Google's CRSI paper: self-improvement memory lifts cross-model continuation by 2.36 points
kimmonismus · x · 2026-11-18
The arXiv preprint (2611.18856) behind the earlier discussion, from Google Cloud AI and DeepMind. It formalizes Continual Recursive Self-Improvement (CRSI)—continuing harness optimization across compatible benchmarks or policy models—and proposes Self-Improvement Memory (SIM) for the harness optimizer: edits grouped by recurring harness problem with recorded rationale, predicted/observed outcomes, and evaluation context. Across four domains, SIM beats harness-only and root-cause-only continuation in cross-dataset and pairwise cross-model comparisons; across ten independent coding runs, paired continuations improve final scores by an average of 2.36 points (p=0.003).
More from Research
- EACL 2027 results: 1,118 papers accepted from 4,136 submissions, main track rate just 13.5% — IAugenstein · 2026-11-18
- PhilBench quizzes 97 LLMs with the 2020 PhilPapers Survey of 1,785 philosophers — pwlot · 2026-11-18
- CADMAGEDDON: open-source agent rebuilds editable CAD locally on a single GPU — Oleg Pavlov · 2026-11-18
- MillBill: local 27B model beats frontier models at generating machining programs — Gennadii Savrasov · 2026-11-18
- The 200-year prehistory of speech synthesis: von Kempelen's 1791 speaking machine to 1970 — ArtificialOther · 2026-11-18
- Ex-DeepMind's Ryutaro Tanno joins AI biotech Pathos as Chief AI Officer — davidstutz92 · 2026-11-18