ProactiveLLM: Learning Active Interaction for Streaming Large Language Models

In three linesProactiveLLM learns when to interact with streaming inputs without external signals. Through monotonic random masking and synchronized privileged self-distillation, the model perceives semantic sufficiency from partial inputs. Reduces interaction latency while maintaining quality on text and speech streaming tasks.

Summary generated by Claude — human-verified