kyry.me/projects/mi-rol

← projects

P-08 · accepted · co-author

Mi-RoL

A multi-tier LLM framework for empathetic conversation on the Pepper robot, grounded in Motivational Interviewing. Tested in a 40-participant study.

Study flow: consent and pre-survey, conversation with Pepper using speech, affect and visual cues, then survey and interview
CH1mi-rolStudy flow from the paper: pre-survey, conversation with Pepper, then survey and interview.

Summary

Co-author. This is joint work with Raiyan Ashraf (lead), Yanni Liu, Sruthi Ganji, Jong-Hoon Kim and Ju Yup Lee.

LLM-based social robots mostly react to the last thing said. Mi-RoL gives the conversation a structure based on Motivational Interviewing (MI), and grounds it in the user’s situation: what the robot sees, what it knows about the user, and what they talked about before.

How it works

The framework splits reasoning from response. At each stage boundary a bigger reasoning model looks at the context, estimates engagement and emotion, picks the strategy and decides when to move on. A fast response model says the actual lines in between. The stages are Intro, Engage, Affirm & Reflect, Evoke and Expand.

Speech is transcribed with Whisper and Silero voice activity detection. GPT-4o picks out the user’s appearance and the objects around them, and that goes into one shared context together with the user’s personal info.

Mi-RoL system overview: personal information, visual cues and speech feed a unified context; a reasoning layer and a response layer drive five MI stages
System overview: the reasoning layer plans and moves between MI stages, the response layer talks.

Nine model pairings were compared first: three reasoning models (GPT-4o, Gemini 2.5 Pro, Llama 3 8B) against three response models (GPT-4o-mini, Gemini 2.5 Flash, Llama 3.2 3B), rated by four researchers. GPT-4o for reasoning with Gemini 2.5 Flash for responses came out on top and was used in the study.

Study

40 participants, between-subject, with a physical Pepper. Each person had a 7–10 minute conversation with either Mi-RoL or a baseline: unstructured, reactive dialogue from a single GPT-4o.

Baseline Mi-RoL effect size d
Perceived empathy 3.46 3.90 0.68
User engagement 3.58 4.06 0.75
Social presence 3.15 3.88 0.88
Intention to use 3.55 4.12 0.65

All four were significantly better with Mi-RoL (5-point scales, p < .05), and social presence improved the most. People liked it most when the robot noticed something they had with them, like a backpack or a tumbler, and connected it to their own stories.

Paper

  • Can Motivational Interviewing Guided Dialogue Improve Empathy in Human–Robot Interaction? A Controlled Study with Baseline ComparisonRaiyan Ashraf, Yanni Liu, Kyrylo Kolesnichenko, Sruthi Ganji, Jong-Hoon Kim, Ju Yup Lee.International Conference on Intelligent Human–Computer Interaction (IHCI 2026), 2026 · accepted
    BibTeX
    @inproceedings{ashraf2026motivational,
  title     = {{Can Motivational Interviewing Guided Dialogue Improve Empathy in Human–Robot Interaction? A Controlled Study with Baseline Comparison}},
  author    = {Ashraf, Raiyan and Liu, Yanni and Kolesnichenko, Kyrylo and Ganji, Sruthi and Kim, Jong-Hoon and Lee, Ju Yup},
  booktitle = {International Conference on Intelligent Human–Computer Interaction (IHCI 2026)},
  year      = {2026}
}