RSI-PDE preprint reports higher equation-solving pass rates with fewer model calls

Researchers report that RSI-PDE, a language-model agent system for discovering and solving partial differential equations from sampled data, improved average pass rates while reducing total recorded model calls by 46.9%. Their preprint evaluates four frontier models on a new benchmark containing 1,001 cases across 42 equation templates. The system ranks candidate equations by residual error, refines coefficients and numerical solvers using feedback, and validates policy and memory updates. Average pass rates improved by 106.5-181.9% relative to single-run baselines and 3.0-19.6% relative to native agent systems. Tests covered clean data, noisy data and unfamiliar equations; the findings remain benchmark results.