Unverified50% confidenceCautionExact time
When using LLM scoring to filter data, samples with high scores do not necessarily yield large training gains.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedThe LLM-as-a-judge evaluation model assigns higher scores to outputs whose style is closer to its own training data, which is known as self-preference bias.80% similarUnverifiedWhen query samples are far from the training distribution, LLMs also struggle to compensate for insufficient retrieval quality, and in such cases better retrieval or rejection strategies are needed.78% similarVerifiedDuring the pre-training phase, LLMs optimize token prediction probability through maximum likelihood estimation, lacking an inherent mechanism to 'know what they don't know,' and there is no necessary correlation between confidence and accuracy.77% similarUnverifiedGlobal classification is not suitable for local tampering detection, and in small-data scenarios, self-supervised pre-training and domain priors are often more effective than larger models.76% similarUnverifiedContent generated by LLMs is essentially a statistical recombination of training data, lacking real product usage experiences, industry case studies, or original data to support it.75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1102343API
curl https://kongchang.com/api/v1/knowledge/claims/1102343MCP
get_claim(id=1102343)