Unverified50% confidenceCautionExact time
工具调用次数过多会导致上下文污染，使模型难以追踪任务目标
1
Sources
50%
Confidence
Long-term
Relevance
7/25/2026
First Seen
Sources
Related Claims
Unverified模型有时会陷入过度思考，在特定任务中产生超过52倍的额外token73% similarUnverifiedData leakage falls into two categories—target leakage and train-test contamination—and leads to systematic overestimation of model performance.72% similarUnverified工具过于细粒度会导致规划步骤爆炸，工具过于粗粒度则会牺牲灵活性，这是Agent系统设计中最常见的权衡问题之一70% similarUnverifiedTraining on a dataset with an overly homogeneous object type causes the model's performance to drop sharply when faced with objects of other materials or forms, which is the domain shift problem.70% similarUnverifiedTool parameter descriptions that are too broad cause the model to hallucinate parameters, while descriptions that are too strict limit the tool's applicability and flexibility.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/615576API
curl https://kongchang.com/api/v1/knowledge/claims/615576MCP
get_claim(id=615576)