← CHI 2026 map

CHI '26 · Honorable mention · full-paper review · confidence medium-high

MindfulAgents: Personalizing Mindfulness Meditation via an Expert-Aligned Multi-Agent System

Mengyuan Wu , Zhihan Jiang , Yuang Fan , Richard Feng , Sahiti Dharmavaram , Mathew Polowitz , Shawn Fallon , Bashima Islam , Lizbeth Benson , Irene Tung , David Creswell , Xuhai "Orson" Xu

DOI PDF Program page Page JSON

MindfulAgents is a credible CHI systems paper: the novelty is in the orchestration of expert-aligned reflection, safety scaffolding, and real-time script personalization rather than in a new meditation theory. The evaluation is reasonably strong for a CHI paper, but the authors themselves acknowledge latency and limited ablation depth, so the contribution is best read as a promising system demonstration rather than a settled clinical advance.

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.

Review Summary

MindfulAgents is strongest as a CHI system contribution that reframes how LLMs should be used in sensitive wellbeing settings. Rather than letting the model freely generate meditation content, the paper builds an expert-aligned multi-agent pipeline that grounds generation in a mindfulness framework, prompts reflection on emotional state and mindfulness skills, and adapts scripts in real time. That design choice is the paper’s main intellectual move: it is not just personalization, but personalization constrained by expert structure and user-state feedback. The evaluation is also more substantial than a simple prototype demo. The abstract reports a formative lab study with N=13 showing significant gains in in-session engagement, self-awareness, and momentary stress, followed by a four-week deployment with N=62 showing improved long-term engagement and mindfulness. That combination gives the paper decent empirical weight for a CHI audience. At the same time, the limitations are important and appropriately stated: the deployment phase was too short to assess sustained psychological benefits, LLM latency affected usability, audio delivery was not personalized even when content was, and the two-arm design limited intermediate ablations. So the paper’s claims are well supported for engagement-oriented personalization in digital mindfulness, but not for broad clinical efficacy or for isolating which component of the multi-agent design matters most. In field terms, this is a solid example of expert-aligned generative interaction design for mental health, with the main value in showing a plausible architecture and early evidence that it can outperform static meditation delivery on user engagement and perceived mindfulness-related outcomes.

What Changed

Canon before

Prior mindfulness apps and LLM-based coaches typically rely on fixed or pre-authored meditation content, with personalization often limited to coarse routing or manual effort rather than real-time script-level adaptation.

Departure from common sense

The paper’s core move is to avoid treating an LLM as a direct meditation-content generator and instead place it inside an expert-aligned multi-agent pipeline with reflection and safety scaffolding. That is a non-obvious design choice because it prioritizes controllability and user-state awareness over raw generative flexibility.

Actual novelty

The novelty is a specific multi-agent architecture for mindfulness meditation that combines expert-established framework grounding, user reflection on emotional state and mindfulness skills, and real-time personalization of guided scripts. The paper explicitly positions this as going beyond routing users to pre-authored meditations toward script-level personalization.

Evidence

The paper reports both a formative lab study and a four-week deployment study. The abstract states that the formative study had N=13 and found significant improvements in in-session engagement, self-awareness, and momentary stress, while the deployment study had N=62 and found increases in long-term engagement and mindfulness. The limitations section also notes latency, non-personalized audio delivery, and a two-arm comparison that prevents intermediate ablations.

Limits

Method limits

The evidence is primarily a two-study evaluation centered on engagement, self-awareness, stress, and mindfulness outcomes. The paper itself notes that the deployment study used a two-arm comparison, limiting assessment of intermediate variants, and that the free practice phase was too brief to measure sustained psychological benefits.

Deployment limits

Real-world use is constrained by LLM processing latency and by the fact that, although content was personalized, audio delivery was not. These constraints matter for responsiveness and for the end-to-end experience of a meditation intervention deployed outside the lab.

Boundary conditions

The claims are best read as applying to digital mindfulness meditation interventions where expert-aligned personalization and reflection can be integrated into the interaction flow. The evidence does not establish broad psychological efficacy across longer horizons or across alternative personalization architectures.

Manual check: No additional manual check is strictly required for schema completion, but a human reviewer may want to verify the exact limitation wording and the full-paper context around the multi-agent architecture and deployment design.

Position in field

This sits at the intersection of HCI, digital mental health, and LLM-mediated personalization. Its contribution is less a new meditation theory than a system-level demonstration that expert-aligned multi-agent orchestration can support more responsive and engaging mindfulness experiences than static or merely routed content.

Abstract

Mindfulness meditation is a widely accessible and evidence-based method for supporting mental health. Despite the proliferation of mindfulness meditation apps, sustaining user engagement remains a persistent challenge. Personalizing the meditation experience is a promising strategy to improve engagement, but it often requires costly and unscalable manual effort. We present MindfulAgents, a multi-agent system powered by large language models that: (1) generates guided meditation scripts based on an expert-established mindfulness framework, (2) encourages users' reflection on emotional states and mindfulness skills, and (3) enables real-time personalization of the mindfulness meditation experience for each user. In a formative lab study (N=13), MindfulAgents significantly improved in-session engagement (p = 0.011) and self-awareness (p = 0.014), as well as reduced momentary stress (p = 0.020). Furthermore, a four-week deployment study (N=62) demonstrated a notable increase (p = 0.002) in long-term engagement and level of mindfulness (p = 0.023). Participants reported that MindfulAgents offered more relevant meditation sessions personalized to individual needs in various contexts, supporting sustained practice. Our findings highlight the potential of LLM-driven personalization for enhancing user engagement in digital mindfulness meditation interventions.