In-Context Symbolic Regression for Robustness-Improved Kolmogorov-Arnold Networks

Symbolic regression can make AI more interpretable: The paper addresses a key challenge in responsible AI by developing methods to convert complex neural networks (KANs) into simple, human-readable mathematical formulas that scientists and regulators can inspect and validate, rather than relying on unexplainable "black-box" predictions.

Current symbolic extraction methods are unreliable: The standard approach of converting each network connection independently is "sensitive to initialisation and non-convex parameter fitting," meaning small changes in setup can produce different formulas—a significant concern for deploying AI in high-stakes scientific and medical applications where consistency matters.

New methods improve robustness through context-aware selection: The paper proposes two approaches (GSR and GMP) that consider how individual formula choices interact across the entire network rather than in isolation, and tests robustness through systematic parameter sweeps to ensure recovered formulas remain consistent and accurate under different conditions.

Robustness testing is essential for trustworthy AI: The authors emphasize measuring "qualitative consistency of recovered formulas" alongside predictive accuracy, recognizing that responsible AI deployment requires not just correct answers, but stable, reproducible explanations that stakeholders can rely on.