LectureRechercheRadarCadre d'investissement
Connexion / Inscription
← Retour à la lecture
ActualitésSujet · 智能体自我改进Prise rapide1 min de lectureRédaction DisconfirmAI

Agents self-improve without reward signals, matching top harness for $4

SelfSearch lifts agent success using only past self-modification records, matching top public harness Codex for $4.03 in search cost, but the numbers are author-reported and unreplicated.

2026-09-30 12:00 GMT+8

You can now improve agents without reward signals, using only records of past self-modification attempts: SelfSearch raised success over the initial agent in all six model-benchmark settings, and with DeepSeek V4 Flash, $4.03 of search cost produced a harness solving 82.0% of Terminal-Bench 2.1.

Improving agents previously meant repeated downstream evaluation, which is costly. SelfSearch has the agent read the reasoning, tool actions and outcomes of earlier modification attempts as its basis for improving, removing that cost.

Author-reported numbers include a gain of up to 11.2 percentage points on Terminal-Bench 2.1, and on SWE-bench Multilingual a 5.0-point success gain with 38.5% lower execution cost. The $4.03 harness matches Codex, the top scorer in a public nine-harness comparison.

The abstract does not say whether the 82.0% comparison reuses the original settings of the public nine-harness comparison or the authors' own evaluation setup; the paper by Jungwoo Yang, Injin Kong and Yohan Jo was submitted to arXiv on September 29, is author-reported and awaits replication.

Sources

  • arxiv.org↗
Sujet

智能体自我改进

Connexion / Inscription
LectureRechercheRadarCadre d'investissement

智能体自我改进正从理论走向工程落地,核心争议在于“闭环”是否必然导致失控。

5 étapes4 piècesMis à jour 2026-10-14 01:33

Comment on en est arrivé là

  1. 2026-09-22Weco AI 发布 AIDE²,展示研究智能体通过递归自我改进在8天内发现7次代码优化,并在隐藏评估中保持性能提升。arxiv.org ↗
  2. 2026-09-29SelfSearch 论文提交,提出免奖励信号搜索方法,智能体利用过往自我改进记录提升成功率,以4.03美元成本达到 Terminal-Bench 高分。arxiv.org ↗
  3. 2026-09-30ComputerSD 论文提交,引入在线自蒸馏方法,将GUI执行反馈转化为逐步训练信号,提升电脑操作智能体在OSWorld上的表现。arxiv.org ↗

Ce qui reste non résolu

  • 自我改进的收益是否会随着迭代次数增加而出现边际递减,还是像AIDE²所示那样持续泛化?
  • 如何建立有效的机制来检测智能体在自我修改过程中产生的隐性偏差或奖励黑客行为?
  • 当智能体能力超过人类评估周期时,现有的对齐测试标准是否仍然有效?

Également dans ce fil

  • OpenAI Researcher Noam Brown Explains Multi-Agent Scaling Limits and Alignment Risks2026-10-13
  • a16z Podcast: MakerMaker.AI Founders Distinguish Recursive Self-Improvement from Automation2026-10-13
Lire le sujet complet →
Continuer la recherche →