Faithful or Fabricated? A Causal Framework for Rationalization Bias in LLM Judges

In three linesStudy on rationalization bias in LLM judges. Researchers test whether model explanations remain stable when non-evidential cues are perturbed (verbosity, confidence). They propose PROOF-BEFORE-PREFERENCE to improve cue invariance and reduce explanation anchoring.

Summary generated by Claude — human-verified