Mathematical Reasoning on AIME 2025 (Avg@8, Avg@16, Avg@32)
25.1Avg@8 ScorePACED Reverse KL
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| PACED Reverse KLTrack=Self-distillation, Backbone=Qwen2.5-Math-7B-Instruct, KL Family=Reverse KL2026.03 | 25.1 | — | — | |
| Hard Filter Reverse KLTrack=Self-distillation, Backbone=Qwen2.5-Math-7B-Instruct, KL Family=Reverse KL2026.03 | 22 | — | — | |
| AKLTrack=Self-distillation, Backbone=Qwen2.5-Math-7B-Instruct, KL Family=Reverse KL2026.03 | 21.5 | — | — | |
| SENTBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.12 | 21.25 | 21.46 | 20.1 | |
| w/ MaskBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.12 | 20.83 | 17.92 | 19.38 | |
| w/ ClipBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.12 | 20.83 | 18.54 | 19.79 | |
| Base modelBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.12 | 19.17 | 19.17 | 17.71 | |
| w/ CovBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.12 | 19.17 | 17.08 | 19.48 | |
| w/ High EnBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.12 | 19.17 | 17.71 | 17.08 | |
| w/ AdvBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.12 | 18.75 | 17.92 | 18.33 | |
| Reverse KL (unweighted)Track=Self-distillation, Backbone=Qwen2.5-Math-7B-Instruct, KL Family=Reverse KL2026.03 | 16.9 | — | — | |
| w/ EnBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.12 | 16.67 | 17.08 | 18.12 | |
| GRPOBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.12 | 15.83 | 15.63 | 13.75 | |
| BaseTrack=Self-distillation, Backbone=Qwen2.5-Math-7B-Instruct, KL Family=Reverse KL2026.03 | 11.5 | — | — |