KERNAUT uses coding agents and QD search to auto-discover interpretable kernel models

sirbayes · x · 2026-10-08

A new alphaXiv paper introduces Kernel Autoresearch (KERNAUT), which treats kernel design as open-ended model discovery: coding agents write kernels as programs, construction contracts guarantee validity of every accepted kernel, a quality-diversity archive retains high-performing kernels with distinct behaviors, and novelty screening steers agents toward functionally new candidates.

The authors frame the classic dilemma: fixed grammars of base kernels guarantee validity but limit expressiveness, while unrestricted programs are expressive but often invalid — in stress tests, 22–58% of LLM-generated kernels passing numerical checks fail at different scales or dimensions.
Interpretability is a key selling point: the best kernel discovered for glucose prediction has the form k(x,x') = phi(x)^T phi(x'), with phi a readable set of 16 scalar functions that humans can refine.
Results: on held-out black-box optimization families, a discovered kernel beats a meta-learned deep kernel trained on the same episodes; kernels learned from ten enzyme-kinetic rate laws achieve lower error than tuned ARD and deep kernel baselines on five unseen mechanisms; human refinement of one kernel further cuts held-out predictive error by 5.7% and optimization regret by 7.8%.
Co-authored by Kevin Murphy; project page and code are open-sourced (richardcsuwandi/kernaut).

Original post →

KERNAUT uses coding agents and QD search to auto-discover interpretable kernel models

More from coding & agent