Recursive Self-Improvement on a Budget: Scoring Code Edits Before They Run

Key points

• Each run fits one linear readout on the parent's steering-split solve rate and eight coordinates of the proposed diff, from 96 scaffold edits labeled by four evaluations of a 16-instance fit split, and then freezes the coefficients.1
• Huxley-Gödel Machine selection draws a parent by Thompson sampling on the steering counts of nodes that have already been evaluated, and it expands the child that parent proposes.1
• Every method then returns the earliest evaluated node whose steering-split solve count is maximal.1
• On that split Huxley-Gödel Machine selection returns 85.1%, the root is at 81.2%, and the calibration-best edit is at 82.8%.1

Sources (1)

• [1]Recursive Self-Improvement on a Budget: Scoring Code Edits Before They Run
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 03:04 AM

  A coding agent that rewrites its own scaffold has to choose which proposed edit to expand before that edit has been evaluated, and the search then has to return one agent.

  Each run fits one linear readout on the parent's steering-split solve rate and eight coordinates of the proposed diff, from 96 scaffold edits labeled by four evaluations of a 16-instance fit split, and then freezes the coefficients.

Extractive summary: sentences quoted from the sources.