Math Reasoning on SVAMP (Accuracy, Reasoning Length)
78.67AccuracyHypLoRA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| HypLoRABackbone=Qwen2.5-3B2026.02 | 78.67 | — | |
| MoSLoRABackbone=Qwen2.5-3B2026.02 | 78 | — | |
| LoRABackbone=Qwen2.5-3B2026.02 | 76.33 | — | |
| HydraLoRABackbone=Qwen2.5-3B2026.02 | 75.33 | — | |
| BaseBackbone=Qwen2.5-3B2026.02 | 74.67 | — | |
| DLR (Full)Backbone=Qwen3-VL-2B-Instruct2026.06 | 67.7 | 3.8 | |
| CoT-SFTBackbone=Qwen3-VL-2B-Instruct2026.06 | 67.3 | 63.4 | |
| CoT-SFTBackbone=LLaMA-3.2-1B2026.06 | 66.7 | 61.8 | |
| DLR (Full)Backbone=LLaMA-3.2-1B2026.06 | 57.3 | 3.5 | |
| RoTBackbone=Qwen3-VL-2B-Instruct2026.06 | 53.7 | 32 | |
| Direct-SFTBackbone=Qwen3-VL-2B-Instruct2026.06 | 52.3 | 0 | |
| ReGuLaRBackbone=LLaMA-3.2-1B-Instruct2026.01 | 50.1 | 2.02 | |
| ReGuLaRBackbone=LLaMA-3.2-1B2026.06 | 50.1 | 2 | |
| CoLaRBackbone=Qwen3-VL-2B-Instruct2026.06 | 49.3 | 3.2 | |
| CoLaRBackbone=LLaMA-3.2-1B-Instruct, Compression Rate=52026.01 | 47.1 | 2.96 | |
| CoLaRBackbone=LLaMA-3.2-1B2026.06 | 47.1 | 3 | |
| CoconutBackbone=Qwen3-VL-2B-Instruct2026.06 | 41.5 | 6 | |
| Direct-SFTBackbone=LLaMA-3.2-1B2026.06 | 41.3 | 0 | |
| CoconutBackbone=LLaMA-3.2-1B-Instruct2026.01 | 39.8 | 6 | |
| CoconutBackbone=LLaMA-3.2-1B2026.06 | 39.8 | 6 | |
| iCoTBackbone=LLaMA-3.2-1B-Instruct2026.01 | 36.4 | 0 | |
| iCoTBackbone=LLaMA-3.2-1B2026.06 | 36.4 | 0 | |
| CODIBackbone=LLaMA-3.2-1B-Instruct2026.01 | 21.7 | 6 | |
| CODIBackbone=LLaMA-3.2-1B2026.06 | 21.7 | 6 | |
| CODIBackbone=Qwen3-VL-2B-Instruct2026.06 | 11.3 | 6 |