Science Question Answering on GPQA (Acc, Peak, Dep)
38.22AccuracyTokenSkip
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TokenSkipModel Series=Qwen2.5-7B, Budget setting=true2026.04 | 38.22 | 5,611 | 21.2 | |
| LThinker++Model Series=Qwen2.5-7B, Budget setting=true2026.04 | 38.22 | 3,265 | 16.1 | |
| LThinker++Model Series=Llama3.1-8B, Budget setting=true2026.04 | 33.67 | 4,101 | 22.9 | |
| VanillaModel Series=Qwen2.5-7B, Budget setting=true2026.04 | 32.49 | 5,493 | 20 | |
| VanillaModel Series=Llama3.1-8B, Budget setting=true2026.04 | 31.82 | 5,773 | 21.7 | |
| LThinker*Model Series=Qwen2.5-7B, Budget setting=true2026.04 | 28.96 | 3,496 | 23.2 | |
| LThinker*Model Series=Llama3.1-8B, Budget setting=true2026.04 | 28.28 | 4,271 | 34.5 | |
| CoTModel Series=Qwen2.5-7B, Budget setting=true2026.04 | 26.76 | 968 | 0.5 | |
| TokenSkipModel Series=Llama3.1-8B, Budget setting=true2026.04 | 26.6 | 6,016 | 23.8 | |
| Distill-R1Model Series=Qwen2.5-7B, Budget setting=true2026.04 | 24.75 | 6,718 | 31 | |
| CoTModel Series=Llama3.1-8B, Budget setting=true2026.04 | 24.75 | 1,231 | 0.9 | |
| Distill-R1Model Series=Llama3.1-8B, Budget setting=true2026.04 | 20.2 | 3,972 | 16 |