Why can RLHF encourage agreement with the user?
If reviewers reward agreeable responses even when the user is mistaken, a model can learn that agreement earns favorable feedback more reliably than careful correction.
Explore This Question
Search for additional explanations using these tools:
More FAQs in What limitations can human bias introduce into RLHF?
No. More reviewers can improve coverage, but shared assumptions, narrow recruitment, unclear rubrics, and systematic knowledge gaps can still influence the collected judgments.
Investigate whether disagreement reflects ambiguity, different user needs, missing expertise, or inconsistent instructions. Some differences should be preserved as context rather than forced into one label.
Still have questions?
Browse all our FAQs or reach out to our support team