ICLR 2026 · 2026 · Oral

Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine

Wenyi Wang^†, Piotr Piękos^†, Li Nanbo, Firas Laakom, Yimeng Chen, Mateusz Ostaszewski, Mingchen Zhuge, Jürgen Schmidhuber

Paper
Code

A strong coding agent is not necessarily a strong starting point for further self-improvement. Huxley-Gödel Machine (HGM) studies this mismatch between an agent’s current performance and its ability to produce better descendants, and uses it to guide the search over self-modifying agents.

HGM project illustration of a robot pulling itself upwards by its own hair. — The HGM project illustration: a visual metaphor for an agent improving its own implementation. View full size · Source

Learning which agents can improve

HGM grows a tree of coding-agent implementations through successive code modifications. Its central quantity, clade metaproductivity (CMP), aggregates benchmark performance across an agent’s descendants to estimate the future potential of that branch. Search can therefore favor a lineage that produces useful improvements even when its current agent is not the highest-scoring one.

Experiments on SWE-bench Verified and Polyglot show stronger coding-agent development than the compared self-improvement systems while using fewer allocated CPU hours. The paper also examines transfer across coding datasets and backbone models.