DeepResearch Bench
Search · 2025-06-13
DeepResearch Bench (USTC) asks deep-research agents to write analyst-grade reports for 100 PhD-level research tasks (50 Chinese, 50 English) across 22 fields. The headline RACE score is an LLM-judged, reference-relative report-quality score on a 0-100 scale (50 matches the reference report), weighted per task across comprehensiveness, insight, instruction following and readability. The judge model belongs to the evaluation configuration: these scores use the legacy Gemini-2.5-Pro judge, not the GPT-5.5 judge adopted in May 2026.
Top models (higher is better)
| Model | Score |
|---|---|
| Tongyi DeepResearch 30B A3B | 40.5 |
| Sonar Reasoning Pro | 37.8 |
| Sonar Reasoning | 37.8 |
| Claude 3.7 Sonnet | 36.6 |
| Sonar Pro | 36.2 |
| Gemini 2.5 Pro Preview 05-06 | 31.9 |
| GPT-4o Search Preview | 30.7 |