Multi-turn RL navigation on Sokoban held-out (val)
52.3Success RateTSR (Beam Search)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TSR (Beam Search)Model scale=Qwen2.5-3B2026.02 | 52.3 | 152 | 3.6 | |
| TSR (Lookahead)Model scale=Qwen2.5-3B2026.02 | 49.5 | 158 | 3.8 | |
| TSR (Best-of-N)Model scale=Qwen2.5-3B2026.02 | 47.7 | 165 | 4 | |
| Instance FilteringModel scale=Qwen2.5-3B2026.02 | 43.7 | 161 | 4.1 | |
| TSR (Beam Search)Model scale=Qwen2.5-0.5B2026.02 | 38.3 | 98 | 3.8 | |
| TSR (Lookahead)Model scale=Qwen2.5-0.5B2026.02 | 36.1 | 101 | 4 | |
| TSR (Best-of-N)Model scale=Qwen2.5-0.5B2026.02 | 33.3 | 103 | 4.3 | |
| Instance FilteringModel scale=Qwen2.5-0.5B2026.02 | 29 | 105 | 4.4 | |
| Base ModelModel scale=Qwen2.5-3B2026.02 | 16 | 272 | 4.6 | |
| Base ModelModel scale=Qwen2.5-0.5B2026.02 | 8.9 | 293 | 4.7 |