Hard Reasoning on GPQA (Score, Speedup)
55.4ScoreCoSpec
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CoSpectype=Collaborative Speculative Decoding2026.05 | 55.4 | 2.27 | |
| Target Modeltype=Standard Decoding2026.05 | 54.6 | — | |
| SpStype=Vanilla Speculative Sampling2026.05 | — | 1.31 |