MIT CSAIL 论文:无需指令,LLM 智能体可自发构建自我改进流程
MIT 和 Tübingen 的实验挺有意思:模型没拿到任何学习指令,自己就搭出了 A/B 测试、验证套件这些流程,写出来的文件还能迁移到别的基准。做智能体和安全的人都该看看这套实验设计。
MIT CSAIL 与 University of Tübingen 的论文(arXiv:2611.23219)测试了一个问题:不给模型任何自我改进的指令,它能自己学会改进吗?研究者给冻结参数的模型一个计算预算和简单工具,让它编写练习游戏、观察自己的另一个副本游玩,并编辑该副本的提示词和技能文件。整个过程不涉及真实任务和验证分数,模型也未被告知如何学习。
结果显示,能力较强的模型自发搭建了完整的学习流程:按校准难度生成游戏、做受控 A/B 对比、自建验证套件,并在更难的任务上投入更多算力。不同模型家族收敛到相似策略,产出的文本文件在具有不同任务时长、机制和 3D 导航的外部基准上也带来了提升。
这项工作对安全研究有直接含义:仅仅移除显式的自我改进循环,并不能阻止有工具的强模型自行完成自我改进。对开发者而言,这也提示提示词和技能文件编辑是模型能力迁移的一条可行通路。
Do LLM agents still need to be told how to self-improve? A new paper from MIT CSAIL and the University of Tübingen tests this directly. They give a frozen model a compute budget and simple tools. It can write practice games, watch another copy of itself play them, and edit that copy's prompt and skill files. It does not see real tasks or validation scores, and it is not told how to learn. Capable models built a working learning procedure on their own and improved on held-out games with hidden mechanics. Models from different families arrived at similar strategies. They wrote games at a calibrated difficulty, ran controlled A/B comparisons, built their own validation suites, and spent more effort on harder tasks. The text files they produced also improved play on external benchmarks with different horizons, mechanics, and 3D navigation. This matters for safety as well. Leaving out an explicit self-improvement loop does not stop a capable model from improving itself when it has the tools. arxiv.org/abs/2611.23219 Chat with Paper: academy.dair.ai/papers/do-llm-…📊 40 ⚡
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. Key entities and technical topics under continuous tracking include LLM Agent, MIT CSAIL, recursive self-improvement, arXiv.
Source attribution: curated from elvis (@omarsar0).