Benchmark RadarMatPhaseBench
MatPhaseBench evaluates vision-language models on understanding materials phase diagrams, using 200 diagram-text pairs from 3681 papers. It targets complex scientific image understanding, with tasks requiring deep comprehension and open-ended responses, covering 189 material systems and 70 elements.
- Released
- 2026-07-03
- Readiness
- Paper only
- Primary field
- Science & Research
Why it matters
This benchmark addresses the gap in evaluating VLMs on logically complex scientific diagrams that require mechanistic reasoning. It measures capabilities beyond surface perception, helping assess practical value for AI-assisted materials science analysis.
Motivation
Materials phase diagrams are a core knowledge representation in materials science, encoding temperature,composition, phase stability, and phase transformation pathways, with their full understanding requiring thermodynamic mechanism analysis and scientific reasoning.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.