PhyEvolve: Multi-Agent State-Evolution Planning for Physics-Aware Video Generation

The curtain only changes what the camera sees. The foam keeps expanding while hidden, so when the curtain rises the scene must reveal the state it has reached — not restart.

Four methods in each row: Minimax-H3 (original prompt), NEWTON, Co-Director and PhyEvolve. Each row has its own shared playback controls.

Two events per row. Each pair compares Minimax-H3 on the original prompt with PhyEvolve. Events are distinct from the comparisons above.

NEWTON, Co-Director and PhyEvolve all run on the same frozen Minimax-H3 generator; PhyEvolve only replaces the prompt. All scores are percentages, higher is better, and the best value in each column is bold.

Main results

Detailed results

Table 6: Component ablation on StEvo-Bench (%).

Variant	Evo. Success ↑	Prog. ↑	Phys. ↑	Coh. ↑	Obs. Ctrl. ↑	Act. Ctrl. ↑
Baseline (original prompt)	17.7	25.3	70.6	88.2	87.6	81.5
w/o Director	45.3	51.2	93.7	94.6	82.4	82.4
w/o Physical Evolver	29.0	38.0	72.3	94.5	88.4	78.6
w/o Composer	44.3	51.5	93.7	94.6	89.3	84.9
PhyEvolve	45.7	51.2	96.8	94.6	89.5	85.2

Table 7: Agent-backbone comparison on PhyGenBench (%).

Agent backbone	Mechanics ↑	Optics ↑	Thermal ↑	Material ↑	Average ↑
InternVL3.5-30B	71.7	87.3	76.7	70.8	77.3
Qwen3.5-9B	67.5	84.0	81.1	75.8	77.3
Qwen3-VL-32B	73.3	84.7	82.2	77.5	79.6

Copy@misc{zhou2026phyevolvemultiagentstateevolutionplanning,
  title         = {PhyEvolve: Multi-Agent State-Evolution Planning for Physics-Aware Video Generation},
  author        = {Yang-Hao Zhou and Lixing Tan and Zhiqi Li and Guang Yang and Xiao Yang},
  year          = {2026},
  eprint        = {2610.30717},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.30717}
}