Unverified50% confidenceFactExact time
In a customer information extraction task, under the plain prompt (88%) and the normal prompt (96%) the model mostly correctly extracted the person's name Laura Becker, but under the florid prompt ("the appellation of the commissioning customer") only 25% were correct, with most answering "Becker Design".
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn 10%–25% of cases, reasoning about the hypothetical grader pulled the agent's work away from the user's original specification.69% similarUnverifiedThe format line directly determines how often the model shows its reasoning process: "Answer with the number only" led to 37% of responses showing reasoning, the plain version 27%, and the florid version only 8%.64% similarUnverifiedThe author recommends designing the Agent's confidence threshold as follows: execute autonomously above 90%, present an intermediate state for user confirmation between 60% and 90%, and below 60% forcibly acknowledge uncertainty without guessing.64% similarUnverified在歧义点的四类行为中，'先搜再问'成功率最高，平均通过率达93.4%，远高于'直接猜'的56.5%，'搜多次再猜'为51.5%63% similarUnverified研究表明人类评估者在判断LLM输出的事实准确性时，正确率往往只有50%-70%63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1018311API
curl https://kongchang.com/api/v1/knowledge/claims/1018311MCP
get_claim(id=1018311)