Awesome-RL-for-Video-Generation
🤖 Introduction
Welcome to the GitHub repository for Awesome-RL-for-Video-Generation! This repository serves as a curated collection of research, resources, and tools related to Reinforcement Learning (RL) for Video Generation. Our goal is to provide an up-to-date and comprehensive overview of RL techniques used in video generation, focusing on the latest advancements. We aim to bridge the gap between RL theory and real-world applications in video generation tasks, offering a solid foundation for future research and development in this field. We hope this repository will serve as a valuable resource for anyone interested in exploring RL applications in video generation!
🔥 News
- [February 14, 2025] We have developed an agent that automatically collects and analyzes the latest papers in the RL-based Video Generation field. It will update the Related Papers daily at 1:00 AM UTC+8.
🔍 Related Papers
We are committed to offering researchers the latest advancements in the field. By regularly reviewing and evaluating recent research studies, we ensure that the list of papers stays up-to-date.
⚠️ The paper analysis may not be accurate and is for reference only!
Date
Paper
Contribution
Available Link
Apr 2026
ARGen: Affect-Reinforced Generative Augmentation towards Vision-based Dynamic Emotion Perception
[](http://arxiv.org/pdf/2604.12255)
• Affiliation: Fudan University
• Method Name: ARGen, Base Model: Qwen2.5-7B-VL, Strategy: policy gradient method
Apr 2026
ARGen: Affect-Reinforced Generative Augmentation towards Vision-based Dynamic Emotion Perception
[](http://arxiv.org/pdf/2604.12255)
• Affiliation: Fudan University
• Method Name: ARGen, Base Model: ModelScopeT2V, Strategy: Policy Gradient
Apr 2026
LPM 1.0: Video-based Character Performance Model
[](http://arxiv.org/pdf/2604.07823)
[](https://large-performance-model.github.io)
• Affiliation: International Digital Economy Academy
• Method Name: LPM 1.0, Base Model: Wan2.1-I2V (16B), Strategy: DPO
• Benchmark Name: LPM-Bench, Data Number: 1000, Evaluation Metric: Motion Dynamics, Identity Consistency, Text Controllability, Audio-Video Synchronization
Apr 2026
LPM 1.0: Video-based Character Performance Model
[](http://arxiv.org/pdf/2604.07823)
[](large-performance-model.github.io)
• Affiliation: Not specified
• Method Name: LPM 1.0, Base Model: Wan2.1-I2V (16B), Strategy: DPO
• Benchmark Name: LPM-Bench, Data Number: 1000, Evaluation Metric: Good/Same/Bad (GSB), Likert scale
Apr 2026
Evolution of Video Generative Foundations
[](http://arxiv.org/pdf/2604.06339)
[](https://github.com/sjtuplayer/Awesome-Video-Foundations)
• Affiliation: Shanghai Jiao Tong University
• Paper Number: 301
Apr 2026
OP-GRPO: Efficient Off-Policy GRPO for Flow-Matching Models
[](http://arxiv.org/pdf/2604.04142)
• Affiliation: Zhejiang University
• Method Name: OP-GRPO, Base Model: Wan2.1-1.4B, Strategy: GRPO
Apr 2026
VERTIGO: Visual Preference Optimization for Cinematic Camera Trajectory Generation
[](http://arxiv.org/pdf/2604.02467)
[](http://vertigo.magic-lab.tech/)
• Affiliation: Shanghai University
• Method Name: VERTIGO, Base Model: Qwen 2.5-VL, Strategy: DPO
• Benchmark Name: LenScript, Data Number: 120000, Evaluation Metric: FCD, CS, Precision, Recall, Density, Coverage, Missing Rate, Consistency, Aesthetic
Mar 2026
Wan-R1: Verifiable-Reinforcement Learning for Video Reasoning
[](http://arxiv.org/pdf/2603.27866)
• Affiliation: Iowa State University
• Method Name: Wan-R1, Base Model: Wan2.2-TI2V-5B, Strategy: GRPO
Mar 2026
Persistent Robot World Models: Stabilizing Multi-Step Rollouts via Reinforcement Learning
[](http://arxiv.org/pdf/2603.25685)
[](https://www.jaibardhan.com/persistworld)
• Affiliation: Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague
• Method Name: PersistWorld, Base Model: Ctrl-World, Strategy: DiffusionNFT
Mar 2026
AnyID: Ultra-Fidelity Universal Identity-Preserving Video Generation from Any Visual References
[](http://arxiv.org/pdf/2603.25188)
[](https://johnneywang.github.io/AnyID-webpage)
• Affiliation: Xi’an Jiaotong University
• Method Name: AnyID, Base Model: Wan-5B, Strategy: DPO
• Benchmark Name: AnyID Benchmark, Data Number: 50 reference sets, 100 prompts, Evaluation Metric: Holi-Arc, Holi-Cur, Ele-CLIP, Ele-DINO, App., Mot., Bg., Sta., Dyn.
Mar 2026
OmniWeaving: Towards Unified Video Generation with Free-form Composition and Reasoning
[](http://arxiv.org/pdf/2603.24458)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Mar 2026
See, Remember, Explore: A Benchmark and Baselines for Streaming Spatial Reasoning
[](http://arxiv.org/pdf/2603.23864)
[](https://github.com/showlab/InstructVideo)
[](https://showlab.github.io/InstructVideo/)
• Affiliation: Zhejiang University
• Method Name: InstructVideo, Base Model: , Strategy: PPO
Mar 2026
ABot-PhysWorld: Interactive World Foundation Model for Robotic Manipulation with Physics Alignment
[](http://arxiv.org/pdf/2603.23376)
[](https://github.com/amap-cvlab/ABot-PhysWorld)
[](https://github.com/amap-cvlab/ABot-PhysWorld)
• Affiliation: AMAP CV Lab, Alibaba Group
• Method Name: ABot-PhysWorld, Base Model: Wan2.1-I2V-14B, Strategy: DPO
• Benchmark Name: EZSbench, Data Number: , Evaluation Metric: Physical Score (Sv)
Mar 2026
Manifold-Aware Exploration for Reinforcement Learning in Video Generation
[](http://arxiv.org/pdf/2603.21872)
• Affiliation: HKUST
• Method Name: SAGE-GRPO, Base Model: HunyuanVideo1.5, Strategy: GRPO
Mar 2026
VIGOR: VIdeo Geometry-Oriented Reward for Temporal Generative Alignment
[](http://arxiv.org/pdf/2603.16271)
[](https://vigor-geometry-reward.site/)
• Affiliation: Nankai University
• Method Name: VIGOR, Base Model: Wan2.1-T2V-1.3B, Strategy: DPO
• Benchmark Name: GB3DV-25k, Data Number: 25,600, Evaluation Metric: PSNR, SSIM, LPIPS, EPI, RPX, RPT
Mar 2026
LICA: Layered Image Composition Annotations for Graphic Design Research
[](http://arxiv.org/pdf/2603.16098)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Mar 2026
LatSearch: Latent Reward-Guided Search for Faster Inference-Time Scaling in Video Diffusion
[](http://arxiv.org/pdf/2603.14526)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Mar 2026
GenState-AI: State-Aware Dataset for Text-to-Video Retrieval on AI-Generated Videos
[](http://arxiv.org/pdf/2603.14426)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Mar 2026
DreamVideo-Omni: Omni-Motion Controlled Multi-Subject Video Customization with Latent Identity Reinforcement Learning
[](http://arxiv.org/pdf/2603.12257)
[](https://dreamvideo-omni.github.io)
• Affiliation: Fudan University
• Method Name: DreamVideo-Omni, Base Model: Wan2.1-1.3B T2V, Strategy: Reward Feedback Learning
• Benchmark Name: DreamOmni Bench, Data Number: 1027, Evaluation Metric: ['CLIP-T', 'R-CLIP', 'R-DINO', 'Face-S', 'mIoU', 'EPE']
Mar 2026
Risk-Controllable Multi-View Diffusion for Driving Scenario Generation
[](http://arxiv.org/pdf/2603.11534)
[](https://github.com/venshow-w/RiskMV-DPO)
• Affiliation: Tsinghua University
• Method Name: RiskMV-DPO, Base Model: , Strategy: RA-DPO
Mar 2026
SPIRAL: A Closed-Loop Framework for Self-Improving Action World Models via Reflective Planning Agents
[](http://arxiv.org/pdf/2603.08403)
• Affiliation: Zhejiang University
• Method Name: SPIRAL, Base Model: Qwen3-VL-8B, Strategy: GRPO
• Method Name: SPIRAL, Base Model: Qwen3-VL-8B, Strategy: DPO
• Benchmark Name: ActWM-Bench, Data Number: 300, Evaluation Metric: Action Completeness, Motion Smoothness, Object Interaction, Physical Fidelity
Mar 2026
Advances in GRPO for Generation Models: A Survey
[](http://arxiv.org/pdf/2603.06623)
• Affiliation: SJTU
• Paper Number: 198
Mar 2026
Place-it-R1: Unlocking Environment-aware Reasoning Potential of MLLM for Video Object Insertion
[](http://arxiv.org/pdf/2603.06140)
• Affiliation: HKUST
• Method Name: Place-it-R1, Base Model: Qwen-VL 2.5-7B, WAN 1.3B, Strategy: DPO
• Benchmark Name: FlexInsert, Data Number: 100, Evaluation Metric: CLIP-I, DINO-I, Smoothness, Aesthetics, PC, PR, PP
Mar 2026
PhyPrompt: RL-based Prompt Refinement for Physically Plausible Text-to-Video Generation
[](http://arxiv.org/pdf/2603.03505)
• Affiliation: Northwestern University
• Method Name: PhyPrompt, Base Model: Qwen2.5, Strategy: GRPO
Mar 2026
PhyPrompt: RL-based Prompt Refinement for Physically Plausible Text-to-Video Generation
[](http://arxiv.org/pdf/2603.03505)
• Affiliation: Northwestern University
• Method Name: PhyPrompt, Base Model: Qwen2.5-Instruct, Strategy: GRPO
Mar 2026
Phys4D: Fine-Grained Physics-Consistent 4D Modeling from Video Diffusion
[](http://arxiv.org/pdf/2603.03485)
[](https://sensational-brioche-7657e7.netlify.app/)
• Affiliation: Northwestern University
• Method Name: Phys4D, Base Model: WAN2.2-5b, Strategy: PPO
• Method Name: Phys4D, Base Model: CogVideoX-5b, Strategy: PPO
• Method Name: Phys4D, Base Model: Open-Sora V1.2, Strategy: PPO
• Benchmark Name: 4D world consistency evaluation, Data Number: , Evaluation Metric: 4D Chamfer Distance, Worldline L2 Error, Mean Drift, Final Drift, Fail Rate, Trajectory Length, Novel-Time Depth Err, Novel-Time Warp Err
Mar 2026
StepVAR: Structure-Texture Guided Pruning for Visual Autoregressive Models
[](http://arxiv.org/pdf/2603.01757)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Mar 2026
FlowPortrait: Reinforcement Learning for Audio-Driven Portrait Video Generation
[](http://arxiv.org/pdf/2603.00159)
• Affiliation: Johns Hopkins University
• Method Name: FlowPortrait, Base Model: , Strategy: GRPO
Feb 2026
The Trinity of Consistency as a Defining Principle for General World Models
[](http://arxiv.org/pdf/2602.23152)
• Affiliation: University of Chinese Academy of Sciences
• Benchmark Name: CoW-Bench, Data Number: 1485, Evaluation Metric: 0–2 scale
• Paper Number: 516
Feb 2026
SPATIALALIGN: Aligning Dynamic Spatial Relationships in Video Generation
[](http://arxiv.org/pdf/2602.22745)
• Affiliation: Nanyang Technological University
• Method Name: SPATIALALIGN, Base Model: Wan2.1-1.3B, Strategy: DPO
• Benchmark Name: DSR-DATASET, Data Number: 620, Evaluation Metric: DSR-SCORE, Correctness@0.7, ID Consistency, CLIP-IQA, Imaging Quality
Feb 2026
SPATIALALIGN: Aligning Dynamic Spatial Relationships in Video Generation
[](http://arxiv.org/pdf/2602.22745)
• Affiliation: Nanyang Technological University
• Method Name: SPATIALALIGN, Base Model: Wan2.1-1.3B, Strategy: DPO
• Benchmark Name: DSR-DATASET, Data Number: 620, Evaluation Metric: DSR-SCORE
Feb 2026
JavisDiT++: Unified Modeling and Optimization for Joint Audio-Video Generation
[](http://arxiv.org/pdf/2602.19163)
[](https://JavisVerse.github.io/JavisDiT2-page)
• Affiliation: Zhejiang University
• Method Name: JavisDiT++, Base Model: Wan2.1-1.3B-T2V, Strategy: DPO
Feb 2026
JavisDiT++: Unified Modeling and Optimization for Joint Audio-Video Generation
[](http://arxiv.org/pdf/2602.19163)
[](https://JavisVerse.github.io/JavisDiT2-page)
• Affiliation: Zhejiang University
• Method Name: JavisDiT++, Base Model: Wan2.1-1.3B-T2V, Strategy: DPO
Feb 2026
EchoTorrent: Towards Swift, Sustained, and Streaming Multi-Modal Video Generation
[](http://arxiv.org/pdf/2602.13669)
• Affiliation: Ant Group
• Method Name: EchoTorrent, Base Model: InfiniteTalk, Strategy: GRPO
Feb 2026
ReTracing: An Archaeological Approach Through Body, Machine, and Generative Systems
[](http://arxiv.org/pdf/2602.11242)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Feb 2026
Smaller is Better: Generative Models Can Power Short Video Preloading
[](http://arxiv.org/pdf/2602.09484)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Feb 2026
SemanticMoments: Training-Free Motion Similarity via Third Moment Features
[](http://arxiv.org/pdf/2602.09146)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Feb 2026
Generalizing Sports Feedback Generation by Watching Competitions and Reading Books: A Rock Climbing Case Study
[](http://arxiv.org/pdf/2602.08996)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Feb 2026
DynFOA: Generating First-Order Ambisonics with Conditional Diffusion for Dynamic and Acoustically Complex 360-Degree Videos
[](http://arxiv.org/pdf/2602.06846)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Feb 2026
EgoAVU: Egocentric Audio-Visual Understanding
[](http://arxiv.org/pdf/2602.06139)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Feb 2026
Euphonium: Steering Video Flow Matching via Process Reward Gradient Guided Stochastic Dynamics
[](http://arxiv.org/pdf/2602.04928)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Feb 2026
Tiled Prompts: Overcoming Prompt Underspecification in Image and Video Super-Resolution
[](http://arxiv.org/pdf/2602.03342)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Feb 2026
Unified Personalized Reward Model for Vision Generation
[](http://arxiv.org/pdf/2602.02380)
[](https://codegoat24.github.io/UnifiedReward/flex)
• Affiliation: Fudan University
• Method Name: UnifiedReward-Flex, Base Model: Wan2.1-T2V-14B, Strategy: GRPO
Feb 2026
FSVideo: Fast Speed Video Diffusion Model in a Highly-Compressed Latent Space
[](http://arxiv.org/pdf/2602.02092)
[](https://kingofprank.github.io/fsvideo/)
• Affiliation: ByteDance
• Method Name: FSVideo, Base Model: Wan2.1-14B-I2V, Strategy: GRPO
• Method Name: FSVideo, Base Model: Wan2.1-14B-I2V, Strategy: ReFL
Feb 2026
PISCES: Annotation-free Text-to-Video Post-Training via Optimal Transport-Aligned Rewards
[](http://arxiv.org/pdf/2602.01624)
• Affiliation: Microsoft
• Method Name: PISCES, Base Model: HunyuanVideo, Strategy: GRPO
• Method Name: PISCES, Base Model: VideoCrafter2, Strategy: GRPO
Feb 2026
PISCES: Annotation-free Text-to-Video Post-Training via Optimal Transport-Aligned Rewards
[](http://arxiv.org/pdf/2602.01624)
• Affiliation: Microsoft
• Method Name: PISCES, Base Model: HunyuanVideo, Strategy: GRPO
• Method Name: PISCES, Base Model: VideoCrafter2, Strategy: GRPO
Jan 2026
SketchDynamics: Exploring Free-Form Sketches for Dynamic Intent Expression in Animation Generation
[](http://arxiv.org/pdf/2601.20622)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Jan 2026
The Script is All You Need: An Agentic Framework for Long-Horizon Dialogue-to-Cinematic Video Generation
[](http://arxiv.org/pdf/2601.17737)
[](https://github.com/Tencent/digitalhuman/tree/main/ScriptAgent)
• Affiliation: Tencent Hunyuan Multimodal Department
• Method Name: ScripterAgent, Base Model: Qwen-Omni-7B, Strategy: GRPO
• Benchmark Name: ScriptBench, Data Number: 1750, Evaluation Metric: Visual-Script Alignment (VSA)
Jan 2026
SkyReels-V3 Technique Report
[](http://arxiv.org/pdf/2601.17323)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Jan 2026
A Mechanistic View on Video Generation as World Models: State and Dynamics
[](http://arxiv.org/pdf/2601.17067)
• Affiliation: Hong Kong University of Science and Technology (Guangzhou)
• Paper Number: 188
Jan 2026
From Generative Engines to Actionable Simulators: The Imperative of Physical Grounding in World Models
[](http://arxiv.org/pdf/2601.15533)
• Affiliation: University of Oxford
• Paper Number: 49
Jan 2026
MVGD-Net: A Novel Motion-aware Video Glass Surface Detection Network
[](http://arxiv.org/pdf/2601.13715)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Jan 2026
CroBIM-V: Memory-Quality Controlled Remote Sensing Referring Video Object Segmentation
[](http://arxiv.org/pdf/2601.12076)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Jan 2026
PhysRVG: Physics-Aware Unified Reinforcement Learning for Video Generative Models
[](http://arxiv.org/pdf/2601.11087)
[](https://lucaria-academy.github.io/PhysRVG/)
• Affiliation: Zhejiang University
• Method Name: PhysRVG, Base Model: Wan2.2 5B, Strategy: GRPO
• Benchmark Name: PhysRVGBench, Data Number: 700, Evaluation Metric: Intersection over Union (IoU), Trajectory Offset (TO)
Jan 2026
TAGRPO: Boosting GRPO on Image-to-Video Generation with Direct Trajectory Alignment
[](http://arxiv.org/pdf/2601.05729)
• Affiliation: The University of Hong Kong
• Method Name: TAGRPO, Base Model: Wan 2.2, Strategy: GRPO
• Method Name: TAGRPO, Base Model: HunyuanVideo-1.5, Strategy: GRPO
• Benchmark Name: TAGRPO-Bench, Data Number: 200, Evaluation Metric: Q-Save (Visual Quality, Dynamic Quality, Image Alignment), HPSv3
Jan 2026
TAGRPO: Boosting GRPO on Image-to-Video Generation with Direct Trajectory Alignment
[](http://arxiv.org/pdf/2601.05729)
• Affiliation: The University of Hong Kong
• Method Name: TAGRPO, Base Model: Wan 2.2, Strategy: GRPO
• Method Name: TAGRPO, Base Model: HunyuanVideo-1.5, Strategy: GRPO
• Benchmark Name: TAGRPO-Bench, Data Number: 200, Evaluation Metric: Q-Save
• Benchmark Name: TAGRPO-Bench, Data Number: 200, Evaluation Metric: HPSv3
Jan 2026
Diffusion-DRF: Differentiable Reward Flow for Video Diffusion Fine-Tuning
[](http://arxiv.org/pdf/2601.04153)
• Affiliation: Northeastern University
• Method Name: Diffusion-DRF, Base Model: Wan2.1-1.3B-T2V, Strategy: Differentiable Reward Fine-tuning
Jan 2026
Diffusion-DRF: Differentiable Reward Flow for Video Diffusion Fine-Tuning
[](http://arxiv.org/pdf/2601.04153)
• Affiliation: Northeastern University
• Method Name: Diffusion-DRF, Base Model: Wan2.1-1.3B-T2V, Strategy: Differentiable Reward Fine-tuning
Jan 2026
Mind the Generative Details: Direct Localized Detail Preference Optimization for Video Diffusion Models
[](http://arxiv.org/pdf/2601.04068)
• Affiliation: Harbin Institute of Technology
• Method Name: LocalDPO, Base Model: Wan2.1-1.3B, Strategy: DPO
• Method Name: LocalDPO, Base Model: CogVideoX-2B, Strategy: DPO
• Method Name: LocalDPO, Base Model: CogVideoX-5B, Strategy: DPO
Jan 2026
Mind the Generative Details: Direct Localized Detail Preference Optimization for Video Diffusion Models
[](http://arxiv.org/pdf/2601.04068)
• Affiliation: Harbin Institute of Technology
• Method Name: LocalDPO, Base Model: Wan2.1-1.3B, Strategy: DPO
• Method Name: LocalDPO, Base Model: CogVideoX-2B, Strategy: DPO
• Method Name: LocalDPO, Base Model: CogVideoX-5B, Strategy: DPO
Jan 2026
Thinking with Frames: Generative Video Distortion Evaluation via Frame Reward Model
[](http://arxiv.org/pdf/2601.04033)
• Affiliation: University of Science and Technology of China
• Method Name: REACT, Base Model: Qwen2.5-VL-7B, Strategy: GRPO
• Benchmark Name: REACT-Bench, Data Number: 2600, Evaluation Metric: F1-score
Jan 2026
Thinking with Frames: Generative Video Distortion Evaluation via Frame Reward Model
[](http://arxiv.org/pdf/2601.04033)
• Affiliation: University of Science and Technology of China
• Method Name: REACT, Base Model: Qwen2.5-VL-7B, Strategy: GRPO
• Benchmark Name: REACT-Bench, Data Number: 2600, Evaluation Metric: Accuracy, F1-score, Precision, Recall
Jan 2026
A Versatile Multimodal Agent for Multimedia Content Generation
[](http://arxiv.org/pdf/2601.03250)
• Affiliation: University of Rochester
• Method Name: MultiMedia-Agent, Base Model: MiniCPM-V2, Strategy: DPO
• Benchmark Name: 18 real world task types, Data Number: 1260, Evaluation Metric: Dover Score, Pick Score, Human Alignment, Aesthetic Score, Psychological Appealing, Audio Video Alignment
Jan 2026
A Versatile Multimodal Agent for Multimedia Content Generation
[](http://arxiv.org/pdf/2601.03250)
• Affiliation: University of Rochester
• Method Name: MultiMedia-Agent, Base Model: Minicpm-v2, Strategy: DPO
Dec 2025
PhyGDPO: Physics-Aware Groupwise Direct Preference Optimization for Physically Consistent Text-to-Video Generation
[](http://arxiv.org/pdf/2512.24551)
[](https://caiyuanhao1998.github.io/project/PhyGDPO)
• Affiliation: Meta Superintelligence Labs
• Method Name: PhyGDPO, Base Model: Wan2.1-T2V-14B, Strategy: DPO
• Benchmark Name: PhyVidGen-135K, Data Number: 135K, Evaluation Metric:
Dec 2025
SoliReward: Mitigating Susceptibility to Reward Hacking and Annotation Noise in Video Generation Reward Models
[](http://arxiv.org/pdf/2512.22170)
• Affiliation: Huazhong University of Science and Technology
• Method Name: SoliReward, Base Model: HunyuanVideo, Strategy: GRPO
• Method Name: SoliReward, Base Model: HunyuanVideo, Strategy: DPO
• Benchmark Name: subject deformity and physical plausibility benchmark, Data Number: 50000, Evaluation Metric: RM Accuracy
Dec 2025
DreaMontage: Arbitrary Frame-Guided One-Shot Video Generation
[](http://arxiv.org/pdf/2512.21252)
[](https://dreamontage.github.io/DreaMontage)
• Affiliation: ByteDance
• Method Name: DreaMontage, Base Model: Seedance 1.0, Strategy: DPO
Dec 2025
VIVA: VLM-Guided Instruction-Based Video Editing with Reward Optimization
[](http://arxiv.org/pdf/2512.16906)
[](https://viva-paper.github.io/)
• Affiliation: Brown University
• Method Name: VIVA, Base Model: HunyuanVideo-T2V-13B, Strategy: GRPO
Dec 2025
Kling-Omni Technical Report
[](http://arxiv.org/pdf/2512.16776)
[](https://app.klingai.com/global/omni/new)
• Affiliation: Kuaishou Technology
• Method Name: Kling-Omni, Base Model: , Strategy: DPO
Dec 2025
What Happens Next? Next Scene Prediction with a Unified Video Model
[](http://arxiv.org/pdf/2512.13015)
[](https://nextsceneprediction.github.io/)
• Affiliation: Pennsylvania State University
• Method Name: unified video model, Base Model: Qwen-VL, LTX, Strategy: GRPO
• Benchmark Name: NSP dataset, Data Number: 0.97M samples for SFT, 8K samples for RL, 1K samples for test, Evaluation Metric: causal consistency
Dec 2025
OmniPerson: Unified Identity-Preserving Pedestrian Generation
[](http://arxiv.org/pdf/2512.02554)
[](https://github.com/showlab/assistgui)
[](https://showlab.github.io/assistgui/)
• Affiliation: Zhejiang University
• Method Name: RL-Video-Gen, Base Model: Qwen2-VL-7B, Strategy: GRPO
• Benchmark Name: VideoGenBench, Data Number: 5000, Evaluation Metric: FID
Dec 2025
PhyDetEx: Detecting and Explaining the Physical Plausibility of T2V Models
[](http://arxiv.org/pdf/2512.01843)
[](https://github.com/Zeqing-Wang/PhyDetEx)
• Affiliation: Sun Yat-sen University
• Method Name: Physical-Aware DPO, Base Model: WanX2.1 1.3B, Strategy: DPO
• Benchmark Name: PID (Physical Implausibility Detection) dataset, Data Number: 3088, Evaluation Metric: F1 Score
Nov 2025
McSc: Motion-Corrective Preference Alignment for Video Generation with Self-Critic Hierarchical Reasoning
[](http://arxiv.org/pdf/2511.22974)
[](https://qiushiyang.github.io/McSc/)
• Affiliation: Tongyi Lab, Alibaba Group
• Method Name: McSc, Base Model: Qwen2-VL-7B-Instruct, Strategy: GRPO
• Method Name: McDPO, Base Model: Wan2.1-T2V-1.3B, Strategy: DPO
Nov 2025
Diverse Video Generation with Determinantal Point Process-Guided Policy Optimization
[](http://arxiv.org/pdf/2511.20647)
[](https://diverse-video.github.io)
• Affiliation: Virginia Tech
• Method Name: DPP-GRPO, Base Model: Qwen2-7b-Instruct, Strategy: GRPO
• Benchmark Name: diverse video-prompt dataset, Data Number: 30,000, Evaluation Metric: TIE, TCE, CLIP
Nov 2025
Growing with the Generator: Self-paced GRPO for Video Generation
[](http://arxiv.org/pdf/2511.19356)
• Affiliation: University of Science and Technology of China
• Method Name: Self-Paced GRPO, Base Model: Wan2.1-T2V, Strategy: GRPO
• Method Name: Self-Paced GRPO, Base Model: HunyuanVideo, Strategy: GRPO
Nov 2025
PhysCorr: Dual-Reward DPO for Physics-Constrained Text-to-Video Generation with Automated Preference Selection
[](http://arxiv.org/pdf/2511.03997)
• Affiliation: Beijing Institute of Technology
• Method Name: PhysCorr, Base Model: , Strategy: DPO
• Method Name: PhyDPO, Base Model: , Strategy: DPO
Nov 2025
PhysCorr: Dual-Reward DPO for Physics-Constrained Text-to-Video Generation with Automated Preference Selection
[](http://arxiv.org/pdf/2511.03997)
• Affiliation: Beijing Institute of Technology
• Method Name: PhysCorr, Base Model: , Strategy: DPO
• Method Name: PhysicsRM, Base Model: LLaVA-Video-Qwen2-7B, Strategy: supervised learning with Huber loss
• Method Name: PhyDPO, Base Model: , Strategy: reweighted DPO
Nov 2025
Reg-DPO: SFT-Regularized Direct Preference Optimization with GT-Pair for Improving Video Generation
[](http://arxiv.org/pdf/2511.01450)
• Affiliation: ByteDance
• Method Name: Reg-DPO, Base Model: Wan2.1-I2V-14B-720P, Strategy: DPO
Nov 2025
CueBench: Advancing Unified Understanding of Context-Aware Video Anomalies in Real-World
[](http://arxiv.org/pdf/2511.00613)
[](https://github.com/Mia-YatingYu/Cue-R1)
• Affiliation: Northwestern Polytechnical University, Xi’an Shaanxi, 710129, China
• Method Name: CUE-R1, Base Model: Qwen2.5-VL-3B, Strategy: GRPO
• Benchmark Name: CUEBENCH, Data Number: 2950, Evaluation Metric: hierarchy score
Nov 2025
ID-Composer: Multi-Subject Video Synthesis with Hierarchical Identity Preservation
[](http://arxiv.org/pdf/2511.00511)
• Affiliation: Peking University
• Method Name: ID-COMPOSER, Base Model: Wan-Video-1.3B, Strategy: Flow-GRPO
• Benchmark Name: OpenS2V-Nexus, Data Number: 218230, Evaluation Metric: NexusScore
Nov 2025
ID-Composer: Multi-Subject Video Synthesis with Hierarchical Identity Preservation
[](http://arxiv.org/pdf/2511.00511)
• Affiliation: Peking University
• Method Name: ID-COMPOSER, Base Model: Wan-Video-1.3B, Strategy: Flow-GRPO
• Benchmark Name: OpenS2V-Nexus, Data Number: 218230, Evaluation Metric: NexusScore
Nov 2025
World Simulation with Video Foundation Models for Physical AI
[](http://arxiv.org/pdf/2511.00062)
[](https://github.com/nvidia-cosmos/cosmos-predict2.5)
• Affiliation: NVIDIA
• Method Name: Cosmos-Predict2.5, Base Model: Cosmos-Reason1, Strategy: GRPO
Nov 2025
World Simulation with Video Foundation Models for Physical AI
[](http://arxiv.org/pdf/2511.00062)
[](https://github.com/nvidia-cosmos/cosmos-predict2.5)
• Affiliation: NVIDIA
• Method Name: Cosmos-Predict2.5, Base Model: Cosmos-Reason1, Strategy: GRPO
Oct 2025
Emu3.5: Native Multimodal Models are World Learners
[](http://arxiv.org/pdf/2510.26583)
[](https://github.com/baaivision/Emu3.5)
[](https://emu.world)
• Affiliation: BAAI
• Method Name: Discrete Diffusion Adaptation, Base Model: Qwen3, Strategy: GRPO
Oct 2025
Omni-Reward: Towards Generalist Omni-Modal Reward Modeling with Free-Form Preferences
[](http://arxiv.org/pdf/2510.23451)
[](https://github.com/HongbangYuan/OmniReward)
• Affiliation: School of Artificial Intelligence, University of Chinese Academy of Sciences
• Method Name: Omni-RewardModel-BT, Base Model: MiniCPM-o-2.6, Strategy: Bradley-Terry
• Method Name: Omni-RewardModel-R1, Base Model: Qwen2.5-VL-7B-Instruct, Strategy: GRPO-based reinforcement learning
• Benchmark Name: Omni-RewardBench, Data Number: 3725, Evaluation Metric: accuracy
Oct 2025
Omni-Reward: Towards Generalist Omni-Modal Reward Modeling with Free-Form Preferences
[](http://arxiv.org/pdf/2510.23451)
[](https://github.com/HongbangYuan/OmniReward)
[](https://hf.co/datasets/HongbangYuan/OmniRewardBench)
• Affiliation: School of Artificial Intelligence, University of Chinese Academy of Sciences
• Method Name: Omni-RewardModel-BT, Base Model: MiniCPM-o-2.6, Strategy: Bradley-Terry
• Method Name: Omni-RewardModel-R1, Base Model: Qwen2.5-VL-7B-Instruct, Strategy: GRPO
• Benchmark Name: Omni-RewardBench, Data Number: 3725, Evaluation Metric: accuracy
Oct 2025
LongCat-Video Technical Report
[](http://arxiv.org/pdf/2510.22200)
[](https://github.com/meituan-longcat/LongCat-Video)
• Affiliation: Meituan
• Method Name: LongCat-Video, Base Model: , Strategy: GRPO
Oct 2025
LongCat-Video Technical Report
[](http://arxiv.org/pdf/2510.22200)
[](https://github.com/meituan-longcat/LongCat-Video)
• Affiliation: Meituan
• Method Name: LongCat-Video, Base Model: WAN2.1 VAE, Strategy: GRPO
Oct 2025
Epipolar Geometry Improves Video Generation Models
[](http://arxiv.org/pdf/2510.21615)
[](https://epipolar-dpo.github.io)
• Affiliation: University of Oxford
• Method Name: Epipolar-DPO, Base Model: Wan-2.1, Strategy: DPO
• Benchmark Name: large dataset of over 162,000 generated videos annotated with 3D scene consistency metrics, Data Number: 162000, Evaluation Metric: Sampson epipolar error
Oct 2025
RealDPO: Real or Not Real, that is the Preference
[](http://arxiv.org/pdf/2510.14955)
[](Vchitect.github.io/RealDPO-Project)
• Affiliation: University of Electronic Science and Technology of China
• Method Name: RealDPO, Base Model: CogVideoX-5B, Strategy: DPO
• Benchmark Name: RealAction-5K, Data Number: 5000, Evaluation Metric: Visual Alignment, Text Alignment, Motion Quality, Human Quality
Oct 2025
RealDPO: Real or Not Real, that is the Preference
[](http://arxiv.org/pdf/2510.14955)
[](Vchitect.github.io/RealDPO-Project)
• Affiliation: University of Electronic Science and Technology of China
• Method Name: RealDPO, Base Model: CogVideoX-5B, Strategy: DPO
• Benchmark Name: RealAction-5K, Data Number: 5000, Evaluation Metric: Visual Alignment, Text Alignment, Motion Quality, Human Quality
Oct 2025
ImagerySearch: Adaptive Test-Time Search for Video Generation Beyond Semantic Dependency Constraints
[](http://arxiv.org/pdf/2510.14847)
• Affiliation: UCAS
• Method Name: ImagerySearch, Base Model: Wan2.1, Strategy: adaptive test-time search strategy
• Benchmark Name: LDT-Bench, Data Number: 2839, Evaluation Metric: ImageryQA
Oct 2025
Identity-GRPO: Optimizing Multi-Human Identity-preserving Video Generation via Reinforcement Learning
[](http://arxiv.org/pdf/2510.14256)
[](Identity-GRPO)
• Affiliation: Alibaba Group
• Method Name: Identity-GRPO, Base Model: Qwen2.5-VL-3B, Strategy: GRPO
• Benchmark Name: multi-human identity-preserving preference benchmark, Data Number: 500, Evaluation Metric: Accuracy
Oct 2025
Identity-Preserving Image-to-Video Generation via Reward-Guided Optimization
[](http://arxiv.org/pdf/2510.14255)
[](https://ipro-alimama.github.io/)
• Affiliation: Taobao & Tmall Group of Alibaba
• Method Name: IPRO, Base Model: Wan 2.2 I2V, Strategy: reward-guided optimization with KL-divergence regularization and facial scoring mechanism
Oct 2025
PhysMaster: Mastering Physical Representation for Video Generation via Reinforcement Learning
[](http://arxiv.org/pdf/2510.13809)
[](https://sihuiji.github.io/PhysMaster-Page/)
• Affiliation: The University of Hong Kong
• Method Name: PhysMaster, Base Model: , Strategy: DPO
Oct 2025
VIST3A: Text-to-3D by Stitching a Multi-view Reconstruction Network to a Video Generator
[](http://arxiv.org/pdf/2510.13454)
[](https://gohyojun15.github.io/VIST3A/)
• Affiliation: ETH Zurich
• Method Name: VIST3A, Base Model: , Strategy: direct reward finetuning
Oct 2025
VIST3A: Text-to-3D by Stitching a Multi-view Reconstruction Network to a Video Generator
[](http://arxiv.org/pdf/2510.13454)
[](https://gohyojun15.github.io/VIST3A/)
• Affiliation: ETH Zurich
• Method Name: VIST3A, Base Model: , Strategy: direct reward finetuning
Oct 2025
Playmate2: Training-Free Multi-Character Audio-Driven Animation via Diffusion Transformer with Reward Feedback
[](http://arxiv.org/pdf/2510.12089)
[](https://playmate111.github.io/Playmate2/)
• Affiliation: Guangzhou Quwan Network Technology
• Method Name: Mask-CFG, Base Model: Wan2.1, Strategy: DPO
Oct 2025
VR-Thinker: Boosting Video Reward Models through Thinking-with-Image Reasoning
[](http://arxiv.org/pdf/2510.10518)
[](https://github.com/qunzhongwang/vr-thinker)
• Affiliation: CUHK MMLab
• Method Name: VR-Thinker, Base Model: Qwen2.5-VL-7B, Strategy: GRPO
Oct 2025
AVoCaDO: An Audiovisual Video Captioner Driven by Temporal Orchestration
[](http://arxiv.org/pdf/2510.10395)
[](https://avocado-captioner.github.io/)
• Affiliation: Kling Team, Kuaishou Technology
• Method Name: AVoCaDO GRPO, Base Model: Qwen2.5-Omni-7B, Strategy: GRPO
Oct 2025
iMoWM: Taming Interactive Multi-Modal World Model for Robotic Manipulation
[](http://arxiv.org/pdf/2510.09036)
[](https://xingyoujun.github.io/imowm/)
• Affiliation: Nanyang Technological University, Singapore
• Method Name: iMoWM, Base Model: , Strategy: model-based RL with DrQ-v2
Oct 2025
Real-Time Motion-Controllable Autoregressive Video Diffusion
[](http://arxiv.org/pdf/2510.08131)
[](https://kesenzhao.github.io/AR-Drag.github.io/)
• Affiliation: Nanyang Technological University
• Method Name: AR-Drag, Base Model: Wan2.1-1.3B, Strategy: GRPO
Oct 2025
Real-Time Motion-Controllable Autoregressive Video Diffusion
[](http://arxiv.org/pdf/2510.08131)
[](https://kesenzhao.github.io/AR-Drag.github.io/)
• Affiliation: Nanyang Technological University
• Method Name: AR-Drag, Base Model: Wan2.1-1.3B, Strategy: GRPO
• Benchmark Name: motion controllability benchmark, Data Number: 206, Evaluation Metric: Motion Consistency
Oct 2025
Presenting a Paper is an Art: Self-Improvement Aesthetic Agents for Academic Presentations
[](http://arxiv.org/pdf/2510.05571)
[](https://evopresent.github.io/)
• Affiliation: University of California, Santa Barbara
• Method Name: PresAesth, Base Model: Qwen-2.5-VL-7B, Strategy: GRPO
• Benchmark Name: EvoPresent Benchmark, Data Number: 650 papers, 2000 slide pairs, Evaluation Metric: Perplexity, ROUGE-L, Layout Balance, Aesthetic Scores, MAE, F1-score, Accuracy
Oct 2025
OpusAnimation: Code-Based Dynamic Chart Generation
[](http://arxiv.org/pdf/2510.03341)
• Affiliation: Opus AI Research, Brown University
• Method Name: Joint-Code-Visual Reward based Group Relative Policy Optimization (JCVR-GRPO), Base Model: Qwen2.5-VL-3B, Strategy: GRPO
• Benchmark Name: DCG-Bench, Data Number: 700, Evaluation Metric: Execution Pass Rate, QA-based Scores
Oct 2025
OpusAnimation: Code-Based Dynamic Chart Generation
[](http://arxiv.org/pdf/2510.03341)
• Affiliation: Opus AI Research, Brown University
• Method Name: JCVR-GRPO, Base Model: Qwen2.5-VL-3B, Strategy: GRPO
• Benchmark Name: DCG-Bench, Data Number: 700, Evaluation Metric: Execution Pass Rate, QA-based Scores
Oct 2025
MultiModal Action Conditioned Video Generation
[](http://arxiv.org/pdf/2510.02287)
[](https://people.csail.mit.edu/yichenl/projects/multimodalvideo/)
• Affiliation: MIT CSAIL
• Method Name: MultiModal Action Conditioned Video Generation, Base Model: I2VGen, Strategy: Video diffusion model with multimodal action conditioning and feature regularization
Oct 2025
MultiModal Action Conditioned Video Generation
[](http://arxiv.org/pdf/2510.02287)
[](https://people.csail.mit.edu/yichenl/projects/multimodalvideo/)
• Affiliation: MIT CSAIL
• Method Name: MultiModal Action Conditioned Video Generation, Base Model: , Strategy: Latent space projection and regularization with diffusion-based video generation
Oct 2025
Self-Forcing++: Towards Minute-Scale High-Quality Video Generation
[](http://arxiv.org/pdf/2510.02283)
[](https://self-forcing-plus-plus.github.io/)
• Affiliation: UCLA
• Method Name: Self-Forcing++, Base Model: Wan2.1-T2V-1.3B, Strategy: GRPO (Group Relative Policy Optimization)
Oct 2025
VidGuard-R1: AI-Generated Video Detection and Explanation via Reasoning MLLMs and RL
[](http://arxiv.org/pdf/2510.02282)
[](https://VidGuard-R1.github.io)
• Affiliation: Department of Computer Science, University of Texas at Austin, Austin, TX, USA
• Method Name: VidGuard-R1, Base Model: Qwen2.5-VL-7B, Strategy: GRPO
• Benchmark Name: VidGuard-R1-CoT-30k, Data Number: 30000, Evaluation Metric: Top-1 accuracy
• Benchmark Name: VidGuard-R1-RL-100k, Data Number: 100000, Evaluation Metric: Top-1 accuracy
Oct 2025
InfoMosaic-Bench: Evaluating Multi-Source Information Seeking in Tool-Augmented Agents
[](http://arxiv.org/pdf/2510.02271)
[](https://github.com/DorothyDUUU/Info-Mosaic)
• Affiliation: Shanghai Jiao Tong University
• Benchmark Name: InfoMosaic-Bench, Data Number: 621, Evaluation Metric: Accuracy, Pass Rate
Oct 2025
Poolformer: Recurrent Networks with Pooling for Long-Sequence Modeling
[](http://arxiv.org/pdf/2510.02206)
• Affiliation: nan
• Method Name: Poolformer, Base Model: , Strategy: Recurrent neural networks with pooling operations for long-sequence modeling
Oct 2025
EvoStruggle: A Dataset Capturing the Evolution of Struggle across Activities and Skill Levels
[](http://arxiv.org/pdf/2510.01362)
[](https://github.com/FELIXFENG2019/EvoStruggle)
• Affiliation: University of Bristol
• Benchmark Name: EvoStruggle, Data Number: 2793 videos, 5385 annotated temporal struggle segments, Evaluation Metric: mAP at different IoU thresholds (0.3, 0.5, 0.7)
Oct 2025
LVTINO: LAtent Video consisTency INverse sOlver for High Definition Video Restoration
[](http://arxiv.org/pdf/2510.01339)
• Affiliation: Laboratoire MAP5, UMR 8145, Université Paris Cité, CNRS
• Method Name: LATINO, Base Model: , Strategy: Bayesian Langevin posterior sampling with Video Consistency Models (VCMs) and Image Consistency Models (ICMs)
Oct 2025
LVTINO: LAtent Video consisTency INverse sOlver for High Definition Video Restoration
[](http://arxiv.org/pdf/2510.01339)
• Affiliation: Laboratoire MAP5, UMR 8145, Université Paris Cité, CNRS
• Method Name: LATINO, Base Model: , Strategy: Langevin posterior sampling with stochastic auto-encoder steps
Oct 2025
EvoWorld: Evolving Panoramic World Generation with Explicit 3D Memory
[](http://arxiv.org/pdf/2510.01183)
• Affiliation: Johns Hopkins University
• Benchmark Name: Spatial360, Data Number: 58000+, Evaluation Metric: FVD, LMSE, LPIPS, PSNR, SSIM, MEt3R, AUC@30
Sep 2025
Stable Cinemetrics : Structured Taxonomy and Evaluation for Professional Video Generation
[](http://arxiv.org/pdf/2509.26555)
[](https://stable-cinemetrics.github.io/)
• Affiliation: Stability AI
• Benchmark Name: StableCinemetrics, Data Number: 20K videos, Evaluation Metric: human evaluation (1-5 scale)
Sep 2025
How Far Do Time Series Foundation Models Paint the Landscape of Real-World Benchmarks ?
[](http://arxiv.org/pdf/2509.26347)
[](https://anonymous.4open.science/r/benchmarking_nature_tsfm-D602)
• Affiliation: University of Luxembourg
• Benchmark Name: REAL-V-TSFM, Data Number: 6130, Evaluation Metric: MAPE, sMAPE, Agg. Relative WQL, Agg. Relative MASE
Sep 2025
V-HUB: A Visual-Centric Humor Understanding Benchmark for Video LLMs
[](http://arxiv.org/pdf/2509.25773)
• Affiliation: Shanghai Jiao Tong University
• Benchmark Name: v-HUB, Data Number: 960, Evaluation Metric: BERTScore, SentBERT, METEOR
Sep 2025
Visual Jigsaw Post-Training Improves MLLMs
[](http://arxiv.org/pdf/2509.25190)
[](https://penghao-wu.github.io/visual_jigsaw/)
• Affiliation: S-Lab, Nanyang Technological University
• Method Name: Visual Jigsaw, Base Model: Qwen2.5-VL-7B-Instruct, Strategy: GRPO
Sep 2025
FlashI2V: Fourier-Guided Latent Shifting Prevents Conditional Image Leakage in Image-to-Video Generation
[](http://arxiv.org/pdf/2509.25187)
[](https://pku-yuangroup.github.io/FlashI2V/)
• Affiliation: Peking University, Shenzhen Graduate School
• Method Name: FlashI2V, Base Model: , Strategy: Flow Matching (FM) with Fourier-Guided Latent Shifting
Sep 2025
World-Env: Leveraging World Model as a Virtual Environment for VLA Post-Training
[](http://arxiv.org/pdf/2509.24948)
• Affiliation: School of Computer Science and Engineering, Sun Yat-sen University, China
• Method Name: World-Env, Base Model: OpenVLA-OFT, Strategy: PPO
Sep 2025
Fidelity-Aware Data Composition for Robust Robot Generalization
[](http://arxiv.org/pdf/2509.24797)
• Affiliation: UCAS-Terminus AI Lab, University of Chinese Academy of Sciences
• Method Name: Coherent Information Fidelity Tuning (CIFT), Base Model: Cosmos-Predict2-2B-Video2World, Strategy: Feature-Space Signal-to-Noise Ratio optimization for data composition
• Method Name: Multi-View Video Augmentation (MV Aug), Base Model: Cosmos-Predict2-2B-Video2World, Strategy: Latent diffusion transformer with periodic cross-view attention for video-to-video synthesis
Sep 2025
IWR-Bench: Can LVLMs reconstruct interactive webpage from a user interaction video?
[](http://arxiv.org/pdf/2509.24709)
[](https://github.com/L-O-I/IWR-Bench)
• Affiliation: Shanghai AI Lab, Zhejiang University
• Benchmark Name: IWR-Bench, Data Number: 113, Evaluation Metric: Interactive Functionality Score (IFS) and Visual Fidelity Score (VFS)
Sep 2025
Can you SPLICE it together? A Human Curated Benchmark for Probing Visual Reasoning in VLMs
[](http://arxiv.org/pdf/2509.24640)
[](https://huggingface.co/datasets/prokajevo/splice-benchmark)
• Affiliation: Institute of Cognitive Science, Osnabrück University, Osnabrück, Germany
• Benchmark Name: SPLICE, Data Number: 3381, Evaluation Metric: Binary Accuracy, Hamming Accuracy, Longest Common Subsequence, Edit Distance
Sep 2025
PoseDiff: A Unified Diffusion Model Bridging Robot Pose Estimation and Video-to-Action Control
[](http://arxiv.org/pdf/2509.24591)
[](https://haozhuo-zhang.github.io/PoseDiff-project-page/)
• Affiliation: The University of Manchester
• Method Name: PoseDiff, Base Model: , Strategy: DDPM (Denoising Diffusion Probabilistic Model)
Sep 2025
NeMo: Needle in a Montage for Video-Language Understanding
[](http://arxiv.org/pdf/2509.24563)
[](https://lavi-lab.github.io/NeMoBench/)
• Affiliation: The Chinese University of Hong Kong
• Benchmark Name: NeMoBench, Data Number: 31,378, Evaluation Metric: Recall@1x, tIoU=0.7, Recall@1x, tIoU=0.5, Average mAP
Sep 2025
Training Agents Inside of Scalable World Models
[](http://arxiv.org/pdf/2509.24527)
[](danijar.com/dreamer4)
• Affiliation: Google DeepMind
• Method Name: Dreamer 4, Base Model: , Strategy: PMPO (Preference optimization as probabilistic inference) with task-conditioned policy and reward modeling
Sep 2025
Rethinking JEPA: Compute-Efficient Video SSL with Frozen Teachers
[](http://arxiv.org/pdf/2509.24317)
• Affiliation: Apple
• Method Name: SALT (Static-teacher Asymmetric Latent Training), Base Model: , Strategy: Two-stage self-supervised learning with frozen teacher for video representation learning
Sep 2025
Reinforcement Learning with Inverse Rewards for World Model Post-training
[](http://arxiv.org/pdf/2509.23958)
• Affiliation: Microsoft Research
• Method Name: Reinforcement Learning with Inverse Rewards (RLIR), Base Model: , Strategy: Group Relative Policy Optimization (GRPO)
Sep 2025
AssemblyHands-X: Modeling 3D Hand-Body Coordination for Understanding Bimanual Human Activities
[](http://arxiv.org/pdf/2509.23888)
• Affiliation: The University of Tokyo, Tokyo, Japan
• Benchmark Name: AssemblyHands-X, Data Number: , Evaluation Metric:
Sep 2025
ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis
[](http://arxiv.org/pdf/2509.23652)
• Affiliation: Alibaba Group
• Method Name: ReWatch-R1, Base Model: Qwen2.5-VL-7B, Strategy: GRPO (Group Relative Policy Optimization)
Sep 2025
WorldSplat: Gaussian-Centric Feed-Forward 4D Scene Generation for Autonomous Driving
[](http://arxiv.org/pdf/2509.23402)
[](https://wm-research.github.io/worldsplat/)
• Affiliation: Nankai University
Sep 2025
VideoScore2: Think before You Score in Generative Video Evaluation
[](http://arxiv.org/pdf/2509.22799)
[](https://TIGER-AI-Lab.github.io/VideoScore2/)
• Affiliation: University of Illinois Urbana-Champaign
• Method Name: VIDEOSCORE2, Base Model: Qwen2.5-VL-7B-Instruct, Strategy: Group Relative Policy Optimization (GRPO)
• Benchmark Name: VIDEOSCORE-BENCH-V2, Data Number: 500, Evaluation Metric: Accuracy, Relaxed Accuracy, PLCC
Sep 2025
Learning Human-Perceived Fakeness in AI-Generated Videos via Multimodal LLMs
[](http://arxiv.org/pdf/2509.22646)
[](https://deeptracereward.github.io/)
• Affiliation: Princeton University
• Benchmark Name: DEEPTRACEREWARD, Data Number: 4334, Evaluation Metric: Accuracy, Explanation score, BBox IoU, BBox Distance, Time Distance
Sep 2025
WoW: Towards a World omniscient World model Through Embodied Interaction
[](http://arxiv.org/pdf/2509.22642)
[](wow-world-model.github.io)
• Affiliation: Beijing Innovation Center of Humanoid Robotics
• Method Name: WoW, Base Model: Cosmos2, Strategy: GRPO
• Method Name: SOPHIA, Base Model: , Strategy: Self-optimizing framework with critic-refiner loop
• Benchmark Name: WoWBench, Data Number: 606, Evaluation Metric: FVD, SSIM, PSNR, DINO, Dreamsim, Mask-guided Regional Consistency, Instruction Understanding, Physical common sense, Planning and Task Decomposition
Sep 2025
Drag4D: Align Your Motion with Text-Driven 3D Scene Generation
[](http://arxiv.org/pdf/2509.21888)
• Affiliation: KAIST
• Method Name: Local-Global DragAnything, Base Model: , Strategy: Motion-conditioned video diffusion with part-augmented trajectory guidance
• Benchmark Name: Drag4D-30, Data Number: 30, Evaluation Metric: CLIP-Score, Sharp, Colorful, Quality, PSNR, SSIM
Sep 2025
StableDub: Taming Diffusion Prior for Generalized and Efficient Visual Dubbing
[](http://arxiv.org/pdf/2509.21887)
[](https://stabledub.github.io)
• Affiliation: nan
• Method Name: StableDub, Base Model: , Strategy: Diffusion-based visual dubbing with lip-habit-modulated mechanism and occlusion-aware training strategy
Sep 2025
DiTraj: training-free trajectory control for video diffusion transformer
[](http://arxiv.org/pdf/2509.21839)
[](https://xduzhangjiayu.github.io/DiTraj_Project_Page/)
• Affiliation: Beijing University of Posts and Telecommunications
• Method Name: DiTraj, Base Model: Wan2.1, CogVideoX, Strategy: Foreground-background separation guidance and STD-RoPE position embedding modification
Sep 2025
Can AI Perceive Physical Danger and Intervene?
[](http://arxiv.org/pdf/2509.21651)
[](https://asimov-benchmark.github.io/v2)
• Affiliation: Google DeepMind Robotics
• Benchmark Name: ASIMOV-2.0, Data Number: 319, Evaluation Metric: Latent risk accuracy, Latent risk severity accuracy, Action effect accuracy, Activated risk accuracy
• Benchmark Name: ASIMOV-2.0-Video, Data Number: 287, Evaluation Metric: Injury risk accuracy, Latent risk and severity accuracy, Last intervention timestamp MAE, Intervention rate
• Benchmark Name: ASIMOV-2.0-Constraints, Data Number: 164, Evaluation Metric: Constraint violation rate
Sep 2025
VideoJudge: Bootstrapping Enables Scalable Supervision of MLLM-as-a-Judge for Video Understanding
[](http://arxiv.org/pdf/2509.21451)
• Affiliati