Training Numerical Intelligence via Auto-Diagnosis and Skill Discovery

Overview

AI agents are becoming increasingly capable of generating scientific code — but generating code is not the same as improving the algorithms behind it. For numerical solvers, execution feedback can expose poor performance, yet it rarely reveals the underlying cause or how to address it.

To close this gap, researchers Peter Chen and Wotao Yin introduce Auto-Diagnosis and Skill Discovery (ADSD), a framework that links numerical diagnosis to reusable solver self-improvement. The work was submitted on 2 October 2026 and is published as arXiv:2610.03872 [cs.AI], spanning 20 pages.

The Diagnosis-First Paradigm

ADSD follows a diagnosis-first paradigm that:

Explains why a solver performs poorly, rather than merely detecting that it does.
Uses this diagnosis to guide the discovery of appropriate numerical methods.
Packages the resulting knowledge into reusable solver skills.

This turns solver improvement from trial-and-error editing into a structured process of diagnosis, discovery, and implementation — a meaningful shift as agentic coding tools mature in 2026.

Results Across Four Numerical Domains

ADSD was evaluated on four challenging numerical domains:

Power flow equations
AC optimal power flow control
Stiff ordinary differential equations (ODEs)
Heterogeneous diffusion PDEs

Across all four, ADSD consistently improved solver accuracy, robustness, and efficiency. On the GOC-500 power flow benchmark, for example, ADSD reduced mean solver error by nearly 71×, with improvements further transferring to unseen grid topologies and operating regimes.

Paper Details

| Field | Value |

|---|---|

| Authors | Peter Chen, Wotao Yin |

| arXiv ID | arXiv:2610.03872 [cs.AI] |

| Submitted | 2 October 2026 |

| Length | 20 pages |

| Subjects | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG) |

| DOI | https://doi.org/10.48550/arXiv.2610.03872 (pending registration) |

Key Takeaway

ADSD demonstrates that separating diagnosis from repair allows AI agents to move beyond surface-level code generation toward genuine algorithmic self-improvement — a promising direction for autonomous scientific computing.

via ArXiv AI