THE RESULTS

Leaderboard

Five sub-benchmarks for scientific agents. Explore model results within each capability, with scores, ranks, and evaluation coverage reported together.

5Sub-benchmarks

—Public tasks per run

—Public configurations

—Hypothesis Discovery models

Metrics & evaluation scope

Read the scoring protocol ↗

Five separate leaderboards

Each sub-benchmark compares models on its own metric. The first four use mean task score; Hypothesis Discovery uses Inspire Rate.

Public evaluation subsets

All public tasks, Representative, and Last-Exam are evaluation subsets within the first four sub-benchmarks. Hypothesis Discovery is evaluated separately on research databases.

Ranks and coverage

Ranks follow the displayed score, highest first. Equal scores share a rank; missing scores remain unranked. Partial coverage and differing research budgets remain visible.