CASE Predicts When Hidden-State Selection Beats Majority Voting
August 19, 2026
CASE (Correctness-Axis SElection) uses a decodability measure to predict if selecting candidates via hidden-state scores will outperform majority voting. It trains a linear gate on answer-token hidden states to improve test-time information fusion.
HOW THIS AFFECTS YOU
●
builderYou can improve LLM response accuracy by switching from majority voting to hidden-state selection when decodability is high.
●
researcherYou can use decodability as a metric to determine the reliability of hidden-state selection methods.