A coding agent at PhysicsX rebuilds the simulator behind Safran's PLAID Tensile2D benchmark and tops its leaderboard with zero error

Douglas Boubert of PhysicsX describes how a coding agent reverse-engineered Tensile2D, a benchmark in Safran's PLAID collection of two-dimensional quasistatic structural simulations generated with the Z-set finite-element solver. The benchmark gives the mesh, the applied pressure and five material parameters and withholds the constitutive law. Working from the 500 labelled training simulations, the agent wrote its own simulator and refined the hardening law until its error on held-out samples was around 10⁻⁸ on all eight metrics, at a median of 0.437 seconds per simulation. PhysicsX reports that the simulator ranks first on the public leaderboard with zero error on every metric, and that the agent also inferred, after 85 hypotheses, that only the even-indexed test samples are scored publicly. The post concludes that public benchmarks which withhold simulation details are vulnerable to capable agents.

Share on LinkedIn, X or Bluesky, or copy the link.