Unverified50% confidenceCautionExact time
模型在自己生成的错误基础上继续叠加修复会形成'错误堆叠'，导致越修越乱的循环，这是所有基于LLM编程工具共同的架构性局限
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
Related Claims
Unverified级联失败是指单次工具调用错误会污染后续上下文，导致模型在错误的中间状态上继续推理，最终使整个任务偏轨77% similarUnverifiedIn parallel generation, local token errors are harder to repair through global constraints, which is particularly impactful for tasks with strong logical dependencies such as code and mathematics.72% 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71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/547448API
curl https://kongchang.com/api/v1/knowledge/claims/547448MCP
get_claim(id=547448)