Enhancing Speech Large Language Models through Reinforced Behavior Alignment
Verifiable Invariant Reinforced Behavior Alignment (VIRBA) optimizes Speech Large Language Models by aligning their behavior across different acoustic realizations of the same spoken request. It uses a reinforcement learning framework that rewards models for semantic preference, rule-verifiable correctness, cross-acoustic invariance, and adaptive reasoning. This moves models beyond simple imitation toward robust reasoning policies stable across varied speech conditions.