Bloom: Designing for LLM-Augmented Behavior Change Interactions

Paper Title

Bloom: Designing for LLM-Augmented Behavior Change Interactions

Publication Info

• Topic area: Health behavior change through LLM-integrated mobile applications
• Keywords: Large language models, behavior change, physical activity, health coaching, personalization, motivational interviewing, mobile health, ambient displays, user experience, wearable data

Background and Problem

• Problem / challenge: Existing research on LLMs for behavioral health focuses narrowly on text-only interactions, overlooking the potential to augment established multimodal behavior change strategies. Real-world evaluations of LLM-based systems remain limited, with few studies comparing LLM systems to pre-LLM approaches.
• Significance: Physical inactivity is a global health concern, with scalable, cost-effective interventions urgently needed to reduce barriers to regular physical activity. LLMs offer novel opportunities for personalized and empathetic health coaching, potentially addressing these challenges.
• Motivation and related work: Decades of HCI research have established effective behavior change interactions such as goal setting, self-tracking, ambient feedback, and nudges. Pre-LLM systems often relied on rule-based personalization, which lacked flexibility. Recent advances in LLMs enable dynamic, open-ended conversations and richer personalization using qualitative context, but their integration with multimodal interventions remains underexplored.

Solution

• Proposed approach: Bloom, a mobile application for physical activity promotion, integrates an LLM-based health coaching chatbot (Beebo) with established behavior change interactions, including goal setting, activity tracking, ambient displays, push notifications, and data visualizations.
• Novelty:
  1. Development of Bloom, which combines LLM-driven coaching with multimodal behavior change interactions.
  2. Creation of a safety benchmark dataset with 600 examples to evaluate LLM coaching safety filters.
  3. Results from a four-week randomized field study comparing Bloom to a no-LLM control condition.
  4. Design insights for LLM-augmented interventions, emphasizing the role of qualitative context and relational cues.
• Procedure and key techniques:
  • Bloom operationalizes the Stanford Active Choices Program, adapting its structure for digital delivery.
  • Beebo uses motivational interviewing strategies to conduct onboarding, check-ins, and at-will chats.
  • The system integrates wearable data, qualitative context, and LLM-generated plans and summaries.
  • Safety filters were developed through redteaming interviews and evaluated using a benchmark dataset.

Results

• Concrete findings:
  • Both conditions doubled the proportion of participants meeting recommended physical activity guidelines (from 36% to 72% overall).
  • Treatment participants reported greater improvements in mindset-related outcomes, including enjoyment (+0.3 points), self-confidence, and agency.
  • Mean physical activity levels increased significantly during the study period (+1,680 steps/day, +87.1 kcal/day, +13.2 min/day, +0.76 km/day).
  • Treatment participants created more varied and balanced plans (1.9 vs. 1.4 activity categories per plan) and spent 5.6x more time in the app.
• Advantage over baselines:
  • Treatment participants showed more stable physical activity persistence over time, while control participants exhibited larger initial gains followed by steeper declines.
  • Treatment participants rated user experience, advice quality, and shared decision-making more highly.
• Experiments / evaluation:
  • Four-week randomized field study with N=54 participants (26 treatment, 28 control).
  • Mixed-methods approach combining wearable data, surveys, interviews, and system interaction logs.
  • Safety filters achieved high recall (≥0.96) and reduced harmful outputs to <0.2% in corrected test sets.
• Limitations and future work:
  • Study duration and sample size were insufficient for robust evaluation of long-term behavior change.
  • Usability challenges arose from repetitive messaging and rigid conversational flows.
  • Recruitment skewed toward higher SES participants, limiting generalizability.
  • Future work should explore advanced LLM designs, longer-term studies, and engagement with low-SES populations.

Summary

Bloom integrates an LLM-based health coaching chatbot with multimodal behavior change strategies to promote physical activity. A four-week randomized field study demonstrated that Bloom fostered positive mindsets, personalized plans, and sustained engagement, doubling the proportion of participants meeting physical activity guidelines. While no short-term advantage in physical activity levels was observed for the LLM condition, psychological and motivational outcomes suggest potential for longer-term benefits. Bloom highlights the opportunities for LLMs to leverage qualitative context and relational cues to empower users and sustain behavior change, with implications for future designs and evaluations of LLM-augmented health interventions.