Papers

Google 预印本:AI 记录改造日志实现自我改进

Why we picked it

Google 这篇预印本让 AI 自己改 harness 并写改造日志,换模型也能接着跑,84.75 分对比 73.42 分的差距说明日志真的管用。做智能体工程的人可以直接抄这套日志方法,比每次换模型都重头调参省事多了。

Google 发布一篇预印本论文,让 AI 重写一个智能体 harness,并把每次改动的原因与效果记录成日志。

Body

实验中先由 Opus 5.5 进行改造并留下日志,再换用 Gemini 3.

8 Flash 继续迭代。结果显示,有日志时模型在 workplace tasks 基准上得 84.

75 分,没有日志时只有 73.42 分;

编程任务上,带日志的版本在 10 次运行中有 9 次胜出。论文的核心结论是:更换新模型时不必从零开始改造 harness,改造知识可以通过日志跨模型传递,形成复利效应。

这意味着递归自我改进(Recursive self-improvement)的路径正在变得可行,工程经验从模型资产变成了可迁移的独立资产。

Original · kimmonismus

This is a super interesting paper and Recursive self-improvement keeps coming closer. New Google preprint:

Google had AI rewrite an agents harness and keep a log of every change, why it made it and whether it worked. Then they swapped Opus 5.5 for Gemini 3.8 Flash and let it keep going.

With the log: 84.75 on workplace tasks, without: 73.42. In coding it came out ahead in 9 of 10 runs.

tl;dr a new model no longer means starting from scratch. That's where it compounds.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Gemini 3.8 Flash, Opus 5.5, Google, agents.

Source attribution: curated from kimmonismus (@kimmonismus). Cross-source corroboration: 2 independent sources have covered this development (including Fireworks AI), confirming industry-wide signal strength.