Cambridge's 'Red Queen Gödel Machine' Co-Evolves the Evaluator With the Agent, Reporting 1.78–1.86x Higher Paper Acceptance

University of Cambridge / Hacker News·low signalXBlueskyLinkedInCopy link

Cambridge's Machine Learning Systems Lab (Alex Iacob under Prof. Nic Lane, with NVIDIA, Flower Labs, MBZUAI, and Inria) is attacking the self-improvement plateau by making the grader improve too: 'The test does not merely measure progress, it defines it, so the efficacy of the test becomes a ceiling the agent cannot climb past.' The method alternates fixed-evaluator phases with checkpoints that swap in stronger evaluators, reporting 1.78–1.86x higher acceptance rates on scientific paper writing, 9% better grading accuracy, and ~13x lower search token cost via a hybrid open/proprietary model split. Freshness caveat: the arXiv paper is dated July 21, 2026, so this is the Cambridge press write-up reaching HN (78 points) rather than a new result.