MarkTechPost ★ 109 4 min

Google Research Open-Sources RRSI: AI Agents That Improve Their Own Harness Without Overfitting

Agentic AIAI AgentsAI InfrastructureHarnessLanguage ModelLarge Language ModelMachine LearningOpen SourceSoftware Engineering

🔗 https://www.marktechpost.com/2026/09/29/google-research-open-sources-rrsi-ai-agents-that-improve-their-own-harness-without-overfitting/

📌 【Google Research】RRSI 開源:讓 Agent 修改自己的 Harness 卻不作弊

TL;DR:RRSI 讓 LLM agent 自行改寫 prompt、工具與流程,用正規化避免針對單一評測過擬合,程式碼已開源。

讓 AI agent 自己重寫操作手冊,聽起來像是危險的權限開放,因為它很可能只是學會怎麼在考卷上拿高分——RRSI 想解決的正是這個問題。

🤔 自我改良的 Harness,如何避免變成應試教育

Google Cloud AI Research 與 UNC-Chapel Hill、Stanford、Washington University in St. Louis 合作,釋出 RRSI(Regularized Recursive Self-Improvement)。它讓 LLM agent 重寫自己的 harness,包括 prompt、工具、記憶、控制流程與子 agent,而模型權重本身不變。問題在於,harness 演化迴圈的作法是提出修改、在固定的 evolve set 上評分、保留贏家,而同一批任務會被重複使用,迴圈很容易記住這些任務。研究團隊將這類失敗歸納成三種模式:針對特定 benchmark 過度擬合、追逐雜訊,以及複雜度不斷累積,每一種都會拉大 evolve set 分數與實際遷移表現之間的落差。

🧩 不是限制修改內容,而是正規化搜尋過程本身

RRSI 的做法是保持 harness 的每個元件都可編輯,但對搜尋過程本身加上正規化限制。研究團隊把這些機制類比為經典的正規化方法:編輯預算對應 L0,修剪機制對應 Lasso(L1),成本規則對應 Ridge(L2)。

🧩 怎麼用

程式碼採 Apache 2.0 授權,需要 Python 3.10 以上環境,並接受任意 LiteLLM 模型字串,預設假設在 Vertex AI 上使用 Claude Opus 4.8。

📊 六個 held-out 分割全數進步,token 用量也更省

在全部 6 個 held-out 分割上,RRSI 都帶來進步。以 Gemini 3.5 Flash 作為 policy 模型時,Terminal-Bench 2.1 從 64.6 提升到 78.7,SWE-bench Verified 從 76.8 提升到 79.0。產生出來的 harness 也更精簡:在 agentic workspace instance 上,RRSI 每次試驗用掉 2.42M policy token,未加正規化的演化版本則用掉 3.80M。值得注意的是,論文摘要與專案頁面對這項省幅的說法並不一致,分別寫著減少 30% 與 36%。另外,Meta-Harness 在 Harvey LAB 的 evolve split 上表現領先,而 RRSI 雖然在 evolve set 上的進步幅度是所有方法中最小的,卻是唯一一個 OOD 平均分數(JobBench、GDPval、APEX-Agents 三者平均)比 baseline H0 高出超過 1 分的方法。

💡 evolve set 進步最少,遷移表現卻最好

這組結果點出了 RRSI 設計理念的核心:在自我改良迴圈裡,真正決定能否遷移到未見過任務的,不只是「改了什麼」,更是「怎麼搜尋、怎麼保留改動」這個過程本身有沒有被約束。犧牲一部分在固定評測集上的優化空間,換來的是更真實的跨基準遷移能力。

🎯 實務啟示

對正在打造會自我修改的 agent harness(例如自動化 prompt 工程、工具選擇、子 agent 編排)的團隊來說,RRSI 提醒了一個容易被忽略的風險:光靠固定的 evolve set 評分機制去挑選「贏家」修改,很容易讓系統學會討好評測集而非真正變強。程式碼已開源且相容任意 LiteLLM 模型字串,適合作為在自家 agent 系統上實驗正規化式自我改良的起點。

🔗 來源

#GoogleResearch #AIAgents #OpenSource #SelfImprovement #LLM #AgenticAI #MachineLearning #Regularization #SWEBench #TerminalBench

原始資料 MarkTechPost · 收集於 2026-09-29
來源原標題
Google Research Open-Sources RRSI: AI Agents That Improve Their Own Harness Without Overfitting
作者
Asif Razzaq
原始標籤
Agentic AI · AI Agents · AI Infrastructure · AI Shorts · Applications · Artificial Intelligence · Editors Pick · Harness · Language Model · Large Language Model · Machine Learning · New Releases · Open Source · Software Engineering · Staff · Tech News · Technology
原始連結
https://www.marktechpost.com/2026/09/29/google-research-open-sources-rrsi-ai-agents-that-improve-their-own-harness-without-overfitting/

摘要原文

Google Cloud AI Research , with UNC-Chapel Hill, Stanford and Washington University in St. Louis, has released RRSI (Regularized Recursive Self-Improvement) . It lets an LLM agent rewrite its own harness: prompts, tools, memory, control flow and sub-agents. Model weights never change. RRSI constrains the improvement loop itself, so gains hold on benchmarks the agent never optimized against. Deployable? Yes, as a research framework. The code is Apache 2.0 , needs Python 3.10+, and accepts any LiteLLM model string. Defaults assume Claude Opus 4.8 on Vertex AI. Harness evolution loops propose edits, score them on a fixed evolve set and keep the winner. The same tasks are reused every round, so the loop can memorize them. The RRSI research names 3 failure modes: benchmark-specific fitting, noise chasing and complexity accumulation. Each one widens the gap between evolve-set scores and real transfer. RRSI keeps every harness component editable. It regularizes how the search moves instead. The research team frame these as analogies to classic regularizers. The edit budget maps to L0, pruning to Lasso (L1) and the cost rule to Ridge (L2). All 6 held-out splits improved. With Gemini 3.5 Flash as the policy, Terminal-Bench 2.1 rose from 64.6 to 78.7. SWE-bench Verified rose from 76.8 to 79.0. The harness is also lighter. On the agentic workspace instance, RRSI uses 2.42M policy tokens per trial. Unregularized evolution uses 3.80M. The abstract reports this as 30% fewer; the project page says 36%. Scores come from Table 1 of the RRSI research paper. All methods share the same starting harness, policy, evolve split and candidate budget. *Per the RRSI research team. OOD average is the mean of JobBench, GDPval and APEX-Agents, computed from Table 1 . Meta-Harness leads on the Harvey LAB evolve split. RRSI has the smallest evolve gain but the only OOD average more than 1 point above H0.

由 claude-code/sonnet 自動生成