人工智能中的递归式自我改进：从有界自我修订到自主研究循环

中文速览

AI 正在从修改答案、用自己生成的数据训练，走向改写自身工具乃至参与 AI 研究，但“自我改进”这个说法常把目标和风险截然不同的做法混在一起。作者梳理了 2024—2026 年的 1,250 篇论文，按系统改进的对象和改进过程由人把关还是自行闭环，整理出一套分类，并把评价器——也就是判断“变好没有”的标准和工具——单独拿出来分析。综述发现，大多数已有成果仍是有外部标准、有人监督的有限改进；形式化验证最可靠，单靠模型自评最弱，而评价信号不可靠时容易出现自我确认、数据质量下降和多样性缩水。厘清这些界限很重要：它既能说明当前 AI 自我改进实际走到了哪一步，也能帮助识别真正开放式递归改进的安全风险，并指出该领域最缺的是可用于治理的可靠测量。

摘要

人工智能系统日益参与自身改进：修订输出，在部署过程中调整并演化自身的智能体框架（支架系统），利用自己生成的数据进行训练，并且——在一股不断壮大的研究脉络中——亲自开展人工智能研究。描述这种参与的文献数量激增，但其术语体系（“自我修订”“自我奖励”“自我博弈”“自我演化”）却将本质上不同的目标混为一谈。我们调查了 1,250 篇 arXiv 论文（2024—2026 年），并沿两个维度对其进行梳理：系统改进的对象——部署中的行为、通过训练形成的策略、评估器，或研究过程本身；以及闭环程度——从人在环路中到完全闭环。该分类法将有界自我修订——趋于收敛、可评估且已成为工业实践——与开放式递归自我改进（RSI）区分开来；后者仍受制于落地要求、崩塌动态，以及当前各方证据所能测量到的计算资源限制。该分类法的独特之处在于专设自我评估类别：每个改进闭环都意味着一种主张，即某种信号能够替代人类判断。因此，我们不仅考察其所监督的闭环，也调查评估器的设计空间——评判器、过程奖励模型、验证器、评分标准、元评估——并将这些信号按验证层级排序，从形式化验证器（最强）到内在自我评估（最弱）。我们观察到，已证实的自我改进能力与这一层级相对应；其典型失效模式（自我确认闭环、模型崩塌、多样性崩塌）源于对该层级的违背；而让人类继续参与闭环的“研究方向设定”瓶颈，则分为两类问题：层级所标示的验证问题，以及该层级无法涵盖的先验问题——究竟什么值得评估。我们将技术文献与有关 RSI 极限的理论，以及前沿实验室关于闭环的叙述所引发的安全与治理问题联系起来，并指出，对自我改进进行符合治理需求的测量，是该领域研究最稀缺的方向。

AI systems increasingly participate in their own improvement: revising their outputs, adapting and evolving their own harnesses during deployment, training on data they generate, and — in a growing research thread — conducting AI research itself. The literature describing this participation has exploded, but under a vocabulary (“self-refine,” “self-reward,” “self-play,” “self-evolve”) that conflates fundamentally different ambitions. We survey 1,250 arXiv papers (2024–2026) and organize them along two axes: what the system improves — its behavior in deployment, its policy through training, its evaluator, or the research process itself — and the degree of loop closure (human-in-the-loop to fully closed). The taxonomy separates bounded self-refinement — convergent, evaluable, and already industrial practice — from open-ended recursive self-improvement (RSI), which remains bounded by grounding requirements, collapse dynamics, and compute constraints on every side current evidence can measure. Its distinctive feature is a dedicated category for self-evaluation: every improvement loop is a claim that some signal can substitute for human judgment, so we survey the evaluator design space — judges, process reward models, verifiers, rubrics, meta-evaluation — alongside the loops it supervises, order the signals into a verification hierarchy from formal verifiers (strongest) to intrinsic self-assessment (weakest), and observe that demonstrated self-improvement strength tracks this hierarchy, that its characteristic failure modes (self-confirming loops, model collapse, diversity collapse) follow from its violations, and that the “research direction-setting” bottleneck that keeps humans in the loop divides into a verification problem the hierarchy indexes and a prior one — choosing what deserves evaluation at all — that it does not. We connect the technical literature to the theory of RSI limits and to the safety and governance questions raised by frontier-lab accounts of closing the loop, and identify governance-grade measurement of self-improvement as the field’s most underpopulated niche.

术语表

recursive self-improvement (RSI)

递归式自我改进（RSI）

self-refinement

自我修订

self-rewarding

自我奖励

self-play

自我博弈

self-evolution

自我演化

self-evaluation

自我评估

loop closure

闭环程度

bounded self-refinement

有界自我修订

open-ended recursive self-improvement

开放式递归自我改进

large language model (LLM)

大语言模型（LLM）

agent

智能体

harness (scaffolding)

智能体框架（支架系统）

evaluator

评估器

verifier

验证器

judge

评判器

reward model

奖励模型

process reward model

过程奖励模型

meta-evaluation

元评估

verification hierarchy

验证层级

test-time training (TTT)

测试时训练（TTT）

inference-time refinement

推理时修订

self-generated data

自生成数据

model collapse

模型崩塌

diversity collapse

多样性崩塌

Gödel machine

哥德尔机

intelligence explosion

智能爆炸

FunSearch

FunSearch

AlphaEvolve

AlphaEvolve

Self-Refine

Self-Refine

STaR

STaR