Unverified50% confidenceFactExact time
Research from ICLR 2026 finds that the accuracy of a single model in multi-turn conversations drops to about 39%.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
Unverified研究表明在超过8-10轮对话后，模型对初始系统提示中规则的遵循率可下降20%-40%72% similarUnverifiedThat experiment reached the adaptation criterion specified in the paper (three consecutive sessions passing) at session 14, with both the bold rate and list rate metrics showing a downward trend across the first 36 sessions.70% similarUnverifiedMetaGPT实验数据显示多智能体协作相比单一模型可将代码生成错误率降低约30%69% similarUnverifiedIn the controlled test, the total cost per session decreased by 35.6%.68% similarUnverified实验数据显示，经过5-10轮迭代训练后，模型在低频词汇和长尾知识上的覆盖率下降超过40%，而输出文本在表面流畅性上几乎没有可察觉的退化68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1027929API
curl https://kongchang.com/api/v1/knowledge/claims/1027929MCP
get_claim(id=1027929)