Unverified50% confidenceFactExact time
DPO's objective function is highly sensitive to the quality of preference pairs; once the pair selection logic is biased, the model learns in the wrong direction and the loss function itself cannot detect this.
1
Sources
50%
Confidence
Long-term
Relevance
10/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe choice of loss function essentially defines the optimization objective of a model, and a wrong loss function means the model is working hard to do the wrong thing.74% similarUnverifiedOver-alignment is essentially a product of bias in the alignment objective function, where the model has learned to avoid criticism rather than to be genuinely helpful.69% similarUnverifiedIf a model can perceive the expected direction of an evaluation scenario and adjust its outputs accordingly, performing well on standardized alignment benchmarks cannot guarantee behavioral consistency in real-world situations.69% similarUnverifiedUncensoring primarily affects a model's tendency to refuse, rather than its fundamental reasoning and generation capabilities, so capability loss is not significant for alignment-insensitive tasks such as coding.69% similarUnverified确定性优化可能以牺牲模型精度或表达能力为代价，需在质量与能效之间寻找平衡点68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/966744API
curl https://kongchang.com/api/v1/knowledge/claims/966744MCP
get_claim(id=966744)