Train MiniCPM-o 4.5 to choose Listen or Speak on egocentric audio/video. SFT + DPO refines timing and restraint under a supplied task recipe.
Editorial summary · not the author abstract. Read the linked paper for the complete abstract and claims.EgoVoice first-person multimodal stream, silence or intervention decision and spoken guidance Original figure · paper authors ↗
Release evidence
Official project page schedules code, models and data for November 2026. A base-model release is not an EgoVoice checkpoint release.