Mathematical Reasoning on MinervaMath (Acc, Tokens, EFF)
54.04AccuracyThinkPrune(6k->3k)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ThinkPrune(6k->3k)Prompt=step, Backbone=Qwen3-8B2026.02 | 54.04 | 2,725 | 1.98 | |
| D-SFT->D-RL3k->D-RL6kPrompt=step, Backbone=Qwen3-8B2026.02 | 53.68 | 5,180 | 1.03 | |
| D-SFT->D-RL6kPrompt=step, Backbone=Qwen3-8B2026.02 | 53.68 | 5,076 | 1.05 | |
| ThinkPrune(6k)Prompt=step, Backbone=Qwen3-8B2026.02 | 53.68 | 3,822 | 1.4 | |
| OriginalPrompt=step, Backbone=Qwen3-8B2026.02 | 52.21 | 7,640 | 0.68 | |
| D-SFT->D-RL6kPrompt=adaptive, Backbone=Qwen3-8B2026.02 | 50 | 3,209 | 1.55 | |
| D-SFT->D-RL6kPrompt=draft, Backbone=Qwen3-8B2026.02 | 49.26 | 1,677 | 2.93 | |
| D-SFT->D-RL3k->D-RL6kPrompt=adaptive, Backbone=Qwen3-8B2026.02 | 48.53 | 2,979 | 1.62 | |
| REO-RL(Q-Spec)*Prompt=step, Backbone=Qwen3-8B2026.02 | 46.8 | 3,952 | 1.18 | |
| D-SFT->D-RL3k->D-RL6kPrompt=draft, Backbone=Qwen3-8B2026.02 | 46.32 | 960 | 4.82 | |
| D-SFT->D-RL3kPrompt=draft, Backbone=Qwen3-8B2026.02 | 44.49 | 1,133 | 3.92 | |
| D-SFTPrompt=draft, Backbone=Qwen3-8B2026.02 | 37.87 | 549 | 6.88 | |
| Mean Acc.Backbone Model=Qwen3-8B2025.10 | 34 | — | — | |
| Self-Cert.Backbone Model=Qwen3-8B2025.10 | 34 | — | — | |
| Loc. non-uniBackbone Model=Qwen3-8B2025.10 | 34 | — | — | |
| Loc. uniBackbone Model=Qwen3-8B2025.10 | 34 | — | — | |
| Glob. uniBackbone Model=Qwen3-8B2025.10 | 34 | — | — | |
| High Conf.Backbone Model=Qwen3-8B2025.10 | 33 | — | — | |
| Low Ent.Backbone Model=Qwen3-8B2025.10 | 33 | — | — | |
| Glob. non-uniBackbone Model=Qwen3-8B2025.10 | 33 | — | — | |
| Loc. uniBackbone Model=DS-R1-Distill-Qwen-7B2025.10 | 31 | — | — | |
| Mean Acc.Backbone Model=DS-R1-Distill-Qwen-7B2025.10 | 30 | — | — | |
| Self-Cert.Backbone Model=DS-R1-Distill-Qwen-7B2025.10 | 30 | — | — | |
| High Conf.Backbone Model=DS-R1-Distill-Qwen-7B2025.10 | 30 | — | — | |
| Low Ent.Backbone Model=DS-R1-Distill-Qwen-7B2025.10 | 30 | — | — | |
| Glob. non-uniBackbone Model=DS-R1-Distill-Qwen-7B2025.10 | 30 | — | — | |
| Glob. uniBackbone Model=DS-R1-Distill-Qwen-7B2025.10 | 28 | — | — | |
| Loc. non-uniBackbone Model=DS-R1-Distill-Qwen-7B2025.10 | 26 | — | — | |
| Mean Acc.Backbone Model=DS-R1-Distill-Llama-8B2025.10 | 23 | — | — | |
| High Conf.Backbone Model=DS-R1-Distill-Llama-8B2025.10 | 23 | — | — | |
| Loc. non-uniBackbone Model=DS-R1-Distill-Llama-8B2025.10 | 23 | — | — | |
| Loc. uniBackbone Model=DS-R1-Distill-Llama-8B2025.10 | 23 | — | — | |
| Glob. uniBackbone Model=DS-R1-Distill-Llama-8B2025.10 | 23 | — | — | |
| Self-Cert.Backbone Model=DS-R1-Distill-Llama-8B2025.10 | 22 | — | — | |
| Low Ent.Backbone Model=DS-R1-Distill-Llama-8B2025.10 | 22 | — | — | |
| Glob. non-uniBackbone Model=DS-R1-Distill-Llama-8B2025.10 | 21 | — | — |