Atlas

Benchmarks

SciResearch-Eval (interval score, physics-wrong tasks)

Lawrence Livermore National Laboratory's SciResearch-Eval tests scientific judgment on one laboratory problem: an agent measures a plate's thermal conductivity, heat capacity and rear heat-loss coefficient by designing up to four heating experiments and fitting a supplied heat-conduction simulator, within 60 tool calls, and reports each property with a 95% interval. Hidden from the agent, the simulator may be numerically under-resolved or missing real physics (a glued joint, radiation, heat lost through the sides, or temperature-dependent conductivity), and the truth is exact because the real sample is itself simulated with richer physics. The headline score, which ranks models, is the proper interval score of Gneiting and Raftery on a log10 scale (the interval's width plus 40 times any miss), averaged over the three properties and taken as a geometric mean over the 38 held-out tasks whose simulator is missing physics. Lower is better: about 0.01 is excellent and above 1 is confidently wrong. No language-model judge scores it.

Models
9
Top score
0.0
Leader
GPT-6 Astra
Updated
Nov 16, 2026
Resolution
—
Saturation
—

No signal or difficulty estimate. This benchmark has no declared score range, so it cannot be placed on the 0-1 scale the capability curve requires.

Loading Atlas data…