SCI-AGENT-EXAM
Science · 2026-10-23
120 expert-curated scientific research tasks across 11 fields and seven task types (code generation, error diagnosis, tool orchestration, end-to-end analysis, hypothesis validation, result interpretation, literature synthesis), each requiring executable and textual artifacts. The final score weights a deterministic programmatic check of the executable artifacts and a calibrated LLM-judge rubric of the reports by task type, averaged over three runs per task and macro-averaged across tasks, in percent. Agents run in a common Claude Code harness.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 80.9 |
| Claude Fable 5 | 67.3 |
| Claude Opus 5 | 64.2 |
| Qwen3.8-Max | 62.7 |
| Kimi K3 | 62.6 |
| GPT-5.6 Sol | 59.6 |
| Opus 4.8 | 57.1 |
| GLM-5.3 | 55.9 |
| Grok 4.6 | 52.4 |
| DeepSeek-V4-Pro | 50.9 |