Unverified50% confidenceTradeoffExact time
ReAct的线性思考-行动循环在需要并行探索多条路径的任务时效率低下，且每步推理的Token消耗累积导致长任务计算成本呈二次方增长
1
Sources
50%
Confidence
Long-term
Relevance
7/22/2026
First Seen
Sources
Related Claims
Unverified实验表明ReAct在需要多步检索的问答任务中显著优于纯推理或纯行动基线，但代价是Token消耗的线性增长81% similarUnverifiedReasoning Effort档位越高，模型可消耗的思考Token上限越大，但延迟和Token成本相应线性乃至超线性增长79% similarUnverified长程Agent任务的实现依赖ReAct框架等推理-行动循环，错误会在链条中级联放大，导致完成率远低于单步问答78% similarUnverified长程任务的Token消耗随步骤数呈近似二次方增长，而非线性增长77% similarUnverifiedReflexion由于需要维护跨轮次反思历史，上下文长度会随循环次数线性增长，在长时间运行任务中可能引发显著延迟与成本问题77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/594066API
curl https://kongchang.com/api/v1/knowledge/claims/594066MCP
get_claim(id=594066)