The Inner Thoughts framework for making a conversational AI proactive in multi-party human-AI dialogue, deciding on its own when to speak instead of only replying when addressed or predicting the next speaker from turn-taking cues. In parallel to the visible conversation the AI maintains a continuous covert stream of candidate thoughts; a loop of thought formation, evaluation by an intrinsic-motivation-to-express signal, a participation decision about whether and when to contribute given the turn-taking state, and articulation into an utterance governs its behaviour. Instantiated in two real-time LLM-driven systems (an AI playground web app and a chatbot), it outperforms next-speaker-prediction baselines on anthropomorphism, coherence, perceived intelligence, and turn-taking appropriateness.
This human-computer-interaction paper tackles how a conversational AI can be genuinely proactive in multi-party human-AI conversations, deciding on its own when to speak rather than only replying whe…