MatFlowBench Experimental
Science · 2026-10-09
The 20 experimental MatFlowBench tasks, in which agents choose samples and instrument measurements within a finite budget from a simulated lab that returns real archived measurements, covering optical-film and catalyst selection, feasibility triage and raw instrument-data reduction. The verifier checks the requested measurements behind each recommendation; the score is the share of passing rollouts, four per task.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 62.5 |
| Claude Opus 5.5 | 47.5 |
| GPT-6.1 Sol | 43.8 |
| Claude Sonnet 5.5 | 42.5 |
| Gemini 3.8 Flash | 40.0 |
| Grok 4.7 | 33.8 |
| Hy4 Preview | 27.5 |
| Kimi K3 | 27.5 |
| Muse Spark 1.3 | 27.5 |