[ICML 2026] ProactiveLLM: Learning Active Interaction for Streaming Large Language Models

TL;DR

ProactiveLLM enables LLMs to actively decide when to interact in streaming settings by learning endogenous semantic sufficiency cues from partial inputs, reducing latency without external teachers, timing labels, or annotated reasoning trajectories while maintaining generation quality.

News

• [2026.6] Our paper now can be found at https://arxiv.org/abs/2606.00523.

ToDo

• Release source code.