NTH

PhysLoop: Harnessing LLMs' Physics Awareness for Self-Improving Video Generation Models

AuthorsYuxuan Chen, Xiaojie Li, Hao Chen, Juncheng Ma, Kewei Zhang, Zhongjing Du, Guoqiang Wei, Fangfang Wang, Zhen Dong, Zhiyang Dou, Liming Zhao, Yan Zeng, Jianwei Zhang, Daquan Zhou

Affiliations[

November 16, 2026 2 min read
Watch on YouTube
The one-line take

PhysLoop uses LLMs’ physics knowledge to help video generators create and repair more physically consistent videos.

Key results

100K
PhysLoop-100K accepted videos

Training dataset of physically verified videos.

2.5K
Distinct physical events

Events represented in PhysLoop-100K.

85.00%
Generation success rate

PhysLoop rate; Seedance 2.5 alone reaches 73.00%.

86.00%
Repair success rate

PhysLoop rate; the strongest standalone baseline reaches 19.33%.

52.06
Physics-IQ Verified score

PhysLoop score; Seedance 2.5 scores 48.07.

What the paper found

PhysLoop uses large language models’ physics knowledge to build and repair videos through a simulation-and-verification loop. An agent turns a request into a scene, runs it in Unreal Engine, Taichi, or Genesis, checks physical events and object states, then uses a video model to refine appearance and verifies the result again. The system’s default configuration uses OpenAI’s GPT-6 Astra as the agent and ByteDance’s Seedance 2.5 for refinement; comparisons also include Anthropic’s Claude Opus 5 and Google’s Gemini Omni. Its PhysLoop-100K dataset contains 100K accepted videos spanning 2.5K physical events. On generation tasks, PhysLoop reaches an 85.00% success rate, versus 73.00% for Seedance 2.5 alone. For localized repair, it reaches 86.00%, compared with 19.33% for the strongest standalone baseline. On Physics-IQ Verified, it scores 52.06, versus 48.07 for Seedance 2.5. The work presents a quality-control framework and a step toward self-improving video models; it does not demonstrate a complete repeated cycle of model retraining.

Original abstract

Video generation models produce visually realistic videos but still struggle with physical dynamics and causal consistency. Learning from generated videos requires fine-grained physical supervision, yet reliable quality judgments, correction pairs, and controlled variations are difficult to obtain. Large language models (LLMs) possess physical knowledge, but struggle to reliably guide video generation models in producing their training videos. We present PhysLoop, a unified harness designed to support physically grounded self-improvement in video generation models. The harness guides and constrains LLMs throughout video generation, selection, and correction, distilling their physical knowledge into physically consistent videos. The framework supports controllable training-data construction and inference-time video repair. By identifying and correcting failures, PhysLoop marks a key milestone toward recursive self-improvement (RSI) in video generation models with limited human intervention. We release 100K videos spanning diverse, realistic indoor and outdoor scenes, along with generation tools and two evaluation datasets: PhysLoop-Gen and PhysLoop-Repair. We evaluate controllable generation and localized repair on these datasets, respectively, and assess physical behavior on the external Physics-IQ Verified benchmark.

Read the original paper

More in Generative Models

Browse all 87 papers →
01Generative Model

dNaViT: Discrete Codes are General-Purpose Visual Representations

Xiaotong Li, Zhengrong Yue, Xu Huang, Jialv Zou, Minhao Jing, Wenhao Yan, Haodong Li, Hai Huang, Bo Zhao, Yan Feng, Peng Pei, Xunliang Cai

dNaViT turns images into reusable discrete codes designed to work across visual understanding, generation, and editing.

Read analysis