Evoke-Sim: A Task-Aware LLM Client Simulation for Evaluating Motivational Interviewing Counsellors
A new framework, Evoke-Sim, has been introduced for evaluating Motivational Interviewing (MI) counsellors using Large Language Models (LLMs). The framework, detailed in a paper on arXiv (2608.07499), addresses a gap in prior simulated client models by aligning with the specific tasks of MI therapy, particularly the 'evoking' task. Evoke-Sim is designed for smoking cessation counselling and employs structured client profiles, a three-stage conversation flow, and a reveal policy that controls which client information is disclosed at each stage. The framework aims to better differentiate levels of MI quality compared to existing profile-grounded simulated clients. The paper, announced as a cross-type submission, presents Evoke-Sim as a task-aware, multi-stage client simulation framework. The development and benchmarking of LLM-based MI counsellors increasingly rely on simulated clients, but prior work has not aligned with the fundamental tasks of MI. The evoking task involves the counsellor eliciting the client's ambivalence and strengthening their motivation for change. Evoke-Sim is specifically designed for this task. The framework's effectiveness was demonstrated by its ability to differentiate MI quality levels better than existing simulated clients. The paper is available at https://arxiv.org/abs/2608.07499.
Key facts
- Evoke-Sim is a task-aware, multi-stage LLM-based client simulation framework for evaluating MI counsellors.
- It is designed specifically for the evoking task in Motivational Interviewing.
- The framework is applied to smoking cessation counselling.
- Evoke-Sim uses structured client profiles, a three-stage conversation flow, and a reveal policy.
- It outperforms existing profile-grounded simulated clients in differentiating levels of MI quality.
- The paper is available on arXiv with identifier 2608.07499.
- The announcement type is 'cross'.
- The development of LLM-based MI counsellors relies on simulated clients.
Entities
Institutions
- arXiv