MatFlowBench Pass@1
Science · 2026-10-09
Collinear AI's private 45-task benchmark of materials-research workflows for LLM agents: 25 computational tasks run DFT and related calculations with Quantum ESPRESSO, Wannier90 and Phonopy, and 20 experimental tasks spend a measurement budget on archived instrument data to select materials. A task-specific executable verifier grades each rollout from its calculation files or measurement log. Pass@1 averages each task's share of passing rollouts (two per computational task, four per experimental task) so that every task counts equally.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 60.0 |
| Claude Opus 5.5 | 56.7 |
| Claude Sonnet 5.5 | 50.0 |
| GPT-6.1 Sol | 45.0 |
| Grok 4.7 | 22.8 |
| Kimi K3 | 22.2 |
| Hy4 Preview | 20.0 |
| Gemini 3.8 Flash | 18.9 |
| Muse Spark 1.3 | 16.7 |