Google论文：正则化递归自改进防止Agent过度拟合
Must-read paper from Google on self-improving agent harnesses.
Agent工程必读，RRSI方案用正则化有效解决自动化调优导致的过拟合问题，数据详实可直接参考其防过拟合机制。
Must-read paper from Google on self-improving agent harnesses.
Google 关于自我改进型智能体工具链的必读论文。
If you auto-optimize your agent's harness, your eval score can go up while the agent gets worse on real tasks.
如果你自动优化智能体的工具链，你的评估分数可能会上升，但智能体在实际任务上的表现却会变差。
This paper shows how to prevent that.
这篇论文展示了如何防止这种情况发生。
Of five harness-evolution methods compared on agentic workspace tasks, RRSI scored the lowest on the tasks it evolved against and highest on all three out-of-distribution benchmarks.
在针对智能体工作区任务的五种工具链进化方法中，RRSI 在其进化的目标任务上得分最低，但在所有三个分布外基准测试上得分最高。
Automated harness evolution proposes edits to prompts, control flow, tools and memory, keeps the ones that raise the score, and repeats.
自动化工具链进化会对提示词、控制流、工具和记忆提出修改建议，保留那些能提高分数的修改，并重复此过程。
The authors show this overfits the training tasks. Meta-Harness reached 93.0 on the Harvey LAB evolve split but gained only 0.3 to 1.5 points on JobBench, GDPval and APEX-Agents.
作者表明这种方法会对训练任务过拟合。Meta-Harness 在 Harvey LAB 进化分割集上达到了 93.0 分，但在 JobBench、GDPval 和 APEX-Agents 上仅获得了 0.3 到 1.5 分的提升。
RRSI adds regularization on both sides of the loop.
RRSI 在循环的两端都添加了正则化。
The proposer gets an edit budget that shrinks over time and is pushed toward directions it has not tried. A critic rejects benchmark-specific edits, and a pruner removes edits that are too small, too costly or no longer useful.
提议者拥有一个随时间推移而缩减的编辑预算，并被推向它尚未尝试过的方向。批评家会拒绝针对特定基准的修改，修剪器则会移除那些改动太小、成本过高或不再有用的修改。
RRSI scored 90.5 on the evolve split and gained 3.5 to 4.7 points on the three held-out benchmarks. In the ablation, unregularized evolution used 3.80M tokens per trial against 2.42M for RRSI. With Gemini 3.5 Flash, RRSI raised Terminal-Bench 2.1 from 64.6 to 78.7 and carried a 2.2-point gain over to SWE-bench Verified.
RRSI 在进化分割集上得分为 90.5，在三个预留基准测试上提升了 3.5 到 4.7 分。在消融实验中，无正则化的进化每次试验使用了 380 万 token，而 RRSI 使用了 242 万 token。使用 Gemini 3.5 Flash 时，RRSI 将 Terminal-Bench 2.1 从 64.6 提升至 78.7，并在 SWE-bench Verified 上保持了 2.2 分的优势。
Paper: https://arxiv.org/abs/2609.24972
论文：https://arxiv.org/abs/2609.24972
Chat with Paper: https://academy.dair.ai/papers/rrsi-regularized-recursive-self-improvement-of-agent-harnesses-2609.24972
与论文对话：https://academy.dair.ai/papers/rrsi-regularized-recursive-self-improvement-of-agent-harnesses-2609.24972
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力