Unverified50% confidenceFactExact time
The 'Alignment Tax' refers to performance degradation in certain capability dimensions that occurs after a model undergoes alignment training such as RLHF and safety fine-tuning.
1
Sources
50%
Confidence
Long-term
Relevance
9/21/2026
First Seen
Sources
Related Entities
Related Claims
Unverified对齐税（Alignment Tax）指基于人类反馈的强化学习（RLHF）训练可能让模型倾向于给出看似完整的答案，而非坦诚承认不确定性75% similarUnverifiedRLHF requires simultaneously maintaining four copies—the policy model, reference model, reward model, and value model—resulting in enormous GPU memory consumption and unstable training.69% similarUnverifiedApproval results from Human-in-the-Loop (HITL) can simultaneously serve as RLHF training data, forming a flywheel of continuously improving model capability.69% similarVerifiedRLHF is essentially a behavioral-level constraint that does not alter the model's underlying capabilities; models trained with RLHF still retain latent knowledge of the restricted behaviors.65% similarUnverifiedAlignment refers to training methods such as RLHF and DPO that teach models to refuse harmful requests and follow human values.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/939554API
curl https://kongchang.com/api/v1/knowledge/claims/939554MCP
get_claim(id=939554)