Unverified50% confidenceFactExact time
上下文回贯机制缺失（工具执行后的观测结果observation没有回写到对话上下文中）是最普遍也最隐蔽的死循环问题
1
Sources
50%
Confidence
Long-term
Relevance
7/21/2026
First Seen
Sources
Related Claims
Unverified带诊断式反馈的重试循环将具体错误作为新上下文注入下一轮对话，其成功率显著高于无信息重试，理论基础来自Self-Refine、Reflexion等LLM自我改进研究70% similarUnverified在长对话中早期的错误尝试和调试过程会持续占据上下文窗口，造成上下文污染并干扰后续推理准确性69% similarUnverified传统ReAct模式在复杂场景下存在瓶颈，长链路推理中的误差累积效应会导致任务失败率上升67% similarUnverifiedIn long-context capability tests, models processing inputs of tens of thousands of tokens may exhibit the 'lost in the middle' phenomenon.66% similarUnverified长程Agent任务的实现依赖ReAct框架等推理-行动循环，错误会在链条中级联放大，导致完成率远低于单步问答66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/575815API
curl https://kongchang.com/api/v1/knowledge/claims/575815MCP
get_claim(id=575815)