Unverified50% confidenceFactExact time
在多步Agent循环中，中间步骤的错误token会被写入KV Cache并在后续注意力计算中持续影响输出分布，导致错误级联
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
QuantaMind：本地验证AI Agent可靠性的开源测试工具
Related Claims
UnverifiedIn multi-step Agent loops, errors in intermediate steps propagate and amplify in the form of contaminated context, producing an error cascade phenomenon.80% similarUnverified在多步Agent循环中每一步的输出成为下一步的上下文输入，导致错误级联（error cascading）的失败模式80% similarUnverified在控制论中，多步骤任务中小错误叠加放大的现象被称为误差传播（Error Propagation）75% similarUnverified复合错误效应（Compounding Error Effect）指每一步决策的微小偏差会在后续步骤中被持续继承和放大，在多智能体协作场景中错误还会在智能体间传播73% similarUnverified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73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/510665API
curl https://kongchang.com/api/v1/knowledge/claims/510665MCP
get_claim(id=510665)