Mathematical Reasoning on AIME 24/25 (Acc, Ent)
71.88AccuracyPolicy Split
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Policy SplitBackbone=Qwen3-8B2026.04 | 71.88 | 0.55 | |
| Policy SplitBackbone=Qwen3-4B2026.04 | 71.67 | 0.49 | |
| GRPO w/ Ent RegularizationBackbone=Qwen3-8B2026.04 | 71.06 | 0.48 | |
| GRPOBackbone=Qwen3-8B2026.04 | 70.83 | 0.35 | |
| GRPO w/ Ent AdvantageBackbone=Qwen3-8B2026.04 | 70.62 | 0.51 | |
| OriginalBackbone=Qwen3-8B2026.04 | 70.29 | 0.36 | |
| GRPO w/ Ent AdvantageBackbone=Qwen3-4B2026.04 | 69.79 | 0.58 | |
| GRPO w/ Forking Token OnlyBackbone=Qwen3-8B2026.04 | 69.58 | 0.35 | |
| GRPO w/ Clip-CovBackbone=Qwen3-8B2026.04 | 69.17 | 0.35 | |
| GRPO w/ Ent Driven AdvBackbone=Qwen3-8B2026.04 | 68.87 | 0.35 | |
| GRPO w/ Ent RegularizationBackbone=Qwen3-4B2026.04 | 68.83 | 0.53 | |
| Policy Split + HE promptBackbone=Qwen3-4B2026.04 | 68.12 | 0.73 | |
| GRPO w/ Forking Token OnlyBackbone=Qwen3-4B2026.04 | 67.29 | 0.4 | |
| GRPO w/ Clip-CovBackbone=Qwen3-4B2026.04 | 67.29 | 0.4 | |
| OriginalBackbone=Qwen3-4B2026.04 | 67.08 | 0.39 | |
| GRPOBackbone=Qwen3-4B2026.04 | 67.08 | 0.39 | |
| GRPO w/ Ent Driven AdvBackbone=Qwen3-4B2026.04 | 66.04 | 0.4 | |
| Policy Split + HE promptBackbone=Qwen3-8B2026.04 | 65.88 | 0.73 | |
| Policy SplitBackbone=Qwen3-1.7B2026.04 | 43.25 | 0.68 | |
| GRPO w/ Ent AdvantageBackbone=Qwen3-1.7B2026.04 | 42.92 | 0.74 | |
| GRPO w/ Ent RegularizationBackbone=Qwen3-1.7B2026.04 | 42.29 | 0.64 | |
| GRPOBackbone=Qwen3-1.7B2026.04 | 41.04 | 0.35 | |
| OriginalBackbone=Qwen3-1.7B2026.04 | 40.54 | 0.36 | |
| GRPO w/ Clip-CovBackbone=Qwen3-1.7B2026.04 | 38.33 | 0.36 | |
| GRPO w/ Forking Token OnlyBackbone=Qwen3-1.7B2026.04 | 37.92 | 0.36 | |
| GRPO w/ Ent Driven AdvBackbone=Qwen3-1.7B2026.04 | 36.46 | 0.36 | |
| Policy Split + HE promptBackbone=Qwen3-1.7B2026.04 | 35.33 | 0.78 |