General Reasoning and Coding on OOD Suite (ARC-c, GPQA, MMLU-Pro, LiveCodeBench)
96.6ARC-c (pass@1)w4a16 QaRL TBPO
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| w4a16 QaRL TBPOModel=Qwen3-30B-A3B-Base, Precision=w4a16, RL Algorithm=QaRL TBPO2026.04 | 96.6 | 48.2 | 68 | 55.4 | 67.05 | |
| BF16 GSPOModel=Qwen3-30B-A3B-Base, Precision=BF16, RL Algorithm=GSPO2026.04 | 95.2 | 50.1 | 70.3 | 55.8 | 67.8 | |
| BF16 GRPOModel=Qwen3-8B-Base, Precision=BF16, RL Algorithm=GRPO2026.04 | 93 | 46.3 | 65.1 | 45.7 | 62.5 | |
| w4a16 QaRL TBPOModel=Qwen3-8B-Base, Precision=w4a16, RL Algorithm=QaRL TBPO2026.04 | 92.3 | 45 | 63.8 | 43.4 | 61 | |
| w4a16 rollout GRPOModel=Qwen3-8B-Base, Precision=w4a16, Rollout Strategy=rollout, RL Algorithm=GRPO2026.04 | 91.6 | 43.9 | 61.5 | 41.4 | 59.5 | |
| w4a16 rollout GSPOModel=Qwen3-30B-A3B-Base, Precision=w4a16, Rollout Strategy=rollout, RL Algorithm=GSPO2026.04 | 89.3 | 42.4 | 65.3 | 47.9 | 61.2 | |
| w4a16 QaRL TBPOModel=Qwen2.5-7B-Math, Precision=w4a16, RL Algorithm=QaRL TBPO2026.04 | 81.1 | 37.1 | 45.8 | 13.9 | 44.4 | |
| BF16 GRPOModel=Qwen2.5-7B-Math, Precision=BF16, RL Algorithm=GRPO2026.04 | 80.4 | 38.2 | 46.3 | 14.6 | 44.8 | |
| w4a16 rollout GRPOModel=Qwen2.5-7B-Math, Precision=w4a16, Rollout Strategy=rollout, RL Algorithm=GRPO2026.04 | 73.9 | 31.6 | 40 | 8.7 | 38.5 | |
| Qwen2.5-7B-MathModel=Qwen2.5-7B-Math2026.04 | 62.6 | 28.5 | 32.1 | 8 | 32.8 | |
| Qwen3-30B-A3B-BaseModel=Qwen3-30B-A3B-Base2026.04 | 61.3 | 34.8 | 52.5 | 28.5 | 44.2 | |
| BF16 GRPOModel=Qwen2.5-1.5B-Math, Precision=BF16, RL Algorithm=GRPO2026.04 | 58.9 | 27.8 | 26.7 | 11.9 | 31.3 | |
| w4a16 QaRL TBPOModel=Qwen2.5-1.5B-Math, Precision=w4a16, RL Algorithm=QaRL TBPO2026.04 | 57.6 | 21.1 | 25.3 | 8.1 | 28 | |
| w4a16 rollout GRPOModel=Qwen2.5-1.5B-Math, Precision=w4a16, Rollout Strategy=rollout, RL Algorithm=GRPO2026.04 | 50.1 | 19.8 | 21.2 | 4.3 | 23.8 | |
| Qwen3-8B-BaseModel=Qwen3-8B-Base2026.04 | 44.9 | 31.2 | 49.5 | 23 | 37.1 | |
| Qwen2.5-1.5B-MathModel=Qwen2.5-1.5B-Math2026.04 | 11.7 | 12.4 | 10.4 | 2.7 | 9.3 |