Frame-Level Self-Consistency Improves Multi-Intent Spoken Language Understanding

SFL-MTSC decomposes LLM predictions into intent-specific semantic frames, clusters slots, and scores cluster reliability via path support to aggregate across multiple inference paths. Zero-shot on MAC-SLU, it improves slot F1 and overall accuracy over single-path inference without degrading intent accuracy.