PhysicsX cracks PLAID with 85 guesses

PhysicsX says its agent reached effectively zero error on the hidden Tensile2D track of Safran’s PLAID benchmark after testing 85 hypotheses about which samples the public leaderboard scored. In the company’s account, agents reconstructed the withheld constitutive law from 500 labeled simulations, then discovered the leaderboard evaluated only even-indexed samples; the final simulator ran in a median 0.437 seconds. For teams trusting public evaluation gates alongside simulation-heavy robot testing, the warning is immediate: hold out private tests and vary the scoring protocol, because an agent can optimize the benchmark’s fingerprint rather than the engineering task.