Multi-turn RL navigation on FrozenLake held-out (val)
60.7Success RateTSR (Beam Search)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TSR (Beam Search)Model scale=Qwen2.5-3B2026.02 | 60.7 | 98 | 3.1 | |
| TSR (Lookahead)Model scale=Qwen2.5-3B2026.02 | 57 | 135 | 3.3 | |
| TSR (Best-of-N)Model scale=Qwen2.5-3B2026.02 | 55 | 161 | 3.5 | |
| Instance FilteringModel scale=Qwen2.5-3B2026.02 | 48.7 | 174 | 3.6 | |
| TSR (Beam Search)Model scale=Qwen2.5-0.5B2026.02 | 30 | 152 | 3.5 | |
| TSR (Lookahead)Model scale=Qwen2.5-0.5B2026.02 | 27.8 | 168 | 3.8 | |
| TSR (Best-of-N)Model scale=Qwen2.5-0.5B2026.02 | 25 | 182 | 4.2 | |
| Instance FilteringModel scale=Qwen2.5-0.5B2026.02 | 19.7 | 157 | 4.1 | |
| Base ModelModel scale=Qwen2.5-3B2026.02 | 12.7 | 225 | 3.9 | |
| Base ModelModel scale=Qwen2.5-0.5B2026.02 | 6.4 | 276 | 4.6 |