Multi-Agent Path Finding on MAPF-FrozenLake 4-agent Sum
33.14Valid RateQwen3-4B + GRPO + LLM-as-Environment-Engineer
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Qwen3-4B + GRPO + LLM-as-Environment-EngineerBackbone=Qwen3-4B, Training=GRPO, Environment Design=Ours2026.06 | 33.14 | 21.33 | |
| Kimi-K2.52026.06 | 26.95 | 17.9 | |
| Qwen3-4B + GRPO (random)Backbone=Qwen3-4B, Training=GRPO, Config=single randomly sampled Round-02026.06 | 26.67 | 16.1 | |
| Grok-4.22026.06 | 23.43 | 15.33 | |
| GPT-5.42026.06 | 17.05 | 10.86 | |
| Gemini-3.1-Pro2026.06 | 12.95 | 10.1 | |
| Qwen3-4B (base)Backbone=Qwen3-4B, Status=untrained2026.06 | 3.43 | 3.05 |