2.322.331.562.451.982.13 Vid-PRE2.652.682.122.752.352.51 Wan2.5 None2.702.863.142.482.562.75 Wan PE3.013.183.352.662.742.99 Vid-PRE3.253.413.552.832.933.19 Wan2.6-Flash None2.732.923.232.502.622.80 Wan PE3.053.183.452.682.863.04 Vid-PRE3.283.493.622.883.023.26 Veo3.1 and Wan2.6-Flash. Three human experts independently score each video on the same 1–5 criteria used by the automatic evaluator. As shown in Table 4, Spearman’s ρ ranges from 0.69 to 0.80 and the mean absolute error remains below 0.55 across all criteria. This agreement indicates that the automatic scores track human-perceived reasoning quality rather than only superficial rendering preference, supporting the reliability of the benchmark-level model comparisons. Ablation Study. We ablate the contribution of each training stage on both external benchmarks using Wan2.2 and Wan2.6-Flash as base generators (Ta- ble 5). SFT alone already provides a significant lift over the baselines, showing that structured reasoning and constraint-aware prompt generation are effective even without reinforcement learning. The subsequent GRPO stage consistently pushes performance further, as our text-based reward system explicitly penal- izes spatial hallucinations and logical misalignments. The same trend holds for Wan2.6-Flash. These results confirm that the two stages are complementary: Think-with-Video Reasoning for Video Generative Models13 Table 4: Agreement between human experts and Gemini-3.0-Pro on the VWG-Bench evaluation criteria. CriterionHuman Geminiρ↑ MAE↓ Video Quality3.203.440.69 0.54 Progress Consistency3.423.350.78 0.42 Rule Following3.473.590.73 0.48 Progress Goal3.062.980.75 0.41 Last-Frame Goal2.932.840.80 0.39 SFT provides the structural foundation, and GRPO refines the policy toward tighter constraint satisfaction. Table 5: Ablation on training stages. ModelVariantMME-CoF V-ReasonBench Wan2.2 Base1.7226.65 + Vid-PRE SFT 2.4440.26 + Vid-PRE SFT+RL 2.5844.48 Wan2.6-Flash Base2.6932.10 + Vid-PRE SFT 2.9745.96 + Vid-PRE SFT+RL 3.0952.57 Qualitative Analysis. Figure 3 presents a qualitative comparison of video generation with and without Vid-PRE across diverse reasoning tasks. As illus- trated, base video models fundamentally lack the capacity for zero-shot logi- cal deduction. When given standard prompts, even those explicitly containing strict negative constraints, the base models suffer from severe reasoning fail- ures and visual degradation. For instance, they hallucinate unrelated objects in visual analogies. By contrast, Vid-PRE dramatically enhances both reasoning capability and generation quality by transforming abstract logical problems into explicit, renderable visual instructions. It first employs a CoT to autonomously deduce the correct logical solutions and then rewrites the prompt to provide pre- cise, step-by-step visual guidance. Consequently, Vid-PRE ensures strict physical constraint compliance, prevents visual hallucinations, and successfully executes complex think-with-video tasks. 5 Related Work Image-to-Video Generation. I2V generation has progressed rapidly, driven primarily by diffusion and transformer architectures. Since Stable Video Diffu- sion [2] established the paradigm of image-conditioned generation, subsequent works have scaled this approach considerably. Open-source frameworks such as 14M. Luo et al. Fig. 3: Qualitative comparison of video generation with and without Vid-PRE. While generic constraints (red) fail to prevent logical errors and visual hallucinations in base models, Vid-PRE translates explicit reasoning into precise visual instructions (green), significantly improving both reasoning accuracy and video quality. CogVideoX [41], Wan [35], and HunyuanVideo [20] have democratized high- resolution synthesis through advanced expert transformers and 3D causal VAEs, while commercial systems like Sora [31], Veo [14], Kling [21], and Seedance [12] have achieved photorealistic fidelity and remarkable temporal consistency. De- spite these advances in rendering visual dynamics, their capacity for structured, rule-bound reasoning remains fundamentally underexplored. Prompt Rewriting and Optimization. The distributional gap between concise user inputs and the detailed captions used during model training makes prompt rewriting critical for video generation. Early approaches bridge this gap via zero-shot in-context learning using LLMs (e.g., GLM-4 in CogVideoX [41] or GPT-4o in Open-Sora [43]). To mitigate hallucination and intent drift inherent in zero-shot rewriting, recent methods have shifted toward reward-guided optimiza- tion, employing techniques such as Direct Preference Optimization (Prompt-A- Video [19]), multi-level feedback (VPO [5]), and retrieval-augmented refinement (RAPO [11]). However, these techniques predominantly focus on enhancing aes- thetic quality and generic text-video alignment. In contrast, Vid-PRE specifically targets reasoning enhancement by offloading the cognitive burden to a VLM and training with purely text-based reinforcement learning signals. Benchmarks for I2V Models. Existing evaluation frameworks broadly fall into two categories. The first focuses on perceptual quality and general gen- Think-with-Video Reasoning for Video Generative Models15 eration capabilities: benchmarks like VBench/VBench++ [17, 18] and World- Score [8] assess visual fidelity and scene dynamics but do not probe higher- order logical reasoning. The second category, motivated by recent demonstra- tions of reasoning in video generation [38], evaluates the cognitive limits of video models. Several broad benchmarks have emerged (e.g., MME-CoF [15], Video- ThinkBench [34], TiViBench [4]), alongside specialized ones targeting procedu- ral pathfinding (VR-Bench [40]), strict rule execution (RULER-Bench [16]), and implicit physical and visual rules (e.g., V-ReasonBench [29], RISE-Video [25], MMGR [3]), as shown in Table 1. Despite advancing the evaluation landscape, these benchmarks often isolate specific cognitive skills or rely on holistic, coarse- grained scoring. VWG-Bench addresses these limitations with a comprehensive taxonomy spanning 9 dimensions and 38 tasks, equipped with a fine-grained three-level VLM-as-Judge framework. 6 Conclusion In this work, we introduced VWG-Bench, a comprehensive evaluation framework designed to rigorously assess the think-with-video reasoning capabilities of mod- ern video generative models. Through our three-level VLM-as-Judge evaluation, we revealed a persistent gap: current models excel at visual rendering but strug- gle with logic-heavy, rule-constrained, and long-horizon tasks. To bridge this gap, we proposed Vid-PRE, a model-agnostic prompt enhancer that offloads spatio- temporal reasoning to a dedicated VLM and translates raw instructions into constraint-aware, executable prompts. Trained via SFT and GRPO with purely text-based rewards, Vid-PRE yields substantial and consistent improvements in reasoning accuracy across multiple state-of-the-art generators. References 1. Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025) 2. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) 3. Cai, Z., Qiu, H., Ma, T., Zhao, H., Zhou, G., Huang, K.H., Kordjamshidi, P., Zhang, M., Xiao, W., Gu, J., et al.: Mmgr: Multi-modal generative reasoning. arXiv preprint arXiv:2512.14691 (2025) 4. Chen, H.H., Lan, D., Shu, W.J., Liu, Q., Wang, Z., Chen, S., Cheng, W., Chen, K., Zhang, H., Zhang, Z., et al.: Tivibench: Benchmarking think-in-video reasoning for video generative models. arXiv preprint arXiv:2511.13704 (2025) 5. Cheng, J., Lyu, R., Gu, X., Liu, X., Xu, J., Lu, Y., Teng, J., Yang, Z., Dong, Y., Tang, J., et al.: Vpo: Aligning text-to-video generation models with prompt opti- mization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. p. 15636–15645 (2025) 16M. Luo et al. 6. Cho, J., Puspitasari, F.D., Zheng, S., Zheng, J., Lee, L.H., Kim, T.H., Hong, C.S., Zhang, C.: Sora as an agi world model? a complete survey on text-to-video gener- ation. arXiv preprint arXiv:2403.05131 1(4) (2024) 7. Ding, J., Zhang, Y., Shang, Y., Zhang, Y., Zong, Z., Feng, J., Yuan, Y., Su, H., Li, N., Sukiennik, N., et al.: Understanding world or predicting future? a compre- hensive survey of world models. ACM Computing Surveys 58(3), 1–38 (2025) 8. Duan, H., Yu, H.X., Chen, S., Fei-Fei, L., Wu, J.: Worldscore: A unified evaluation benchmark for world generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. p. 27713–27724 (2025) 9. Fang, X., Ma, L., Chen, Z., Zhou, M., Qi, G.j.: Inflvg: Reinforce inference-time consistent long video generation with grpo. arXiv preprint arXiv:2505.17574 (2025) 10. Feng, K., Gong, K., Li, B., Guo, Z., Wang, Y., Peng, T., Wu, J., Zhang, X., Wang, B., Yue, X.: Video-r1: Reinforcing video reasoning in mllms. arXiv preprint arXiv:2503.21776 (2025) 11. Gao, B., Gao, X., Wu, X., Zhou, Y., Qiao, Y., Niu, L., Chen, X., Wang, Y.: The devil is in the prompts: Retrieval-augmented prompt optimization for text-to-video generation. In: Proceedings of the Computer Vision and Pattern Recognition Con- ference. p. 3173–3183 (2025) 12. Gao, Y., Guo, H., Hoang, T., Huang, W., Jiang, L., Kong, F., Li, H., Li, J., Li, L., Li, X., et al.: Seedance 1.0: Exploring the boundaries of video generation models. arXiv preprint arXiv:2506.09113 (2025) 13. Ghazanfari, S., Croce, F., Flammarion, N., Krishnamurthy, P., Khorrami, F., Garg, S.: Chain-of-frames: Advancing video understanding in multimodal llms via frame- aware reasoning. arXiv preprint arXiv:2506.00318 (2025) 14. Google DeepMind: Veo-3 technical report. Tech. rep., Google DeepMind (May 2025), https://storage.googleapis.com/deepmind-media/veo/Veo-3-Tech- Report.pdf, accessed 25 June 2026 15. Guo, Z., Chen, X., Zhang, R., An, R., Qi, Y., Jiang, D., Li, X., Zhang, M., Li, H., 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) 16. He, X., Fan, Z., Li, H., Zhuo, F., Xu, H., Cheng, S., Weng, D., Liu, H., Ye, C., Wu, B.: Ruler-bench: Probing rule-based reasoning abilities of next-level video gener- ation models for vision foundation intelligence. arXiv preprint arXiv:2512.02622 (2025) 17. Huang, Z., He, Y., Yu, J., Zhang, F., Si, C., Jiang, Y., Zhang, Y., Wu, T., Jin, Q., Chanpaisit, N., et al.: Vbench: Comprehensive benchmark suite for video gener- ative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 21807–21818 (2024) 18. Huang, Z., Zhang, F., Xu, X., He, Y., Yu, J., Dong, Z., Ma, Q., Chanpaisit, N., Si, C., Jiang, Y., et al.: Vbench++: Comprehensive and versatile benchmark suite for video generative models. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025) 19. Ji, Y., Zhang, J., Wu, J., Zhang, S., Chen, S., Ge, C., Sun, P., Chen, W., Shao, W., Xiao, X., et al.: Prompt-a-video: Prompt your video diffusion model via preference- aligned llm. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision. p. 18725–18735 (2025) 20. Kong, W., Tian, Q., Zhang, Z., Min, R., Dai, Z., Zhou, J., Xiong, J., Li, X., Wu, B., Zhang, J., et al.: Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603 (2024) 21. Kuaishou Technology: Kling ai: Next-generation ai creative studio. https:// klingai.com/ (June 2024), accessed 25 June 2026 Think-with-Video Reasoning for Video Generative Models17 22. LeCun, Y., et al.: A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review 62(1), 1–62 (2022) 23. Li, M., Zhong, J., Zhao, S., Zhang, H., Lin, S., Lai, Y., Wei, C., Psounis, K., Zhang, K.: Tir-bench: A comprehensive benchmark for agentic thinking-with-images rea- soning. arXiv preprint arXiv:2511.01833 (2025) 24. Lin, M., Wang, X., Wang, Y., Wang, S., Dai, F., Ding, P., Wang, C., Zuo, Z., Sang, N., Huang, S., et al.: Exploring the evolution of physics cognition in video generation: A survey. arXiv preprint arXiv:2503.21765 (2025) 25. Liu, M., Ma, S., Meng, S., Zhao, X., Zhang, Z., Zhang, S., Zhong, Z., Chen, P., Cao, H., Sun, X., et al.: Rise-video: Can video generators decode implicit world rules? arXiv preprint arXiv:2602.05986 (2026) 26. Liu, X., Xu, Z., Li, M., Wang, K., Lee, Y.J., Shang, Y.: Can world simula- tors reason? gen-vire: A generative visual reasoning benchmark. arXiv preprint arXiv:2511.13853 (2025) 27. Luo, M., Li, B., Xu, S., Zhang, S., Chen, Q., Han, M., Chen, W., Huang, Y., Fei, H., Lee, M.L., et al.: Unveiling the cognitive compass: Theory-of-mind-guided multimodal emotion reasoning. arXiv preprint arXiv:2602.00971 (2026) 28. Luo, M., Wu, S., Jing, L., Ju, T., Zheng, L., Lai, J., Wu, T., Du, X., Li, J., Yan, S., et al.: Dr. v: A hierarchical perception-temporal-cognition framework to diag- nose video hallucination by fine-grained spatial-temporal grounding. International Journal of Computer Vision 134(6), 278 (2026) 29. Luo, Y., Zhao, X., Lin, B., Zhu, L., Tang, L., Liu, Y., Chen, Y.C., Qian, S., Wang, X., You, Y.: V-reasonbench: Toward unified reasoning benchmark suite for video generation models. arXiv preprint arXiv:2511.16668 (2025) 30. Lyu, Q., Havaldar, S., Stein, A., Zhang, L., Rao, D., Wong, E., Apidianaki, M., Callison-Burch, C.: Faithful chain-of-thought reasoning. In: Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Con- ference of the Asia-Pacific Chapter of the Association for Computational Linguis- tics (Volume 1: Long Papers). p. 305–329 (2023) 31. OpenAI: Sora 2 system card. Tech. rep., OpenAI (September 2025), https://cdn. openai.com/pdf/50d5973c-c4f-4c2d-986f-c72b5d0f069/sora_2_system_ card.pdf, accessed 25 June 2026 32. Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024) 33. Su, Z., Xia, P., Guo, H., Liu, Z., Ma, Y., Qu, X., Liu, J., Li, Y., Zeng, K., Yang, Z., et al.: Thinking with images for multimodal reasoning: Foundations, methods, and future frontiers. arXiv preprint arXiv:2506.23918 (2025) 34. Tong, J., Mou, Y., Li, H., Li, M., Yang, Y., Zhang, M., Chen, Q., Liang, T., Hu, X., Zheng, Y., et al.: Thinking with video: Video generation as a promising multimodal reasoning paradigm. arXiv preprint arXiv:2511.04570 (2025) 35. Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., et al.: Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314 (2025) 36. Wang, Y., Liu, X., Pang, W., Ma, L., Yuan, S., Debevec, P., Yu, N.: Survey of video diffusion models: Foundations, implementations, and applications. arXiv preprint arXiv:2504.16081 (2025) 37. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35, 24824–24837 (2022) 18M. Luo et al. 38. Wiedemer, T., Li, Y., Vicol, P., Gu, S.S., Matarese, N., Swersky, K., Kim, B., Jaini, P., Geirhos, R.: Video models are zero-shot learners and reasoners. arXiv preprint arXiv:2509.20328 (2025) 39. Xue, Z., Wu, J., Gao, Y., Kong, F., Zhu, L., Chen, M., Liu, Z., Liu, W., Guo, Q., Huang, W., et al.: Dancegrpo: Unleashing grpo on visual generation. arXiv preprint arXiv:2505.07818 (2025) 40. Yang, C., Wan, H., Peng, Y., Cheng, X., Yu, Z., Zhang, J., Yu, J., Yu, X., Zheng, X., Zhou, D., et al.: Reasoning via video: The first evaluation of video models’ reasoning abilities through maze-solving tasks. arXiv preprint arXiv:2511.15065 (2025) 41. 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) 42. Zhang, X., Du, C., Pang, T., Liu, Q., Gao, W., Lin, M.: Chain of preference op- timization: Improving chain-of-thought reasoning in llms. Advances in Neural In- formation Processing Systems 37, 333–356 (2024) 43. Zheng, Z., Peng, X., Yang, T., Shen, C., Li, S., Liu, H., Zhou, Y., Li, T., You, Y.: Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404 (2024) 44. Zheng, Z., Yang, M., Hong, J., Zhao, C., Xu, G., Yang, L., Shen, C., Yu, X.: Deepeyes: Incentivizing" thinking with images" via reinforcement learning. arXiv preprint arXiv:2505.14362 (2025) 45. Zhou, S., Vilesov, A., He, X., Wan, Z., Zhang, S., Nagachandra, A., Chang, D., Chen, D., Wang, X.E., Kadambi, A.: Vlm4d: Towards spatiotemporal awareness in vision language models. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 8600–8612 (2025) Supplementary Material for “From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models” Meng Luo et al. A Additional Evaluation Details A.1 Physics Generalization and Per-Dimension Analysis Table 1 reports the physics-focused results, and Table 2 provides the complete dimension-level breakdown. Table 1: Vid-PRE on physics-related reasoning tasks. ModelPhysics (MME-CoF) Physical Dynamics (V-RB) Wan2.22.2530.90 Wan2.2 with Vid-PRE3.31 (+47%)47.29 (+53%) Wan2.6-Flash3.2533.74 Wan2.6-Flash with Vid-PRE3.45 (+6%)53.68 (+59%) Table 2: Per-dimension breakdown with Vid-PRE on VWG-Bench. Values in paren- theses are relative gains over the corresponding base generator. “–” indicates that a dimension has no rule-following annotation. DimensionVQPCRFPGLG Symbolic2.91 (+9.7%) 2.36 (+13.5%)–2.94 (+17.6%) 2.93 (+16.7%) Graphical3.43 (+14.3%) 3.35 (+9.5%) 4.18 (+11.5%) 4.39 (+15.5%) 3.05 (+16.4%) Embodied3.96 (+4.2%) 4.73 (+2.8%) 4.87 (+3.6%) 4.49 (+5.4%) 3.86 (+5.5%) GUI3.32 (+14.5%) 3.29 (+15.4%)–4.25 (+16.4%) 3.02 (+18.4%) Counterfactual 4.32 (+3.4%) 4.84 (+2.7%)–2.08 (+6.4%) 3.54 (+4.8%) Spatial3.86 (+5.1%) 4.15 (+5.2%) 1.78 (+9.5%) 3.89 (+6.1%) 2.72 (+7.0%) Temporal4.06 (+4.2%) 4.31 (+5.0%)–4.32 (+5.3%) 4.05 (+4.9%) Physical3.87 (+5.3%) 4.58 (+5.9%)–3.36 (+5.4%) 3.06 (+5.0%) Social3.74 (+3.6%) 4.32 (+5.0%)–4.32 (+6.0%) 3.83 (+5.9%) A.2 Additional Commercial Model and Reward Sensitivity Table 3 compares Seedance 2.0 with Veo3.1. 2M. Luo et al. Table 3: Seedance 2.0 and Veo3.1 on VWG-Bench. ModelVQ PC RF PG LG Overall Veo3.13.54 3.78 3.95 3.45 3.05 3.55 Seedance2.0 3.67 3.52 3.87 3.38 3.17 3.52 GRPO uses four reward components; format compliance is always active to ensure output parsability. Table 4 ablates the remaining three components. Solution correctness is the main source of improvement, while intent preservation and image consistency provide complementary gains. Table 4: Reward-component ablation using Wan2.2. r intent r image r correct MME-CoF V-ReasonBench ✗2.4440.26 ✗✓✗2.4037.85 ✓✗2.4239.10 ✓✗2.4138.45 ✗✓2.5343.35 ✓✗✓2.5644.05 ✗✓2.5543.75 ✓2.5844.48 A.3 Inference Cost Vid-PRE prompt rewriting takes approximately 3–5 seconds per sample, whereas video generation typically takes more than 60 seconds in the evaluated I2V pipelines. The resulting overhead is below 8% of end-to-end latency. B Task Descriptions Task 1: Arithmetic Definition: Given a set of arithmetic expressions, the model is required to compute their results and write them in the corresponding positions. Task 2: Bubble Sorting Definition: Given a set of bubbles labeled with numbers, the model is required to make them rise one by one in ascending or descending order according to the numbers on the bubbles. Supplementary Material for Think-with-Video Reasoning3 Fig. 1: First-frame input used for I2V generation. Prompt: “Watch the equations appearing in this video. What is the value of x in the last equation?” Fig. 2: First-frame input used for I2V generation. Prompt: “Make the bubbles float out of the frame in ascending order according to their corresponding numbers.” 4M. Luo et al. Task 3: Coding Definition: Given a coding problem and an input, the model is required to compute the output and write it in the corresponding position. Fig. 3: First-frame input used for I2V generation. Prompt: “Write the correct answer after Öutput ̈ .” Task 4: Sudoku Definition: Given an incomplete Sudoku puzzle, the model is required to fill in the puzzle correctly. Fig. 4: First-frame input used for I2V generation. Prompt: “Fill in this Sudoku puz- zle.” Supplementary Material for Think-with-Video Reasoning5 Task 5: Tic-Tac-Toe Definition: Given a Tic-Tac-Toe board and several pieces, the model is required, as either “X” or “O”, to make the next move that can either secure an immediate win or block the opponent’s immediate win. Fig. 5: First-frame input used for I2V generation. Prompt: “Given the current state of the Tic-Tac-Toe board in the image, predict the next move that results in a win for the player X. Output the position of the winning move.” Task 6: Box Matching Definition: Given a set of objects and several boxes marked with certain at- tributes, the model is required to place each object into the corresponding box according to its attribute. Fig. 6: First-frame input used for I2V generation. Prompt: “Place the objects into the corresponding boxes according to the categories labeled on the boxes.” 6M. Luo et al. Task 7: Color Connecting Definition: Given several colored circles, the model is required to draw lines connecting circles of the same color. Fig. 7: First-frame input used for I2V generation. Prompt: “Connect circles of the same color with lines, and ensure that the lines have no intersections or overlaps.” Task 8: Sequence Completion Definition: Given a sequence of continuous images and several blank slots, the model is required to infer the underlying pattern from the image sequence and fill the blank slots with the correct images. Fig. 8: First-frame input used for I2V generation. Prompt: “Finish the incomplete frames of this shape-sequence puzzle. Do not change given patterns. Each row contains equal-sized square cells outlined in black on white. Shapes are solid black arrows, dots, or stripes that obey a clear mathematical rule (rotation, translation, scaling, or toggling). Produce the missing panels so the sequence remains consistent and has a single logical solution.” Supplementary Material for Think-with-Video Reasoning7 Task 9: Shape Fitting Definition: Given several 3D shapes and a board with multiple shape-shaped slots, the model is required to place each 3D shape into the corresponding slot based on its shape. Fig. 9: First-frame input used for I2V generation. Prompt: “The scene shows some colored pieces, and a panel with some holes. Each colored piece fits into one and only one hole. A hand grabs each colored piece and puts it into an empty hole that has the exact same shape.” Task 10: Visual Analogy Definition: Given a row of patterns and the patterns after a rule-based trans- formation, the model is required to apply the same transformation rule to a new row of patterns and produce the result. Fig. 10: First-frame input used for I2V generation. Prompt: “Create a smooth ani- mation to generate the missing object in the lower right region and solve the visual analogy. The original three objects must remain still. Static shot, no zoom no pan no dolly.” 8M. Luo et al. Task 11: Embodied Action Execution Definition: Given an embodied scene and an action instruction, the model is required to control a robotic arm or a human hand to execute the action. Fig. 11: First-frame input used for I2V generation. Prompt: “From a first-person perspective, execute this instruction with human/robotic arm(s): closing the rightest cabinet door.” Task 12: Embodied Trajectory Reasoning Definition: Given an embodied scene, an action instruction, and several can- didate trajectories, the model is required to control a robotic arm or a human hand to execute the action along one of the candidate trajectories. Fig. 12: First-frame input used for I2V generation. Prompt: “From a first-person perspective, execute this instruction with human/robotic arm(s): drag the cloth to the middle of a spoon and a mushroom. There are four possible trajectories (T1 T4) shown in the image; please operate by following the trajectory you consider to be the most correct.” Supplementary Material for Think-with-Video Reasoning9 Task 13: Implicit Goal Reasoning Definition: Given an embodied scene and an implicit action goal, the model is required to infer and execute the true action instruction. Fig. 13: First-frame input used for I2V generation. Prompt: “From a first-person perspective, how do I operate the robotic arm to complete the instruction: ̈ help me clean up the tissues on the table ̈ ?” Task 14: Intention Causality Definition: Given a scene and several visual cues, the model is required to infer the behavioral goal from the visual cues and execute it. Fig. 14: First-frame input used for I2V generation. Prompt: “Inserting the key into the keyhole.” 10M. Luo et al. Task 15: Procedural Knowledge Definition: Given a scene and a complex action instruction, the model is re- quired to understand, decompose, and correctly execute the instruction. Fig. 15: First-frame input used for I2V generation. Prompt: “Folding the letter, in- serting it into the envelope, and sealing it.” Task 16: GUI Instruction Execution Definition: Given a GUI interface and an instruction, the model is required to perform the operation and correctly present the subsequent changes in the interface. Fig. 16: First-frame input used for I2V generation. Prompt: “Execute this instruction: Open the ParEval Learderboard.” Supplementary Material for Think-with-Video Reasoning11 Task 17: GUI Click Reasoning Definition: Given a GUI interface and a specific location, the model is required to execute a click at that location and plausibly present the subsequent changes in the interface. Fig. 17: First-frame input used for I2V generation. Prompt: “Click Ä prominent play button labeled ’Latest Episode’ displayed in the middle section of the podcast page ̈ .” Task 18: Physics Inversion Definition: Given a standard physical scene, the model is required to simulate a counterintuitive physical evolution process in which fundamental physical laws are reversed. Fig. 18: First-frame input used for I2V generation. Prompt: “The solid granite rocks gently lift off the table and float upwards into the sky.” 12M. Luo et al. Task 19: Causal Intervention Definition: Given a causal sequence, the model is required to alter one key in- tervention variable, infer the resulting downstream chain reactions, and generate the subsequent outcomes. Fig. 19: First-frame input used for I2V generation. Prompt: “The hammer hits the book hard. The book must remain completely still for several seconds after being hit. After the delay, the book suddenly flies away violently.” Task 20: Fictional Rule Execution Definition: Given a set of fictional operational rules, the model is required to drive objects through state transitions strictly according to those rules. Fig. 20: First-frame input used for I2V generation. Prompt: “The solid gold finger touches the wooden chair, transforming the entire chair into solid gold.” Supplementary Material for Think-with-Video Reasoning13 Task 21: Surreal Attribute Mapping Definition: Given objects with surreal properties, the model is required to sim- ulate their anomalous physical behavior under specific forces. Fig. 21: First-frame input used for I2V generation. Prompt: “The tissue paper is resting on the foam. Despite its light appearance, the leaf behaves as if it weighs 1000kg.” Task 22: Reverse Temporal Logic Definition: Given an object, the model is required to generate its physical evolution under reversed time. Fig. 22: First-frame input used for I2V generation. Prompt: “The ball rolling back- ward, accelerating, and jumping precisely back into the person’s hand.” 14M. Luo et al. Task 23: Maze Definition: Given a maze and a pair of start and end points, the model is required to move a block from the start point to the end point. Fig. 23: First-frame input used for I2V generation. Prompt: “Move the green cube along the maze to the red cube.” Task 24: Real World Spatial Reasoning Definition: Given a real-world scene, the model is required to complete instruc- tions or reasoning involving real-world objects in the scene. Fig. 24: First-frame input used for I2V generation. Prompt: “A blue arrow from the catcher in the red jersey toward the batter in the green jersey.” Supplementary Material for Think-with-Video Reasoning15 Task 25: Rotation Reasoning Definition: Given a rotated image, the model is required to rotate it back by a fixed angle and direction and use a bounding box to mark the corresponding object. Fig. 25: First-frame input used for I2V generation. Prompt: “Rotate the video frame 90 degrees clockwise in the 2D plane, then draw bounding boxes around donut.” Task 26: Spatial Geometry Definition: Given a 3D image, the model is required to apply the specified 3D transformation to the object in the image. Fig. 26: First-frame input used for I2V generation. Prompt: “The view switching smoothly around the still car.” 16M. Luo et al. Task 27: Short-term Temporal Dynamics Definition: Given a scene and several objects, the model is required to generate their state changes over the course of several seconds. Fig. 27: First-frame input used for I2V generation. Prompt: “The signal switching to green.” Task 28: Medium-term Evolutionary Processes Definition: Given a scene and several objects, the model is required to generate their state changes over timescales ranging from minutes to months. Fig. 28: First-frame input used for I2V generation. Prompt: “A time-lapse of the candle burning over an hour.” Supplementary Material for Think-with-Video Reasoning17 Task 29: Long-term Temporal Progression Definition: Given a scene and several objects, the model is required to generate their long-term evolution on a yearly timescale. Fig. 29: First-frame input used for I2V generation. Prompt: “A time-lapse over five years.” Task 30: Physics-based Reasoning Definition: Given a physical scene, the model is required to generate physics- related changes of objects in the scene, involving mechanics and dynamics, fluid dynamics, material deformation, optics, thermodynamics, and acoustics. Fig. 30: First-frame input used for I2V generation. Prompt: “The knife cutting through the block.” 18M. Luo et al. Task 31: Biological Kinematics Definition: Given a biological-entity scene, the model is required to generate its natural life motion while adhering to physiological structure and kinematic constraints. Fig. 31: First-frame input used for I2V generation. Prompt: “The plant growing to- wards light.” Task 32: Physical Commonsense Reasoning Definition: Given a scene, the model is required to generate physical changes governed by everyday physical common sense. Fig. 32: First-frame input used for I2V generation. Prompt: “The man exhales a warm breath onto the glass.” Supplementary Material for Think-with-Video Reasoning19 Task 33: Life & Healthcare Knowledge Definition: Given a medical scene, the model is required to generate changes based on biological responses and basic medical logic. Fig. 33: First-frame input used for I2V generation. Prompt: “The skin’s reaction as the intense exercise continues.” Task 34: Geography & Earth Science Definition: Given a scene, the model is required to generate changes involving macroscopic phenomena such as river formation, celestial systems, and weather evolution. Fig. 34: First-frame input used for I2V generation. Prompt: ‘Showing tectonic plates pushing against each other.” 20M. Luo et al. Task 35: Specialized Sports Kinematics Definition: Given a sports scene and several people, the model is required to make them perform sports actions with professional characteristics. Fig. 35: First-frame input used for I2V generation. Prompt: “The athlete jumps up- ward.” Task 36: Social Rules Definition: Given a social scene, the model is required to control the people in the scene or a first-person protagonist to carry out instructions grounded in social contracts, norms, and the implicit “script” of the specific scenario. Fig. 36: First-frame input used for I2V generation. Prompt: “The man should dispose of the empty bottle.” Supplementary Material for Think-with-Video Reasoning21 Task 37: Cultural Customs Definition: Given a real-world scene, the model is required to carry out in- structions involving specific cultural symbols, festival customs, and metaphorical abstract concepts. Fig. 37: First-frame input used for I2V generation. Prompt: “The couple greets the visitor.” Task 38: Emotion Recognition Definition: Given a person, the model is required to make them perform be- haviors involving emotional changes, micro-expressions, and body language. Fig. 38: First-frame input used for I2V generation. Prompt: “Showing their natural reaction as the dull lecture drags on.” 22M. Luo et al. C VWG-Bench Prompt Templates Last Frame Evaluation You are a professional engineer in the field of image-to-video generation. You will be given the last frame of a video generated by an image-to-video model for the task_name task, along with a pre-designed “Last frame goal” that describes what should appear in the last frame. Your task is to evaluate how consistent this last frame is with the given “Last frame goal”, assign a score from 1 to 5 (from least consistent to most consistent), and provide the reason for the score you assign. The “Last frame goal”: last_frame_goal Output your result in the following format: "last_frame_goal_score": "<your score here>", "reason_for_last_frame_goal_score": "<your reason here>" Video Evaluation You are a professional engineer in the field of image-to-video genera- tion. You will be given a video generated by an image-to-video model for the task_name task with the image-to-video prompt: “user_prompt”. Your task is to evaluate the following metrics based on the correspond- ing scoring criteria: video_quality, progress_consistency, (Optional) implicit_rule, and (Optional) progress_goal. First, evaluate the overall quality of the video, focusing only on the visual and temporal coherence of the frames themselves (not semantics, prompt adherence, or aesthetics). Assess stability and continuity, including but not limited to: abrupt visual shifts, jump cuts, frame drops or repeats, jitter or stutter, flicker (brightness, color, or texture), temporal inconsistency (identity, shape, or size drifting), warping or melting, ghosting or double edges, motion discontinuities, camera instability, rolling-shutter-like dis- tortions, aliasing or shimmering, compression or blocking artifacts, noise bursts, sudden blur changes, and background “breathing” or deformation. Penalize any brief but severe failures strongly. Assign a score from 1 to 5 (1 = unusable or very unstable with frequent severe artifacts; 2 = major artifacts that substantially disrupt viewing; 3 = noticeable artifacts but generally watchable; 4 = minor occasional artifacts with mostly stable frames; 5 = very stable, clean, and temporally coherent), and provide the reason for the score you assign. Second, evaluate the overall temporal consistency of the video, that is, whether objects or scenes in the video undergo changes that should not oc- cur. The specific rule is “foreground” (Optional), “background”. Based on how consistent the video is with this rule, assign a score from 1 to 5 Supplementary Material for Think-with-Video Reasoning23 (from least consistent to most consistent), and provide the reason for the score you assign. (Optional) Third, evaluate how well the video understands the task, that is, whether the video follows the implicit rules of the task. The implicit rule of this task is “implicit_rule”. Based on how consistent the video is with this implicit rule, assign a score from 1 to 5 (from least consistent to most consistent), and provide the reason for the score you assign. (Optional) Finally, evaluate whether the process shown in the video meets expectations, that is, whether the video aligns with the predefined “Progress goal”. This “Progress goal” is: “progress_goal”. Based on the consistency between the video and the “Progress goal”, assign a score from 1 to 5 (from least consistent to most consistent), and provide the reason for the score you assign. Output your evaluation result in the following format: "video_quality_score": "<your video_quality score here>", "reason_for_video_quality_score": "<your reason here>", "progress_consistency_score": "<your progress_consistency score here>", "reason_for_progress_consistency_score": "<your reason here>", (Optional) "implicit_rule_score": "<your implicit_rule score here>", (Optional) "reason_for_implicit_rule_score": "<your reason here>", (Optional) "progress_goal_score": "<your progress_goal score here>", (Optional) "reason_for_progress_goal_score": "<your reason here>" Last Frame Goal Annotation You will be given a prompt for an image-to-video task and its correspond- ing first-frame image. Your task is to obtain a “last frame goal”, that is, by imagining or reasoning, determine what the final frame of the video gener- ated from the given first-frame image and image-to-video prompt should look like (for example, if the first frame shows the interior of a house that is dimly lit with a light switch in the center of the image, and the image- 24M. Luo et al. to-video prompt is to turn on the light, then the “last frame goal” should be “the switch is turned on and the interior becomes bright”). The format should be concise, in the form of a single sentence. The image-to-video prompt: user_prompt Output your “last frame goal” in the following format: "last_frame_goal": "<your last frame goal here>" Text-to-Image Output Filtering Please determine whether the prompt in this text-to-image output per- fectly matches the corresponding generated image. That is, all elements appearing in the prompt must be present in the image, and no elements not mentioned in the prompt should appear in the image. If there is no perfect match, set result to 0 and provide a reason; otherwise, set result to 1 and set the reason to null. The text-to-image prompt: t2i_prompt Output your result and reason in the following format: "result": "<your result here>", "reason": "<your reason here>" or null Supplementary Material for Think-with-Video Reasoning25 D Vid-PRE Prompt Templates Chain-of-Thought and Optimized Prompt Annotation You are a professional image-to-video prompt engineer. Your task is to produce a chain-of-thought and an optimized prompt, given ONE static image (first frame) and ONE original prompt for an image-to-video model. Write the chain-of-thought in a natural, human style: free-form prose, varied sentence lengths, no numbered lists or headings, and no rigid section titles. It should read like someone genuinely thinking through the problem, so light hesitations such as “hmm”, “oh”, “well”, and “okay, so” are allowed as long as they do not dominate the text. Do not follow a fixed template; let the narrative flow adapt to each case. In the chain-of-thought, casually but clearly cover: what you infer from the image at a high level; how you reinterpret and clarify the user’s intent; any minimal assumptions you decide to make; the key rules and constraints that should govern motion, camera or viewport behavior, and what must stay unchanged; when the sequence should end; and what counts as success versus failure for this image-to-video generation. You may be detailed and reflective, but keep the chain-of-thought within about 384 tokens. Avoid bullet points and explicit enumeration such as “1), 2), 3)”. Then write the optimized prompt as a single, self-contained instruction for a commercial image-to-video model: imperative, unambiguous, and concise compared to the chain-of-thought, while still fully specifying the needed behavior. Explicitly encode the semantics you rely on from the image, motion or behavior rules, camera behavior, invariants, and the end condi- tion. Avoid meta-commentary, filler words, and examples in the optimized prompt. General rules: do not request extra inputs; if you assume something, state it in the chain-of-thought and reflect it succinctly in the optimized prompt; do not invent new visual assets that contradict the image or user intent; prefer operational, checkable wording over purely subjective adjectives. Avoid using unescaped double quotes inside the JSON strings. Output strictly in JSON with two fields. Original prompt: original_prompt Output your chain-of-thought and optimized prompt in the following for- mat: "chain_of_thought": "<your chain-of-thought here>", "optimized_prompt": "<your optimized prompt here>" 26M. Luo et al. Rejection Sampling This is a task to determine whether an image-to-video prompt is appro- priate and valid. You will be given an image as the first frame of this image-to-video task, a text describing the task objective (which can be considered the ground truth), and an annotated image-to-video prompt. You need to examine the image-to-video prompt for: whether all content mentioned in the ground truth text is implemented in the image-to-video prompt (i.e., whether it is consistent with the ground truth text); and whether the spatial relationships mentioned in the image-to-video prompt, such as “left” and “bottom right”, are consistent with the spatial relation- ships in the image (i.e., consistency with the first frame). Finally, please give the image-to-video prompt a comprehensive score on both aspects of consistency, using a scale from 1 to 5 (from extremely inappropriate to extremely appropriate), and then provide a reason to support your score. The ground truth text: gt_text The image-to-video prompt: optimized_prompt Output your score in the following format: "score": "<your score here>", "score_reason": "<your reason here>" Rule-Based Final Prompt Optimization This is a task to optimize an image-to-video prompt based on given rules. You will be given an image-to-video prompt and a text containing a set of rules designed to improve the prompt’s output. You need to append these rule texts naturally to the image-to-video prompt with minimal modifica- tions, adding the final “patch” to the prompt. The rule text: rule The image-to-video prompt: optimized_prompt Output your optimized prompt in the following format: "optimized_prompt": "<your optimized image-to-video prompt here>" Reward Scoring for Reinforcement Learning You are a strict-but-pragmatic evaluator for image-to-video prompt rewrit- ing. You will be given a first-frame image, an original image-to-video prompt, and a rewritten prompt. The rewritten prompt should be an ex- pansion or refinement of the original prompt. EVALUATION PHILOSOPHY (VERY IMPORTANT): Supplementary Material for Think-with-Video Reasoning27 A) Prioritize CONTENT requirements over CAMERA terminology. B) Be tolerant to camera wording when the original prompt is ambiguous or contradictory. C) The most important improvement is adding TASK-CRITICAL DE- TAILS (decision-making details), not generic quality constraints. Step 1: Extract the ORIGINAL prompt into two lists: 1) NON-NEGOTIABLE CONTENT REQUIREMENTS (highest prior- ity): key objects or regions to target, required edits or overlays (arrows, boxes, glow, text), and any measurement or fixed-scale constraints. 2) SECONDARY CAMERA REQUIREMENTS (lower priority): static, zoom, pan, or dolly wording. Step 2: Check coverage in the rewritten prompt: – For each NON-NEGOTIABLE CONTENT requirement, mark: pre- served, partially preserved, missing, or substituted. – For camera requirements, mark: consistent, acceptable reinterpreta- tion, or conflicting. Step 3: Identify NEW DETAILS added by the rewritten prompt and classify each as either: – TASK-CRITICAL detail (high value): concrete, verifiable decisions that directly determine successful completion (e.g., exact moves or paths, which cell gets which number, which puzzle piece goes to which slot or region, exact arrow or box placement). – SUPPORTING detail (medium or low value): generic quality con- straints or anti-failure constraints (e.g., keep background fixed, only object X moves, smooth motion, preserve lighting, maintain geome- try). SCORING RULES: 1) Intent preservation (1–5): Score mainly based on whether NON- NEGOTIABLE CONTENT requirements are preserved. If ANY key con- tent requirement is missing or substituted, the intent preservation score should be≤3. Do NOT strongly penalize minor camera wording differ- ences when the content goal is preserved. 2) Image consistency (1–5): Determine whether the mentioned objects or regions are actually visible and correctly referenced. Penalize hallucina- tions or incorrect spatial relations. 3) Solution correctness (1–5): focus on TASK-CRITICAL NEW DETAILS and whether they are correct. Use this rubric: 28M. Luo et al. – 5: Adds MULTIPLE correct TASK-CRITICAL details that substan- tially reduce ambiguity and make the task almost directly executable. – 4: Adds at least ONE correct TASK-CRITICAL detail that clearly improves executability, with no major harmful errors. – 3: Neutral or modest refinement: adds only SUPPORTING details, or adds correct but non-critical details that do not materially change executability. – 2: Adds incorrect or conflicting details (including incorrect task-critical details) that would likely reduce success. – 1: The added details are largely wrong or make the task fail. IMPORTANT coupling rule: If intent_preservation_score≤ 3 because key content requirements are missing or substituted, then solution_correctness_score MUST also be≤3. The original image-to-video prompt: origin_user_prompt The rewritten prompt: user_prompt Output JSON in the following format (strings must be properly escaped): "intent_preservaion_score": <1-5>, "reason_for_intent_preservaion_score": "<cite Step 2 coverage>", "image_consistency_score": <1-5>, "reason_for_image_consistency_score": "<reason>", "solution_correctness_score": <1-5>, "reason_for_solution_correctness_score": "<whether NEW details are task-critical vs supporting, and whether they are correct>"