RSI-PDEbench: A Benchmark for Self-Improving Agents in PDE Solving

Binghang Lu, Lang Cao, Changhong Mou, Yige Sun, Cindy Xiangrui Kong, Zheyuan Deng, Tian Yuan, Bing Hu, Runyu Zhang, Na Lu, Guang LinarXiv 2026 · arXiv

Scientific discovery is often a messy "identify-then-solve" loop. You start with sparse, noisy data from an experiment and try to work backward to find the underlying Partial Differential Equation (PDE). Once you think you have the equation, you still have to estimate its physical coefficients and build a stable numerical solver to make predictions. This workflow typically requires a PhD-level mix of physics intuition and computational expertise.

A new paper, "RSI-PDEbench: A Benchmark for Self-Improving Agents in PDE Solving," asks a bold question: Can Large Language Model (LLM) agents automate this entire scientific pipeline? To find out, the authors developed a massive new benchmark and a novel "self-improving" agent framework that significantly outperforms standard AI coding assistants.

TL;DR

The authors introduce RSI-PDEbench, a benchmark of 1,001 cases designed to test if LLM agents can identify and solve PDEs from noisy data. They also debut RSI-PDE, a self-improving agent harness that uses physics-based "residual guidance" to boost accuracy by up to 181% while nearly halving the number of expensive model calls.

What Problem Does This Paper Address?

While LLMs are increasingly capable at writing general code, scientific computing is a different beast. Identifying a physical law from data is an "inverse problem"—it's mathematically "ill-posed," meaning many different equations might look like they fit the data, especially when that data is sparse or noisy.

Existing benchmarks often simplify this by giving the agent the equation upfront, focusing only on the coding of the solver. This paper tackles the much harder, more realistic task: The agent is given only the data and a library of possible physical laws. It must correctly identify the physics, tune the parameters, and deliver a working, accurate solver.

Methods

The core innovation is the RSI-PDE harness, a structured environment that wraps around a frontier LLM (like Claude or GPT). Instead of a single-shot attempt, the harness enables a recursively self-improving loop:

  1. Residual Guidance: The agent doesn't just guess. It uses a "residual screen" to calculate how well various equation templates fit the observed data points. This narrows the search to a shortlist of the five most likely candidates.
  2. Identify-then-Solve Workflow: For each candidate, the agent estimates coefficients and implements a numerical solver in a sandbox.
  3. Validation Gate: This is the "self-improvement" engine. The agent only keeps updates to its solver code, instruction policy, or memory if they pass a strict validation gate—meaning they must measurably improve performance on known training cases without breaking previous successes.
  4. Evolving Memory: The agent maintains a "lessons learned" memory and a self-written "instruction policy" that it refines over multiple rounds of experimentation.
The RSI-PDE workflow: (1) Unknown PDE data enters, (2) the LLM writes a program, (3) candidates are screened by residuals, and (4) a validation gate decides what to keep.

Datasets

The authors introduced RSI-PDEbench, which they claim is the largest joint PDE/Fractional PDE (FPDE) benchmark for coding agents.

  • Scope: 1,001 cases across 42 templates, including classical physics (Heat, Wave, Navier-Stokes) and complex fractional dynamics (which model materials with "memory").
  • Evaluation Settings:
    • Full: Clean, noise-free data to test baseline logic.
  • Noisy: Adds a massive 50% RMS Gaussian noise—a "brutal" level of corruption that tests if the agent can see the physics through the static.
  • Hard-OOD: Out-of-distribution tests with unseen equation templates and shifted coefficient ranges.
Visual examples of the diverse physical fields included in the benchmark, from fluid flow to electromagnetics.

Key Insights and Ideas

  • Physics Priors are Essential: The "aha moment" here is that LLMs shouldn't just be treated as coders; they need to be grounded in mathematical consistency. By using equation residuals (how much the data "violates" a candidate equation) as a primary signal, the agent avoids wasting time on physically impossible solutions.
  • Validation Prevents Hallucination: Recursive self-improvement can be dangerous; an agent might "improve" its way into a local minimum of nonsense. The Validation Gate acts as a scientific peer-review process, ensuring only genuine progress is retained.
  • Structured Loops Save Money: Surprisingly, the more complex, iterative approach was actually more efficient. By failing fast in a sandbox and only validating promising leads, RSI-PDE reduced total model calls by 46.9% compared to standard agentic workflows.

Key Results

The researchers tested the harness across four "frontier" models (using future-dated versions like Opus 5.5 and GPT-6.1).

  • Massive Accuracy Gains: RSI-PDE improved pass rates by 106.5% to 181.9% over single-run baselines.
  • Outperforming Native Agents: It beat "native" agent harnesses (like Claude Code) by 3.0% to 19.6% across the different settings.
  • High-Noise Resilience: Even with 50% noise, the system maintained a strong lead, proving that the residual-based identification is a robust way to handle "messy" real-world data.
Final accuracy and learning curves. The red lines (RSI-PDE) consistently reach higher pass counts faster than the black dashed lines (native harnesses).

Limitations and Future Work

  • Synthetic Data: While the benchmark is diverse, it is still synthetic. Real laboratory data often contains systematic biases or non-Gaussian noise that might confuse the current residual-ranking system.
  • Closed Hypothesis Space: The agent currently selects from a library of 42 templates. It cannot yet "invent" entirely new mathematical forms (symbolic discovery) from scratch.
  • Identifiability: At very high noise levels, different physical laws can become mathematically indistinguishable. The authors note that a low residual doesn't guarantee the agent found the "true" law, only a consistent one.

Future work will likely focus on expanding the agent's ability to perform open-ended equation discovery and testing these agents on real-world experimental datasets.

Summary updated

Abstract

Identifying and solving partial differential equations (PDEs) from sampled solution data often requires substantial expert effort in equation discovery, coefficient estimation, and numerical solver configuration, especially under noise and distribution shifts. Can large language model (LLM) agents automate this workflow accurately and efficiently? To investigate this question, we introduce \bench{}, a benchmark of 1,001 cases spanning 42 equation templates, including both classical and fractional PDEs. It provides three evaluation settings: Full evaluates performance on noise-free solution data; Noisy tests robustness to Gaussian noise with an RMS equal to 50\% of the clean-signal RMS; and Hard-OOD tests challenging out-of-distribution generalization to unseen equation templates and shifted coefficient ranges. We further introduce RSI-PDE, a recursively self-improving agent harness that shortlists candidate equations by residual loss, uses relative $L^2$ feedback to refine coefficients and numerical solvers, and applies a validation gate to policy and memory updates. We compare RSI-PDE with single-run baselines and native agent harnesses (Claude Code or Codex) across four frontier models. Across the three settings, RSI-PDE improves the pass rate averaged across the four models by 106.5--181.9\% relative to single-run baselines and by 3.0--19.6\% relative to native harnesses, while reducing total recorded model calls by 46.9\% across the 12 model--setting pairs. These results highlight the potential of validated recursive self-improvement in LLM agents to advance scientific computing.
BibTeX
@misc{lu2026rsi,
  author = {Binghang Lu and Lang Cao and Changhong Mou and Yige Sun and Cindy Xiangrui Kong and Zheyuan Deng and Tian Yuan and Bing Hu and Runyu Zhang and Na Lu and Guang Lin},
  title = {{RSI-PDEbench: A Benchmark for Self-Improving Agents in PDE Solving}},
  year = {2026},
  eprint = {2611.14949},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.14949},
}
Figure