RSI-PDEbench: A Benchmark for Self-Improving Agents in PDE Solving
Summary
- Researchers introduce RSI-PDEbench, a benchmark of 1,001 cases across 42 equation templates designed to evaluate LLM agents in solving partial differential equations under noise and distribution shifts.
- The benchmark covers three settings: Full (noise-free), Noisy (Gaussian noise), and Hard-OOD (out-of-distribution generalization to unseen templates and shifted coefficients).
- They also propose RSI-PDE, a recursively self-improving agent harness that uses residual loss, relative L2 feedback, and a validation gate for policy and memory updates.
- Across four frontier models, RSI-PDE improves pass rates by 106.5 to 181.9 percent over single-run baselines and 3.0 to 19.6 percent over native harnesses, while reducing total model calls by 46.9 percent.
Why it matters
The introduction of RSI-PDEbench and the RSI-PDE harness demonstrates how recursively self-improving LLM agents can dramatically enhance accuracy and efficiency in complex scientific computing tasks like partial differential equation solving.