AI Paper Atlas

Research map / Video / Video generation

$PhyWorldBench$: A Comprehensive Evaluation of Physical Realism in Text-to-Video Models

Jing Gu, Xian Liu, Yu Zeng, Ashwin Nagarajan, Fangrui Zhu, Daniel Hong, Yue Fan, Qianqi Yan, Kaiwen Zhou, Ming-Yu Liu, Xin Eric Wang

ICLR 2026 · Video › Video generation · approach: Compositional text-to-video diffusion

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
tell the MLLM the video is AI-generated (context-aware prompt) (ablation)CAP evaluator (GPT-o1) · PhyWorldBench evaluationROC-AUC (PC): 65.6 -> 75.1Improved
add two-step describe-then-judge chain-of-thought (ablation)CAP evaluator (GPT-o1) · PhyWorldBench evaluationROC-AUC (PC): 73.6 -> 75.1Improved
add CAP to GPT-o1 judgeGPT-o1 · PhyWorldBench evaluationROC-AUC (PC): 61.6 -> 75.1Improved
add CAP to Qwen-VL-2.0 judgeQwen-VL-2.0 · PhyWorldBench evaluationROC-AUC (PC): 59.8 -> 74.1Improved
add CAP to GPT-4o judgeGPT-4o · PhyWorldBench evaluationROC-AUC (SA): 72.1 -> 78.5Improved
fine-tune GPT-4o evaluator on 80% of annotated videosCAP with GPT-4o · PhyWorldBench evaluationPC: 73.9 -> 80.7Improved
use physics-enhanced prompt instead of event promptHunyuan 720p · PhyWorldBenchphysics-following fraction: 0.159 -> 0.198Improved
use physics-enhanced prompt instead of event promptLTX-Video · PhyWorldBenchphysics-following fraction: 0.056 -> 0.065No clear effect
use detailed narrative prompt instead of event promptHunyuan 720p · PhyWorldBenchphysics-following fraction: 0.159 -> 0.155No clear effect
use detailed narrative prompt instead of event promptCogVideoX-1.5 · PhyWorldBenchphysics-following fraction: 0.123 -> 0.168Improved
prompt erratic visual changes vs softened prompts12 T2V models · paired prompts, human evalsuccess rate: 42.1% -> 22.1%Hurt
prompt higher-complexity multi-force interactions vs simpler cases12 T2V models · paired prompts, human evalsuccess rate: 18% -> 12%Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy