Unverified50% confidenceFactExact time
SIRS (Single-Increment Rejection Sampler) randomly samples one increment and uses a single query to estimate the log acceptance probability of the entire block, thereby reducing query cost while preserving unbiasedness.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMRR (mean reciprocal rank) measures how early the first relevant result appears, taking the reciprocal of the rank of the first relevant document for each question and then averaging them, completely ignoring other relevant documents after the first.69% similarUnverified置换检验通过大量随机置换（通常1000到10000次）构建零假设下的经验分布来估计p值，无需参数假设但计算成本相对较高69% similarUnverifiedTop-p (nucleus sampling) filters out extremely low-probability outputs by truncating the long tail of the probability distribution.66% similarUnverifiedFor end-user-facing reply-type models, a maximum token count per sample (e.g., 512) can be set, with the excess truncated or discarded, to suppress the gradient weight of long reasoning samples and let the model learn to complete answers within a limited length.66% similarUnverifiedAs a probabilistic model, the AI reviewer Agent's pass judgment on the same output is random, and multiple reviews may yield different conclusions.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1042095API
curl https://kongchang.com/api/v1/knowledge/claims/1042095MCP
get_claim(id=1042095)