Unverified50% confidenceFactExact time
When the data distribution is balanced, D* approaches a certain baseline value, and the new formula reduces to a result close to the traditional formula; when the data are severely imbalanced or skewed, D* deviates significantly from the baseline, leading to large adjustments in the estimate of predictive performance.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSubsequent 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.73% similarUnverifiedWhen covariates are imbalanced or skewed, some observations become high-leverage points that exert a disproportionate influence on the regression coefficients, so that even with the same N and p, out-of-sample predictive performance can differ greatly.72% similarUnverifiedHigh-variance overfitting should be mitigated through regularization, feature reduction, and adding data; when the data itself is highly volatile, interval forecasting should be used instead.71% similarUnverifiedUnder off-policy sampling conditions, when the preference margin is aligned in direction with the utility gradient, DPO's update shifts from a corrective signal to a reinforcement-style update similar to reinforcement learning.69% similarUnverifiedWhen B is fixed and M increases, the bits allocated per fact decrease and distortion rises, which explains why feeding more data does not necessarily improve accuracy linearly.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1026676API
curl https://kongchang.com/api/v1/knowledge/claims/1026676MCP
get_claim(id=1026676)