Unverified50% confidenceFactExact time
A wider gap between the similarity distributions of positives and negatives does not necessarily translate into better retrieval performance; there is a gap between distribution-level 'looking good' and task-level 'working well' (termed 'distribution illusion').
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe approach of combining reward and diversity into a single aggregate score for optimization has a flaw: high diversity can directly offset low reward, meaning the batch may still contain inferior images that clearly deviate from user preferences.74% similarUnverified少量高质量的偏好数据往往比大量低质量数据能产生更好的对齐效果73% similarUnverifiedA study proposes that differences between representational similarity metrics are far less important than the differences introduced by the objects of comparison (i.e., how data is organized).71% similarUnverifiedThe salience boost is blended into the embedding retrieval score multiplicatively rather than additively, to ensure relevance still dominates and to prevent irrelevant content from being forcibly surfaced.69% similarUnverifiedVector similarity measures semantic proximity, not temporal validity or whether information has been superseded.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1028175API
curl https://kongchang.com/api/v1/knowledge/claims/1028175MCP
get_claim(id=1028175)