# PhyEvolve adds physical planning to text-to-video prompts

> A new multi‑agent framework improves physical consistency in generated videos, boosting benchmark scores by about 8 %.

Oossa · 2026-10-23 · https://oossa.com/en/phyevolve-adds-physical-planning-to-text-to-video-prompts

Researchers led by Lixing Tan introduced PhyEvolve, a three‑agent system that rewrites text‑to‑video prompts to include explicit physical state changes. The agents – Director, Physical Evolver, and Composer – plan causal transitions and feed them to existing video generators without retraining. Tests on four major benchmarks show an average 7.9‑point gain over the previous best agentic method.

## The facts

- Improves benchmark scores by 7.9 percentage points on average
- Paper posted on arXiv on 2026‑10‑23

## Why it matters

Creators can get videos that follow realistic physics without needing new AI models.

## Sources & references

1. [PhyEvolve: Multi-Agent State-Evolution Planning for Physics-Aware Video Generation](https://arxiv.org/abs/2610.30717) – arXiv cs.CV, 2026-10-23

Last updated: 2026-10-23