Chess Move Prediction on Lichess Puzzle Database (held-out positions)
54.5Top-1 AccuracyVPS
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VPSBackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 54.5 | 1,744 | 50.8 | |
| SFT + GRPOBackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 54.1 | 1,670 | 38.5 | |
| VPSBackbone=Qwen3-8B2026.04 | 53.6 | 1,706 | 44.1 | |
| SFT + GRPOBackbone=Qwen3-8B2026.04 | 53.4 | 1,670 | 30.2 | |
| SFT onlyBackbone=DeepSeek-R1-Distill-Llama-8B2026.04 | 21.5 | 996 | 22.6 | |
| SFT onlyBackbone=Qwen3-8B2026.04 | 21.2 | 1,160 | 20.7 |