Google's RRSI adds regularization to recursive self-improvement, +3.3 points OOD
jiqizhixin · x · 2026-10-13
Google and collaborators introduce RRSI (Regularized Recursive Self-Improvement), addressing a core flaw in RSI: agents can memorize benchmark patterns, mistake noise for gains, or pile on prompts, context, and control logic to buy tiny evolve-set improvements with ever more test-time compute.
RRSI regularizes the self-improvement loop so gains transfer out of distribution:
- Proposal-side: annealed update sparsity (budget shrinks over rounds), evidence-aware credit assignment over the full history, structured exploration toward underexplored components
- Selection-side: leakage screening, noise-adjusted performance floors, complexity-aware acceptance (Ridge/L2-style), structural pruning (Lasso/L1-style)
Results: on three fully held-out OOD benchmarks, unregularized evolution averages 40.3 vs 43.6 for RRSI (+3.3, peak +4.7), with every split improving; a harness evolved on Terminal-Bench 2.1 rises from 82.0 to 83.8.
More from coding & agent
- Summer Launches Game Audio Maker for AI-Generated AAA-Quality Game Sound — garrytan · 2026-10-13
- Seroter's reading list: stop installing skills, agent governance beats tool calls — rseroter · 2026-10-13
- Engineer Shares a Blog on Scaling AI Inference Stacks for Many Customers — abhijithneil · 2026-10-13
- Tim Sweeney: AI Boosts Dev Efficiency Up to 5x Today, Nearly Everything to Hit 3x Within a Year — zephyr_z9 · 2026-10-13
- Decision models: the new Hub category that outputs numbers, not paragraphs — erhannah · 2026-10-13
- Swarms API Launches Dynamic Pricing With 15% Service Fee Per Model — KyeGomezB · 2026-10-13