Surface Evolver Bench, created by Yann Henon, evaluates whether LLM agents can write complete datafiles for Surface Evolver, a long-standing physics program for modeling liquid surfaces shaped by surface tension. Each of its 16 tasks asks the agent to model a liquid surface under geometric constraints, contact angles, gravity, and volume conditions, with tools to consult documentation, run the simulator, and submit a final file.

We source results from the public Surface Evolver Bench leaderboard. Our chart reports the mean score, a partial-credit average over the benchmark’s tasks; the stricter binary pass rate is kept alongside it in the data export.

Each task is graded deterministically with partial credit: passing static checks on the datafile, completing a hidden simulation run, and matching hidden physical quantities (volume, area, energy) within tolerances against a reference solution. A task counts as passed only when every check succeeds, which is why the pass rate is lower than the mean score. Models are run once per task in an agentic loop, and reasoning-effort variants of the same model appear as separate entries.