Mathematical Reasoning on AIME25 (pass@32, Overall)
92.6pass@32EqLen-GRPO
Evaluation results
| Method | Links | ||
|---|---|---|---|
| EqLen-GRPOBackbone=Qwen3-Next2026.04 | 92.6 | — | |
| LUSPOBackbone=Qwen3-Next2026.04 | 89.8 | — | |
| DAPOBackbone=Qwen3-Next2026.04 | 89.5 | — | |
| Dr. GRPOBackbone=Qwen3-Next2026.04 | 89.1 | — | |
| GSPOBackbone=Qwen3-Next2026.04 | 88.4 | — | |
| GRPOBackbone=Qwen3-Next2026.04 | 88 | — | |
| BaseBackbone=Qwen3-Next2026.04 | 87.8 | — | |
| EqLen-GRPOBackbone=Qwen3-32B2026.04 | 84.2 | — | |
| LUSPOBackbone=Qwen3-32B2026.04 | 78.8 | — | |
| DAPOBackbone=Qwen3-32B2026.04 | 78.2 | — | |
| Dr. GRPOBackbone=Qwen3-32B2026.04 | 77.4 | — | |
| GSPOBackbone=Qwen3-32B2026.04 | 76.6 | — | |
| GRPOBackbone=Qwen3-32B2026.04 | 75.8 | — | |
| BaseBackbone=Qwen3-32B2026.04 | 72.9 | — | |
| VM-AV-GRPOAlgorithm=VM-AV-GRPO, Backbone=Qwen2.5-Math-1.5B2026.03 | 43.3 | 52.5 | |
| MSA-GRPOAlgorithm=MSA-GRPO, Backbone=Qwen2.5-Math-1.5B2026.03 | 43.3 | 50.7 | |
| AV-GRPOAlgorithm=AV-GRPO, Backbone=Qwen2.5-Math-1.5B2026.03 | 36.7 | 50.9 | |
| FA-GRPOAlgorithm=FA-GRPO, Backbone=Qwen2.5-Math-1.5B2026.03 | 36.7 | 49.9 | |
| PR-GRPOAlgorithm=PR-GRPO, Backbone=Qwen2.5-Math-1.5B2026.03 | 36.7 | 49.4 | |
| SVE-LNA-GRPOAlgorithm=SVE-LNA-GRPO, Backbone=Qwen2.5-Math-1.5B2026.03 | 30 | 48.7 | |
| CDA-GRPOAlgorithm=CDA-GRPO, Backbone=Qwen2.5-Math-1.5B2026.03 | 30 | 49 | |
| GRPOAlgorithm=GRPO, Backbone=Qwen2.5-Math-1.5B2026.03 | 26.7 | 47.8 | |
| AISModel=Qwen3-8B, Bitwidth=FP82026.05 | — | 43.33 | |
| AISModel=Qwen3.5-9B, Bitwidth=FP82026.05 | — | 73.33 | |
| FlashRL (TIS)Model=Qwen3-8B, Bitwidth=FP82026.05 | — | 34.23 | |
| FlashRL (TIS)Model=Qwen3.5-9B, Bitwidth=FP82026.05 | — | 66.67 | |
| RL (BF16)Model=Qwen3-8B, Bitwidth=BF162026.05 | — | 36.7 | |
| RL (BF16)Model=Qwen3.5-9B, Bitwidth=BF162026.05 | — | 69.5 | |
| RL (FP8 Rollout)Model=Qwen3-8B, Bitwidth=FP82026.05 | — | 26.81 | |
| RL (FP8 Rollout)Model=Qwen3.5-9B, Bitwidth=FP82026.05 | — | 60 |