Science Question Answering on GPQA (Accuracy, Time, Peak)
47.98AccuracyHybridThinker
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| HybridThinkerBackbone=Qwen2.5-7B Series2026.06 | 47.98 | 14.84 | 2,124 | |
| HybridThinkerBackbone=Llama3.1-8B Series2026.06 | 37.88 | 14.55 | 1,971 | |
| VanillaBackbone=Qwen2.5-7B Series2026.06 | 37.37 | 24.5 | 7,184 | |
| LightThinkerBackbone=Qwen2.5-7B Series2026.06 | 37.37 | 13.33 | 1,969 | |
| VanillaBackbone=Llama3.1-8B Series2026.06 | 37.37 | 24.72 | 6,824 | |
| LightThinkerBackbone=Llama3.1-8B Series2026.06 | 37.37 | 15.59 | 1,933 | |
| Distill-R1Backbone=Llama3.1-8B Series2026.06 | 30.81 | 18.16 | 5,640 | |
| CoTBackbone=Qwen2.5-7B Series2026.06 | 28.28 | 1.18 | 998 | |
| CoTBackbone=Llama3.1-8B Series2026.06 | 24.75 | 4.58 | 2,230 | |
| Vanilla + H2OBackbone=Llama3.1-8B Series2026.06 | 23.23 | 53.47 | 1,024 | |
| Distill-R1Backbone=Qwen2.5-7B Series2026.06 | 22.22 | 11.89 | 4,383 | |
| Vanilla + H2OBackbone=Qwen2.5-7B Series2026.06 | 20.2 | 47.98 | 1,024 | |
| Vanilla + SepLLMBackbone=Llama3.1-8B Series2026.06 | 15.66 | 23.69 | 1,024 | |
| Vanilla + SepLLMBackbone=Qwen2.5-7B Series2026.06 | 12.63 | 40.89 | 1,024 |