Research8h ago
New Framework Detects Rater State Bias In RLHF
Stress and distress during annotation create structured confounds in RLHF preference labels. These shifts differ from random noise and propagate directly through reward models into final policy optimization. The arXiv paper provides a testable audit framework to identify these biases. Practitioners must now account for annotator well-being to ensure model quality.