RRSI: Regularized Recursive Self-Improvement of Agent Harnesses research paper by Google, 2026

Google · Sep 21, 2026 · Agents and evaluation · 0 citations · 180 upvotes

RRSI keeps self-improving agent harnesses from memorising their training tasks, so gains carry over to new benchmarks.

Summarised by hand from the abstract.

More from Google

Date	Paper	Topic	Citations
Jan 28, 2022	Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsAsking a model to write out its intermediate steps (chain of thought) sharply improves its math and logic answers.Reasoning · Jan 2022	Reasoning	22.1k
Oct 11, 2018	BERT: Pre-training of Deep Bidirectional Transformers for Language UnderstandingPre-training a Transformer to fill in masked words, then fine-tuning it, beat task-specific models across language understanding tests.Foundation models · Oct 2018	Foundation models	121k
Jun 12, 2017	Attention Is All You NeedIntroduced the Transformer, the attention-only architecture behind nearly every large language model since.Architectures · Jun 2017	Architectures	194k

Topic

Date	Paper	Lab	Citations
Jun 3, 2026	Agents' Last ExamOver 1,000 real, checkable professional tasks written with 250+ industry experts; the hardest tier is far from solved.UC Berkeley · Jun 2026UC Berkeley	UC Berkeley	15
Mar 12, 2026	Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document CollectionsOn 2,250 questions over 800 PDFs, the best agents match human accuracy but lean on brute-force search rather than planning.Snowflake · Mar 2026Snowflake	Snowflake	1

About this paper

Authors: Peng Xia, Rujun Han, Zifeng Wang and 11 more
arXiv: 2609.24972 · PDF
Citations: 0, 0 influential · Semantic Scholar
Upvotes: 180 · Hugging Face
Code: github.com/google-research/rrsi
Lab: Google · on Companies · on Acquisitions · on Paydays · on TechConf · on Releases

Changes

Date	What changed
Sep 24, 2026	New paperAdded to the listSep 24, 2026

Sources: each lab's own papers and arXiv, with citation and upvote counts from Semantic Scholar and Hugging Face. One-line summaries are for orientation, not a substitute for the paper. Logos via logo.dev; trademarks belong to their owners.

What it shows

More from Google

Other agents and evaluation papers

Changes

New papers by email