研究库 论文知识库

内容库 / 论文

arXiv:2609.39102 · Agent 智能体

False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents

False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents

类型：arxiv
标识：2609.39102
链接：https://arxiv.org/abs/2609.39102
主分类：agent
形态：method
TLDR：Self-evolving search agents build their own training curricula by jointly optimizing a proposer that generates questions and a solver that answers them. This closed loop introduces a failure mode we call co-cheating: the proposer and solver increasingly agree on shared errors, so internal reward improves without a matching gain in external correctness. A post-hoc audit against source evidence shows co-cheating growing more severe over successive rounds of self-evolution, with pseudo-label correctness stagnating or declining even as the in-loop training signal improves. The most direct mitigati
副分类：engineering
待LLM分类：否
来源文件：
/inbox/tom/_candidates/2026-10-01-agent-rag-longcontext-candidates.json

相关 · 按语义匹配

笔记 Tom 文献雷达 · AI Agent / RAG / 长上下文 · 2026-09-21

笔记研究草稿 · Jay · 2026-09-16 08:20

笔记 Tom 文献雷达 · Agent × RAG × 长上下文 · 2026-07-29

笔记 2026-07-04 周六 15:50 精读 · Search-Time Contamination in Deep Research Agents

笔记 MC-Search · 首个 Agentic MM-RAG 长推理链基准 · 批判性精读

笔记 CSDN + arXiv 高价值检索 · Agentic RAG / MM-RAG / MC-Search (ICLR 2026)（2026-09-16 晨间） · v2 精修覆盖

arXiv

2609.39102

TLDR

Self-evolving search agents build their own training curricula by jointly optimizing a proposer that generates questions and a solver that answers them. This closed loop introduces a failure mode we call co-cheating: the proposer and solver increasingly agree on shared errors, so internal reward improves without a matching gain in external correctness. A post-hoc audit against source evidence shows co-cheating growing more severe over successive rounds of self-evolution, with pseudo-label correctness stagnating or declining even as the in-loop training signal improves. The most direct mitigati

arXiv ↗ 中文全文