Large Language Model–Based Agents for Physical Activity and Cognitive Training: Scoping Review

TL;DR

This scoping review maps the use of large language model-based agents for physical activity and cognitive training, highlighting early promise and key challenges in their design and evaluation.

Contribution

The first systematic mapping of LLM-based agents in physical activity and cognitive training domains, identifying design practices and methodological gaps.

Findings

LLM-based agents are primarily used in coaching roles for physical activity and companion roles for cognitive training.

Proprietary models like GPT-3.5 and GPT-4 dominate, with limited use of open-weight models and inconsistent prompt documentation.

Outcomes focus on user perception rather than behavioral changes, indicating a need for more rigorous evaluation methods.

Abstract

Large language model (LLM)–based conversational agents have been increasingly used in digital health interventions. However, their specific application to physical activity (PA) and cognitive training—two critical well-being domains—has not been systematically mapped. In fact, these domains share an important need for personalized, adaptive support and conversational engagement, making them relevant targets for examining how LLM-based agents are currently conceptualized and deployed.

This scoping review aimed to map the extent, characteristics, and design practices of LLM-based conversational agents supporting PA or cognitive training, specifically analyzing their application contexts, social roles, and technological features.

Following PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines, we searched Web of Science, Scopus, PubMed, ACM Digital Library, and IEEE Xplore for studies published between January 2018 and December 2024. We included eligible studies that described LLM-based conversational agents designed for PA or cognitive training. Two reviewers independently screened records and extracted data. Descriptive synthesis and framework analysis were used to characterize intervention domains, agent roles, prompting strategies, model types, and reported outcomes.

Of 357 records screened, 10 studies met eligibility criteria (7 on PA and 3 on cognitive training). Applications predominantly involved coaching roles for PA and companion or scaffolding roles in cognitive domains. The agent landscape was dominated by proprietary LLMs (GPT-3.5, GPT-4, and Bard), with limited use of open-weight models. Prompt engineering emerged as a central yet inconsistently documented design mechanism. Reported outcomes mainly focused on perceived usefulness, engagement, or content quality, with few quantitative behavioral outcomes.

LLM-based conversational agents have demonstrated early promise for supporting PA and emerging approaches to cognitive training, yet the current evidence remains exploratory and methodologically limited. Key challenges persist, including inconsistent reporting of prompts, reliance on proprietary models with limited reproducibility, and a lack of standardized outcome measures. More rigorous and transparently documented evaluations of these tools are required to strengthen the evidence base and guide future development.

Figures