Unverified50% confidenceFactExact time
Subsequent research has pointed out that DPO is prone to overfitting spurious correlations in preference data on out-of-distribution samples, and that the degree of over-optimization is harder to monitor due to the lack of an explicit reward model.
1
Sources
50%
Confidence
Long-term
Relevance
9/6/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedSome models carry the risk of over-optimization toward test sets, causing leaderboard scores to deviate from real-world performance (benchmark saturation).76% similarVerified若用含测试集的全量数据计算均值和标准差，会导致预处理泄漏，使评估指标虚高，无法真实反映模型泛化能力74% similarUnverifiedThe overconfidence of a model answering directly while skipping retrieval is technically referred to as miscalibration, meaning there is a systematic deviation between confidence and actual accuracy.73% similarUnverified若将验证集或测试集数据纳入 fit，会造成数据泄露，导致模型评估结果过于乐观72% similarUnverifiedSome vendors tend to over-optimize for evaluation sets (benchmark gaming/benchmark contamination), leading to inflated scores but insufficient generalization ability.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/867158API
curl https://kongchang.com/api/v1/knowledge/claims/867158MCP
get_claim(id=867158)