Benchmark RadarMat-Pref
Mat-Pref evaluates compositional reasoning in inorganic materials via 10,837 ionic-substitution questions across 11 structure families, with splits for in-distribution performance, held-out families, and cross-property transfer.
- Released
- 2026-06-20
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
It isolates generalization types (structural transfer, property transfer, memorization) in scientific reasoning, helping to identify when RLVR improves reasoning over memorization.
Motivation
Reinforcement learning from verifiable rewards (RLVR) has driven rapid progress in mathematical and code reasoning, but when extended to science, existing benchmarks do not decompose what generalizes: do gains reflect structural transfer, property transfer, or memorization?
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.