Before You Say It: Predicting Your Next Move from Everyday Conversations
This presentation explores a novel approach to anticipating how people communicate by mining structured behavioral patterns from longitudinal conversational data collected via smartwatches. Instead of predicting exact words, the system predicts communicative functions like deflection, disclosure, or advisement by learning situation-specific rules with explicit exceptions. The research demonstrates that pattern-conditioned prediction substantially outperforms both zero-shot and full-history prompting, achieving especially strong results on behaviors people want to change.Script
Can an algorithm predict not what you will say, but how you will say it, based on weeks of your everyday conversations? This paper shows that language models can anticipate your communicative behavior by learning the situational rules that govern when you deflect, disclose, or redirect.
The researchers collected over 1,000 hours of naturalistic speech from 14 participants wearing always-on smartwatches for 7 to 10 days. After cleaning and participant review, this yielded nearly 10,000 utterances capturing real interactions across home, work, and social settings.
The core innovation is Situational Reasoning, a method that mines behavioral patterns in the form: if situation, then behavior, but not when exception. A pattern might say the participant deflects when challenged by authority, but not when challenged by a peer. These rules are grounded in specific examples and assigned confidence scores based on supporting and contradictory evidence.
Pattern-conditioned prediction achieved a mean score of 0.597, outperforming zero-shot prediction by 29 percent and full conversation history by 19 percent. On behaviors participants flagged as wanting to change, the gain was even larger: 41 percent over zero-shot and 39 percent over all-in-context prompting.
Prediction accuracy improved by 24 percent as more conversational data accumulated, while baseline methods remained flat. This result challenges the assumption that long-context prompting alone will recover person-specific behavioral regularities. Simply giving the model more data is not enough; it must be organized into retrievable, conditional structure.
This work shows that anticipating behavior requires more than raw conversational history. It requires structured, situation-specific models that users can inspect, correct, and control. To explore how pattern-conditioned prediction could shape proactive conversational systems, visit EmergentMind.com and create your own video from the latest research.
