Scientific Reasoning on GPQA (Accuracy, Gain, Speedup)
27AccuracyFull-step decoding
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Full-step decodingModel=Dream-7B-Instruct, Decoding Strategy=Full-step, Evaluation Setting=Zero-shot2025.08 | 27 | — | — | |
| ProphetModel=Dream-7B-Instruct, Decoding Strategy=Prophet, Evaluation Setting=Zero-shot2025.08 | 26.6 | -0.4 | 1.66 | |
| ProphetModel=LLaDA-8B-Instruct, Decoding Strategy=Prophet, Evaluation Setting=Zero-shot2025.08 | 25.7 | 0.5 | 1.82 | |
| Full-step decodingModel=LLaDA-8B-Instruct, Decoding Strategy=Full-step, Evaluation Setting=Zero-shot2025.08 | 25.2 | — | — |