Scientific Question Answering on HLE-Verified Gold (test)
60AccuracyATLAS-MM
Evaluation results
| Method | Links | |
|---|---|---|
| ATLAS-MMBackbone=Claude Sonnet 4.62026.06 | 60 | |
| Self-Refine (no early stop)Backbone=Claude Sonnet 4.62026.06 | 58 | |
| ATLASBackbone=Claude Sonnet 4.62026.06 | 56 | |
| Self-RefineBackbone=Claude Sonnet 4.62026.06 | 53 | |
| GPT-5.2-HProtocol=Zero-shot2026.06 | 52.48 | |
| Reward-model rerankingBackbone=Claude Sonnet 4.6, Reward Model=Skywork-Reward-V22026.06 | 52 | |
| Budget ForcingBackbone=Claude Sonnet 4.62026.06 | 51 | |
| Opus-4.6Protocol=Zero-shot2026.06 | 50.16 | |
| Gemini-3-ProProtocol=Zero-shot2026.06 | 48.93 | |
| Qwen3-Max-TProtocol=Zero-shot2026.06 | 48.48 | |
| Opus-4.5Protocol=Zero-shot2026.06 | 48.16 | |
| Pass@1Backbone=Claude Sonnet 4.62026.06 | 48 |