AI & Agent Evaluation
2,151total visitsadmin

Huxley-Gödel Machine: Human-Level Coding Agent Development

peer-reviewed paper · source date 2025-10-24 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog

Problems / challenges / motivations

  • Self-modification search scores a variant by its own benchmark performance, but the useful property is whether it leads somewhere — the two come apart.
  • The paper names this the Metaproductivity–Performance Mismatch: a mediocre variant can be the ancestor of the best ones.
  • Expanding the wrong branches burns large amounts of compute.

Key ideas

  • Proposes Clade-Metaproductivity (CMP): score a variant by the aggregate benchmark performance of its descendants rather than its own.
  • Guides tree expansion with Thompson sampling, balancing exploring new branches against deepening promising ones.
  • Beats DGM and SICA on SWE-bench Verified and Polyglot using fewer CPU hours, reaching human-level on SWE-bench Lite.
  • KAUST / MBZUAI (Schmidhuber group); ICLR 2026 oral.

Why it matters for AI engineering

  • It is the most rigorous refinement of the Gödel-machine lineage's search signal, and later work (Red Queen GM) builds directly on CMP.
  • The general principle applies well beyond agent evolution: when selecting among scaffolds or prompts, the metric should reflect downstream potential, not immediate score.
  • Caveat: out-of-window (October 2025), and results are coding-benchmark-specific — "human-level" is relative to SWE-bench Lite. Heavily promoted by the authoring group, with limited independent replication of CMP's advantage.

Comments

No comments yet.