Science Reasoning on GPQA (Pass@1)
70.2Pass@1SwiR
Evaluation results
| Method | Links | |
|---|---|---|
| SwiRBackbone=Qwen3-32B2026.04 | 70.2 | |
| Chain of Mindset (CoM)Base Model=Qwen3-VL-32B-Instruct2026.02 | 69.7 | |
| CoTBackbone=Qwen3-32B, Decoding Strategy=Greedy2026.04 | 69.7 | |
| Soft ThinkingBackbone=Qwen3-32B2026.04 | 67.17 | |
| SeLaRBackbone=Qwen3-32B2026.04 | 67.17 | |
| CoTBackbone=Qwen3-32B, Decoding Strategy=Sampling2026.04 | 66.16 | |
| Chain of Mindset (CoM)Base Model=Gemini-2.0-Flash2026.02 | 65.7 | |
| Direct I/OBase Model=Qwen3-VL-32B-Instruct2026.02 | 63.64 | |
| Direct I/OBase Model=Gemini-2.0-Flash2026.02 | 62.63 | |
| SwiRBackbone=Qwen3-8B2026.04 | 62.63 | |
| w/o MergingModels=Math2026.02 | 61.65 | |
| SeLaRBackbone=Qwen3-8B2026.04 | 61.62 | |
| w/o MergingModels=Code2026.02 | 60.01 | |
| CoTBackbone=Qwen3-8B, Decoding Strategy=Sampling2026.04 | 58.08 | |
| Soft ThinkingBackbone=Qwen3-8B2026.04 | 57.58 | |
| SCF-RKLModels=Fuse2026.02 | 56.36 | |
| CoTBackbone=Qwen3-8B, Decoding Strategy=Greedy2026.04 | 55.05 | |
| FullKV budget=Full Cache2026.02 | 54.6 | |
| ForesightKVKV budget=4K2026.02 | 53.7 | |
| ForesightKV(w/o RL)KV budget=4K2026.02 | 52.4 | |
| Ties MergingModels=Fuse2026.02 | 52.08 | |
| SCEModels=Fuse2026.02 | 51.61 | |
| Task ArithmeticModels=Fuse2026.02 | 51.39 | |
| Dare Task ArithmeticModels=Fuse2026.02 | 51.39 | |
| ForesightKVKV budget=2K2026.02 | 51.3 | |
| ForesightKV(w/o RL)KV budget=2K2026.02 | 51.2 | |
| Dare Ties MergingModels=Fuse2026.02 | 51.07 | |
| R-KVKV budget=4K2026.02 | 50.7 | |
| SnapKVKV budget=4K2026.02 | 49.1 | |
| ForesightKVKV budget=1K2026.02 | 45.2 | |
| ForesightKV(w/o RL)KV budget=1K2026.02 | 44.2 | |
| H2OKV budget=4K2026.02 | 43.1 | |
| R-KVKV budget=2K2026.02 | 40 | |
| CoTBackbone=Qwen3-1.7B, Decoding Strategy=Sampling2026.04 | 39.39 | |
| SwiRBackbone=Qwen3-1.7B2026.04 | 37.88 | |
| SeLaRBackbone=Qwen3-1.7B2026.04 | 35.35 | |
| SnapKVKV budget=2K2026.02 | 34.8 | |
| Soft ThinkingBackbone=Qwen3-1.7B2026.04 | 33.83 | |
| CoTBackbone=Qwen3-1.7B, Decoding Strategy=Greedy2026.04 | 31.31 | |
| H2OKV budget=2K2026.02 | 29.7 | |
| H2OKV budget=1K2026.02 | 25.2 | |
| R-KVKV budget=1K2026.02 | 23 | |
| MATRIX-Gen-ReasonSFT Dataset=MATRIX-Gen-Reason2024.10 | 17.18 | |
| SnapKVKV budget=1K2026.02 | 16.9 | |
| Tulu 3 CodeSFT Dataset=Tulu 3 Code2024.10 | 13.13 | |
| Camel MATHSFT Dataset=Camel MATH2024.10 | 11.11 | |
| Magpie ReasonSFT Dataset=Magpie Reason2024.10 | 10.6 | |
| PersonaHub-MathSFT Dataset=PersonaHub-Math2024.10 | 10.1 | |
| Numina MATHSFT Dataset=Numina MATH2024.10 | 9.09 | |
| Tulu 3 MATHSFT Dataset=Tulu 3 MATH2024.10 | 6.57 |