Math on iGSM
100AccuracySFT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SFTBackbone=Qwen2.5-3B-Instruct2026.05 | 100 | — | |
| SFTBackbone=Llama-3.2-3B-Instruct2026.05 | 100 | 19.1 | |
| Anchored LearningBackbone=Llama-3.2-3B-Instruct2026.05 | 99.4 | 52 | |
| Anchored LearningBackbone=Qwen2.5-3B-Instruct2026.05 | 93.6 | — | |
| Anchored LearningBackbone=Qwen2.5-3B-Instruct, Incremental Learning Stage=Stage 1: iGSM2026.05 | 93.6 | — | |
| Low-SFTBackbone=Qwen2.5-3B-Instruct2026.05 | 93 | — | |
| Low-SFTBackbone=Qwen2.5-3B-Instruct, Incremental Learning Stage=Stage 1: iGSM2026.05 | 93 | — | |
| Anchored LearningBackbone=Qwen2.5-3B-Instruct, Incremental Learning Stage=Stage 2: iGSM -> MedCalc2026.05 | 88.4 | — | |
| Anchored LearningBackbone=Qwen2.5-3B-Instruct, Incremental Learning Stage=Stage 3: iGSM -> MedCalc -> IFEval2026.05 | 70 | — | |
| DFTBackbone=Llama-3.2-3B-Instruct2026.05 | 66 | 24.2 | |
| Low-SFTBackbone=Llama-3.2-3B-Instruct2026.05 | 64.8 | 45.6 | |
| KL-SFTBackbone=Llama-3.2-3B-Instruct2026.05 | 62.4 | 20.7 | |
| Low-SFTBackbone=Qwen2.5-3B-Instruct, Incremental Learning Stage=Stage 2: iGSM -> MedCalc2026.05 | 62.2 | — | |
| KL-SFTBackbone=Qwen2.5-3B-Instruct2026.05 | 60.6 | — | |
| STMBackbone=Qwen2.5-3B-Instruct2026.05 | 59.4 | — | |
| DFTBackbone=Qwen2.5-3B-Instruct2026.05 | 52.4 | — | |
| STMBackbone=Llama-3.2-3B-Instruct2026.05 | 49.2 | 41.8 | |
| Low-SFTBackbone=Qwen2.5-3B-Instruct, Incremental Learning Stage=Stage 3: iGSM -> MedCalc -> IFEval2026.05 | 33 | — | |
| Iter-SFTBackbone=Qwen2.5-3B-Instruct2026.05 | 29.4 | — | |
| Iter-SFTBackbone=Llama-3.2-3B-Instruct2026.05 | 27.2 | 39.6 | |
| Self-SFTBackbone=Qwen2.5-3B-Instruct2026.05 | 25 | — | |
| Self-SFTBackbone=Llama-3.2-3B-Instruct2026.05 | 24.2 | 37.8 | |
| BaseBackbone=Qwen2.5-3B-Instruct2026.05 | 18.4 | — | |
| BaseBackbone=Qwen2.5-3B-Instruct, Incremental Learning Stage=Pre-trained2026.05 | 18.4 | — | |
| BaseBackbone=Llama-3.2-3B-Instruct2026.05 | 16.6 | 39.4 |