Google's RRSI Paper Adds Regularization to Stop Agent Self-Improvement Overfitting
cihangxie · x · 2026-09-23
Google researchers released RRSI (arXiv:2609.24972, with code and project page), showing that recursive self-improvement of agent harnesses overfits training tasks — in-distribution gains shrink on out-of-distribution benchmarks. Their fix constrains both the proposer (annealed edit budget, encouraging unexplored trajectories) and the selector (a critic to screen benchmark-specific proposals, a pruner to drop small/costly/useless edits), validated across 8 benchmarks spanning coding, agentic workspace and engineering.
More from coding & agent
- AutoRAGBench: LLM agents build RAG systems beating the best published human designs — omarsar0 · 2026-11-22
- Stop @-ing your agent in Slack — let Jev decide when it should respond — DavidWells · 2026-11-22
- Intent adds notch launcher, Claude Code and Codex session import — Wattenberger · 2026-11-22
- rakyll explains AX: a low-opinion job orchestrator that agents control themselves — rakyll · 2026-11-22
- Gemini Neon builds an interactive USL lock-contention UI exposing a testing blind spot — tianyin_xu · 2026-11-22
- Claude Code can now be approved from your iPhone, ending the 14-minute permission wait — BLCNYY · 2026-11-22