Unverified50% confidenceFactExact time
For sampled facts whose scores fall below the anchor by more than the margin, the gradient of GMR's hinge loss is the same as the gradient of replaying the correct answers to those facts; facts that have not degraded produce no retention gradient.
1
Sources
50%
Confidence
Long-term
Relevance
10/21/2026
First Seen
Sources
Related Entities
Related Claims
Unverified对比学习中的推远操作只要求负样本相似度低于正样本，并不要求达到-1的极端值72% similarUnverifiedMAPLE's probability-level metrics measure whether the ranking of the correct token (the gold token) shifts among failure cases that were not recovered; even if the answer is ultimately not corrected, a rise in the gold token's ranking indicates that the technique is moving closer to the correct answer.71% similarUnverifiedSAGE outperforms the simple baseline on rank correlation and has a lower error rate.70% similarUnverifiedSetting an absolute failure count ceiling for small-sample slices is more robust than setting a percentage threshold and is insensitive to sample size70% similarUnverifiedIn evaluation regression detection, only if the candidate's lower bound is higher than the baseline's lower bound (or the baseline point estimate) is there statistical grounds to say it has not gotten worse; otherwise it can only be judged inconclusive.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1080348API
curl https://kongchang.com/api/v1/knowledge/claims/1080348MCP
get_claim(id=1080348)