Single-Player Game Playing on 2048, Tetris, Candy Crush, Super Mario Bros Suite
1,589Reward (2048)COS-PLAY
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| COS-PLAYBase Model=Qwen3-8B2026.04 | 1,589 | 510.9 | 648.8 | 948.9 | 924.4 | |
| GPT-5.42026.04 | 1,126.6 | 458.2 | 532.6 | 752 | 717.4 | |
| GPT-OSS-120B2026.04 | 1,029.5 | 358.1 | 334.4 | 968.5 | 672.6 | |
| CLAUDE-4.6-SONNET2026.04 | 945 | 444.2 | 328.6 | 399.5 | 529.3 | |
| GEMINI-3.1-PRO2026.04 | 813.3 | 372.7 | 334.3 | 436.8 | 489.3 | |
| SFTBase Model=Qwen3-8B, Skill Bank Configuration=W/O Skill2026.04 | 516.7 | 28.2 | 356.1 | 736.8 | 409.5 | |
| GRPOBase Model=Qwen3-8B, Skill Bank Configuration=W/O Skill2026.04 | 510 | 96.7 | 163.3 | 669.4 | 359.9 | |
| SFTBase Model=Qwen3-8B, Skill Bank Configuration=1ST Skill2026.04 | 385.5 | 35.8 | 569.6 | 871.9 | 465.7 | |
| GRPOBase Model=Qwen3-8B, Skill Bank Configuration=1ST Skill2026.04 | 152 | 93.7 | 353.8 | 621.2 | 305.2 | |
| QWEN3-8BTraining Protocol=Base2026.04 | 131 | 32 | 519.9 | 835.5 | 379.6 | |
| SFTBase Model=Qwen3-8B, Skill Bank Configuration=Final Skill2026.04 | 64.8 | 24.4 | 554.4 | 794.4 | 359.5 |