An LLM-based Motivation-Aware Framework For AI Coaching For Behaviour Change
This is a credible CHI contribution because it turns a familiar MI idea into a concrete LLM coaching framework with a clear state-based strategy split and a real mixed-methods evaluation. The novelty is incremental but well-argued, and the paper is appropriately bounded by a single-session physical-activity study.
Axes Lens
Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.
Contribution shape
- Knowledge form
- technical knowledge typical · 50/268
- Novelty type
- framework typical · 59/268
- Abstraction level
- system typical · 61/268
- Generalization target
- design family typical · 38/268
- Validation mode
- mixed methods typical · 136/268
Evidence profile
- Evidence strength
- moderate typical · 105/268
- Claim alignment
- strong typical · 231/268
- Overclaim risk
- medium typical · 210/268
Review Summary
From a CHI perspective, the paper’s value is not that it invents motivational interviewing for coaching, but that it makes a sharper design move: it treats motivational state as the key control variable and uses that to decide whether the agent should emphasize motivational support or planning. That is a sensible departure from generic “supportive chatbot” thinking, and it is grounded in the authors’ claim that prior MI systems overemphasized emotional support while underusing goal-setting behaviors. The contribution is therefore best read as a framework-level advance rather than a new interaction primitive or a new model architecture. The evaluation is also aligned with the claim: the authors report a mixed-methods, single-session study with 140 Prolific participants, comparing against a baseline system and measuring readiness to change plus user perceptions of MI adherence and planning effectiveness. That gives the paper reasonable empirical support for short-term interaction effects. At the same time, the limitations are important and well matched to the evidence: the study is single-session, text-only, and tied to a physical-activity scenario, with the authors explicitly noting that Prolific participants may lack genuine motivation for behavior change and that non-verbal cues are ignored. So the paper’s strongest reading is as a bounded but useful design and evaluation package for LLM-based health coaching, not as proof of durable behavior-change efficacy across contexts. The honorable-mention level fits that balance: clear idea, credible study, but limited generalization and deployment scope.
What Changed
Canon before
Prior MI-based coaching systems in HCI are described as emphasizing emotional support and generic motivational interviewing behaviors, with less attention to goal-setting and planning behaviors as a distinct design axis.
Departure from common sense
A fixed, one-size-fits-all coaching script is not the right abstraction here; the paper argues that the agent should adapt its strategy to the user’s motivational state, shifting from motivational support for ambiguous clients to planning-focused guidance for action-ready clients.
Actual novelty
The paper’s main contribution is a motivation-aware MI framework for LLM coaching that explicitly distinguishes ambiguous from action-ready clients and switches between motivational support and planning-focused strategies, addressing a gap the authors identify in prior MI-agent work.
Evidence
The paper proposes an LLM-based coaching framework and evaluates it in a mixed-methods, single-session physical-activity study with 140 Prolific participants against a baseline system. Reported outcomes include increased readiness to change, stronger perceived MI adherence, and better planning effectiveness, with qualitative evidence that agreement on a change plan relates to intrinsic motivation gains.
“ The coaching agent can provide support for low-motivated users and resorts to planning strategies otherwise. We demonstrate our framework using physical activity promotion as a low-risk test case, though it can be adapted to other behavioural change domains by modifying prompts and adding domain-specific knowledge”
actual novelty · Abstract + Introduction (goal-setting overlooked; proposed motivation-aware planning vs motivation support) · confidence 0.78
“ Participants perceived our agent as more adherent to MI principles and more effective at planning than the baseline. Qualitative analysis revealed that users who agreed with the agent on a change plan experienced the greatest increases in intrinsic motivation”
departure from common sense · Introduction / Design Guidances (MI phases tailored by motivational states) · confidence 0.62
“Multiagent Systems, Richland, SC, 966–974. Digital Library Google Scholar [52] Jeanette M. Olsen. 2014. Health Coaching: A Concept Analysis: Health Coaching Concept Analysis. Nursing Forum 49, 1 (Jan. 2014), 18–29. Crossref Google Scholar [53] Jeanette M. Olsen and Bonnie J. Nesbitt. 2010. Health Coaching to Improve Healthy Lifestyle Behaviors: An Integrative Review. American Journal of Health Promotion 25, 1 (Sept. 2010), e1–e12. Crossref Google Scholar [54] SoHyun Park, Jeewon Choi, Sungwoo Lee, Changhoon Oh, Changdai Kim, Soohyun La, Joonhwan Lee, and Bongwon Suh. 2019. Designing a Chatbot for a Brief Motivational Interview on Stress Management: Qualitative”
limitation · Limitations section · confidence 0.97
“ We conducted a mixed-methods evaluation using a single-session study design, recruiting 140 participants via the Prolific online research platform”
validation scope · Abstract + Methodology (single-session study; physical activity case study; baseline comparison) · confidence 0.80
Limits
Method limits
Evidence comes from a single-session mixed-methods evaluation with a baseline comparison, so the study supports short-term interaction effects more than durable behavior change or long-term coaching efficacy.
Deployment limits
The system is text-only and depends on textual motivational-state detection; it was tested in a physical-activity coaching scenario with Prolific participants, so deployment beyond that setting and population remains uncertain.
Boundary conditions
The framework is framed around motivational state detection and phase-specific turn limits; the authors also note that participants may not have had genuine motivation for behavior change, which constrains interpretation of the observed effects.
Position in field
This sits at the intersection of MI systems and LLM-based health coaching, extending prior emotionally oriented coaching agents toward a more explicit planning-aware framework and offering a design rationale grounded in motivational state distinctions.