Multi-Agent Path Finding on 3-agent 5x5 (evaluation set)
62.67Valid RateQwen3-4B + GRPO + Ours
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen3-4B + GRPO + OursBackbone=Qwen3-4B, Training=GRPO + Adaptive-weight reward2026.06 | 62.67 | 35.33 | |
| Kimi-K2.52026.06 | 57.33 | 34 | |
| Qwen3-4B + GRPO (random)Backbone=Qwen3-4B, Training=GRPO (random)2026.06 | 50.67 | 29.33 | |
| GPT-5.42026.06 | 42.67 | 31.33 | |
| Gemini-3.1-Pro2026.06 | 35.33 | 20 | |
| Grok-4.22026.06 | 29.33 | 18.67 | |
| Qwen3-4B (base)Backbone=Qwen3-4B2026.06 | 18 | 16.67 |