Atlas

Benchmarks

← All benchmarks

PG-LLM-50 Temporal (59 assays)

Science

PG-LLM's temporally held-out protein-variant ranking condition uses 59 assays from 19 studies whose measured fitness tables became available on or after February 1, 2026. Each text-only prompt contains 50 variants without tools or measured fitness; three fixed draws are scored by Spearman correlation, averaged within study-group and draw and then across study-groups. This differs from the 217-assay ProteinGym nested-macro condition. Published coverage varies by model and effort; reported standard errors are across draws.

Top models (higher is better)

ModelScore
GPT-5.6 Sol0.3
Opus 4.80.3
Gemini 3.5 Flash0.3
Loading Atlas data…