Back to Devoured - September 23, 2026
DEVOURED
Google Publishes RRSI for Self-Improving AI Agents

Google Publishes RRSI for Self-Improving AI Agents

AI research Regularized-rsi.com
Google's new RRSI method forces AI agents to learn without overfitting, successfully improving performance across eight different benchmarks.
What: Researchers from Google Cloud AI and several universities developed Regularized Recursive Self-Improvement (RRSI). By constraining the feedback loops during agent training, the method prevents performance degradation on new, out-of-distribution tasks, outperforming prior methods by up to 22.9%.
Why it matters: This approach addresses the tendency of self-improving agents to exploit specific benchmark patterns rather than learning robust, transferable skills.
Deep dive
  • Regularization of the search loop: Instead of just editing agent code, RRSI regulates the evolution process by managing the edit budget and evaluating candidate changes against a 'noise floor'.
  • Leakage critic: An automated check that flags and rejects benchmark-specific logic or data before it is scored.
  • Cost-aware evolution: The algorithm requires that any performance gain must pay for the additional inference tokens required by the change; otherwise, the component is pruned.
  • Performance stability: Unlike prior self-improvement methods that often show regression on held-out tasks, RRSI maintained consistent gains across all tested domains.
Decoder
  • Recursive self-improvement: An iterative process where an AI agent modifies its own code or parameters to achieve higher performance.
  • Out-of-distribution (OOD): Data or tasks that differ significantly from the training set, testing the model's ability to generalize.
  • Policy tokens: The compute resources or 'budget' consumed by the agent's logic during execution.
Original article

Evolved harnesses overfit the benchmark they are scored on. RRSI transfers.

RRSI improves every out-of-distribution benchmark without overfitting the split it evolves on. Prior methods do the opposite: large evolve-set gains that shrink or vanish once the benchmark changes, two of them ending below the harness they started from.

Regularize the search, not the harness

Every harness component stays editable. RRSI constrains the loop that edits it: how much one proposal may change, and which measured gains are allowed to stick.

Proposal side

Annealed edit budget: Early rounds may bundle a few coordinated edits to find a mechanism; late rounds get one attributable change.

Evidence-aware credit: Every candidate is logged with its hypothesis, diff, score and cost change, so the proposer builds on what worked and stops re-testing what failed.

Structured exploration: When progress stalls inside the noise band, budget is redirected to components the run has never touched.

Selection side

Leakage critic: Task names, entities, answers or benchmark-specific logic are rejected before a candidate is ever scored.

Noise-adjusted floor: A gain must clear the variance measured on the unchanged base harness.

Cost rule: Extra inference tokens have to be paid for by measured gain.

Pruning: Components that stop earning their place are flagged for deletion.

Every held-out benchmark improves

Evolve on one suite per domain, then run the harness unchanged everywhere else. Same tools, judge, trials and window as H0; the policy is Claude Opus 4.8.

Two rules act on cost directly

The cost rule refuses growth that is not paid for when it is proposed; pruning removes growth that stopped paying for itself since. No prior method carries either.

Watch the harness evolve, round by round

Four real runs, every candidate: what it proposed, what the critic said, why the gate kept or dropped it, and the exact diff.

BibTeX

@article{xia2026rrsi,
  title={RRSI: Regularized Recursive Self-Improvement of Agent Harnesses},
  author={Xia, Peng and Han, Rujun and Wang, Zifeng and Chen, Yanfei and Zhuang, Yufan and Lee, Yoonho and Huang, Chengsong and Yu, Han and CuiZhu, Zhongying and Ming, Yifei and Yao, Huaxiu and Gokturk, Burak and Pfister, Tomas and Lee, Chen-Yu},
  journal={arXiv preprint arXiv:2609.24972},
  year={2026}
}
Explore all 56 articles in today's digest

Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi