Unverified50% confidenceOpinionExact time
The history concatenation approach leads to context bloat and dilution of key information as the number of dialogue turns increases.
1
Sources
50%
Confidence
Long-term
Relevance
10/29/2026
First Seen
Sources
Related Entities
Related Claims
Unverified在多轮对话场景中，随着历史消息累积，系统指令相对业务内容的比重下降，格式遵从度随之降低79% similarUnverified在一个对话窗口里，提问轮次越多，后面每一轮消耗的Token就越大，因为历史对话会被拼接重新发送77% similarVerified大语言模型在长时间对话中会因上下文窗口限制而丢失早期信息75% similarUnverified传统对话历史模式的局限性包括：无法原生表达工具调用的中间状态、缺乏时序元数据、不支持并发事件流、在上下文窗口有限时难以进行选择性压缩和摘要75% similarUnverifiedThe topic drift phenomenon that occurs in long-text generation by large language models is academically referred to as Semantic Drift, and its root cause lies in the autoregressive generation mechanism.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1112737API
curl https://kongchang.com/api/v1/knowledge/claims/1112737MCP
get_claim(id=1112737)