Unverified50% confidenceFactExact time
Mechanism two: in the intermediate regime, pooling the generation counts of different questions is more effective than estimating each answer distribution separately, and only logarithmic gains can be obtained from additional questions.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedPractical implication: in test-time ensembles, when questions are scarce one should prioritize expanding the question set rather than blindly increasing the number of generations; when questions are abundant, the improvement from the number of generations encounters an identification ceiling.75% similarUnverifiedPooling is superior to per-problem tuning: when there is a common shape to borrow across problems, sharing information can significantly reduce overall calibration error, while tuning each problem independently incurs a clear cost when labels are few.74% similarUnverifiedThe statistical intuition of pooling: merging the generation counts of all questions to estimate shared weights is equivalent to using more data to estimate fewer parameters to reduce variance, but when the relative merits of models differ greatly across questions, forced sharing introduces bias.74% similarUnverified在垂直领域（如法律、医疗）表现不佳时，通过收集高质量问答数据进行微调往往比无限优化Prompt更有效且更可持续72% similarUnverifiedEqual numbers of tokens do not equal equal computation; a conversational scheme may appear better simply because it consumed more sequential reasoning steps.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1046166API
curl https://kongchang.com/api/v1/knowledge/claims/1046166MCP
get_claim(id=1046166)