Mathematical Reasoning on AIME24, AIME25, AMC23, MATH, Olympiad (Aggregate)
47.3Average ScoreGRPO + coupled rhythm credit
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPO + coupled rhythm creditBackbone=Qwen3-4B-Base, Context Length=8K2025.10 | 47.3 | |
| GRPO + local-chunk creditBackbone=Qwen3-4B-Base, Context Length=8K2025.10 | 47.2 | |
| GRPO + global-anchor creditBackbone=Qwen3-4B-Base, Context Length=8K2025.10 | 47.1 | |
| GRPOBackbone=Qwen3-4B-Base, Context Length=8K2025.10 | 44.8 | |
| GRPO + random creditBackbone=Qwen3-4B-Base, Context Length=8K2025.10 | 44.7 | |
| GRPO + high-entropy creditBackbone=Qwen3-4B-Base, Context Length=8K2025.10 | 44.5 | |
| GRPO + coupled rhythm creditBackbone=Qwen3-8B-Base, Context Length=1K2025.10 | 43.2 | |
| GRPO + global-anchor creditBackbone=Qwen3-8B-Base, Context Length=1K2025.10 | 42.1 | |
| GRPO + local-chunk creditBackbone=Qwen3-8B-Base, Context Length=1K2025.10 | 41.3 | |
| GRPO + high-entropy creditBackbone=Qwen3-8B-Base, Context Length=1K2025.10 | 40.1 | |
| GRPO + random creditBackbone=Qwen3-8B-Base, Context Length=1K2025.10 | 39.6 | |
| GRPOBackbone=Qwen3-8B-Base, Context Length=1K2025.10 | 39.4 | |
| GRPO + coupled rhythm creditBackbone=Qwen3-4B-Base, Context Length=1K2025.10 | 39.2 | |
| GRPO + local-chunk creditBackbone=Qwen3-4B-Base, Context Length=1K2025.10 | 38.6 | |
| GRPO + global-anchor creditBackbone=Qwen3-4B-Base, Context Length=1K2025.10 | 38.2 | |
| GRPO + high-entropy creditBackbone=Qwen3-4B-Base, Context Length=1K2025.10 | 37.2 | |
| GRPOBackbone=Qwen3-4B-Base, Context Length=1K2025.10 | 37.1 | |
| GRPO + random creditBackbone=Qwen3-4B-Base, Context Length=1K2025.10 | 37.1 |