Unverified50% confidenceFactExact time
The research community refers to the phenomenon of the same model giving answers of significantly varying quality in response to subtly differently worded instructions as 'prompt sensitivity'.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified显式反馈依赖用户主动参与导致样本偏差，隐式反馈解读需要因果推断，成熟团队会叠加两类信号并结合A/B测试76% similarUnverifiedSensitivity stemming from the evaluation method may arise from the scoring rubric favoring certain phrasings, a single reference answer format, or an automatic evaluator that parses output format too rigidly.74% similarUnverifiedOn a benchmark that is sensitive to prompt wording, leaderboard rankings may more often reflect how well each model happens to match the evaluator's preferences rather than true capability.73% similarUnverifiedThere is a systematic bias in annotator preferences, with higher scores given to polite, clearly structured, and risk-averse responses.70% similarUnverifiedCUE supports two evaluation modes: user-conditioned session replay, and metric consistency under new persona sampling.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1000345API
curl https://kongchang.com/api/v1/knowledge/claims/1000345MCP
get_claim(id=1000345)