Mathematical Reasoning on Minerva (Pass@1 (Avg@16), Tokens)
41.7Pass@1 (Avg@16)Mixed
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| MixedBackbone=Qwen3-8B-Base2026.01 | 41.7 | — | — | |
| Ours (Shorten Prefix)Backbone=Qwen3-4B (non-think)2026.01 | 41.7 | — | — | |
| GRPOBackbone=Qwen3-8B-Base2026.01 | 41.5 | — | — | |
| RTSBackbone=Qwen3-4B (non-think)2026.01 | 41.3 | — | — | |
| MixedBackbone=Qwen3-4B (non-think)2026.01 | 40.8 | — | — | |
| GRPOBackbone=Qwen3-4B (non-think)2026.01 | 40.6 | — | — | |
| Ours (Shorten Prefix)Backbone=Qwen3-8B-Base2026.01 | 40.4 | — | — | |
| DLER-R1-7B-ResearchBackbone=DeepSeek-R1-Distill-Qwen-7B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 39.5 | 1,798 | 0.85 | |
| RTSBackbone=Qwen3-8B-Base2026.01 | 39.5 | — | — | |
| Ours (Shorten Prefix)Backbone=Qwen3-4B-Base2026.01 | 39.1 | — | — | |
| L1-Qwen-7B-MaxBackbone=DeepSeek-R1-Distill-Qwen-7B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 38.9 | 2,120 | 0.73 | |
| Laser-DE-L4096-7BBackbone=DeepSeek-R1-Distill-Qwen-7B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 38.9 | 1,850 | 0.79 | |
| MixedBackbone=Qwen3-4B-Base2026.01 | 38.5 | — | — | |
| APR-7B (Ours)Backbone=DeepSeek-R1-Distill-Qwen-7B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 38.4 | 1,647 | 0.79 | |
| GRPOBackbone=Qwen3-4B-Base2026.01 | 38.3 | — | — | |
| BaseBackbone=Qwen3-4B (non-think)2026.01 | 38.1 | — | — | |
| TrainingEfficient DS-7BBackbone=DeepSeek-R1-Distill-Qwen-7B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 37.7 | 2,940 | 0.44 | |
| RTSBackbone=Qwen3-4B-Base2026.01 | 36.9 | — | — | |
| Original ModelBackbone=DeepSeek-R1-Distill-Qwen-7B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 36.2 | 4,266 | — | |
| AdaptThink-7B-delta0.05Backbone=DeepSeek-R1-Distill-Qwen-7B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 35.1 | 2,570 | 0.25 | |
| SB DS7B alpha 2Backbone=DeepSeek-R1-Distill-Qwen-7B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 31.6 | 901 | 0.15 | |
| DLER-R1-1.5B-ResearchBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 31.5 | 2,225 | 0.57 | |
| Laser-L8192-1.5BBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 31.1 | 3,638 | 0.19 | |
| Original ModelBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 30.3 | 4,082 | — | |
| APR-1.5B (Ours)Backbone=DeepSeek-R1-Distill-Qwen-1.5B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 30.2 | 1,985 | 0.5 | |
| L1-Qwen-1.5B-MaxBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 29.4 | 2,779 | 0.17 | |
| Laser-DE-L4096-1.5BBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 29 | 2,289 | 0.22 | |
| DS-1.5B-thinkprune-iter2kBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 27.6 | 2,118 | 0.04 | |
| TrainingEfficient DS-1.5BBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 27 | 3,018 | -0.28 | |
| AdaptThink-1.5B-delta0.05Backbone=DeepSeek-R1-Distill-Qwen-1.5B, Sampling Temperature=0.6, Max Response Length=8192, Rollout samples=162026.01 | 24.7 | 1,637 | -0.33 | |
| BaseBackbone=Qwen3-8B-Base2026.01 | 23.7 | — | — | |
| BaseBackbone=Qwen3-4B-Base2026.01 | 21.6 | — | — |