Expert-level Science Question Answering on GPQA (Accuracy)
36.4AccuracyUniSD*
Evaluation results
| Method | Links | |
|---|---|---|
| UniSD*Base Model=Qwen2.5-7B2026.05 | 36.4 | |
| Agree (Tok.)Base Model=Qwen2.5-7B, Protocol=Token-level agreement, Context Retrieval=True2026.05 | 36.2 | |
| Match (Repr.)Base Model=Qwen2.5-7B, Protocol=Representation-only matching2026.05 | 35.9 | |
| Agree (Seq.)Base Model=Qwen2.5-7B, Protocol=Sequence-level agreement, Context Retrieval=True2026.05 | 35.7 | |
| EMABase Model=Qwen2.5-7B, Protocol=EMA teacher2026.05 | 35.3 | |
| ContrastBase Model=Qwen2.5-7B, Protocol=Token-level contrastive learning2026.05 | 34.6 | |
| SDFTBase Model=Qwen2.5-7B2026.05 | 34.2 | |
| SSDBase Model=Qwen2.5-7B2026.05 | 33.7 | |
| Match (Joint)Base Model=Qwen2.5-7B, Protocol=Joint logit-representation matching2026.05 | 33.5 | |
| GKDBase Model=Qwen2.5-7B2026.05 | 33 | |
| OPSDBase Model=Qwen2.5-7B2026.05 | 31.5 | |
| ClipBase Model=Qwen2.5-7B, Protocol=Divergence clipping2026.05 | 31.5 | |
| RawBase Model=Qwen2.5-7B2026.05 | 31 | |
| SFTBase Model=Qwen2.5-7B2026.05 | 30.6 |