PaperlayerView exact revision on arXiv

Source-linked AI summary

Scoping the Landscape of Large Language Model-based Affective Personal Informatics: Review, Reflection, Future Direction

Minjie Liu, Lihuan Li, Xuran Li, Flora Salim, Benjamin Tag

arXiv:2610.20084v1cs.HC

TL;DR

Affective PI uses personal data to support emotional self-awareness and well-being, but no dedicated synthesis had mapped how LLMs are integrated into these tools. This scoping review analyzed 45 papers and found recurring challenges in processing pipelines, interaction, and participant recruitment, informing future design directions.

  • Problem

    Existing surveys cover related emotion tasks, empathy, or clinical diagnosis, but not how LLMs are designed and integrated into affective PI tools.

    Source: §2.2 · p. 4

  • Method

    The authors conducted a scoping review of 45 papers and synthesized macro trends, system taxonomies, and research methodologies.

    Source: §1 · p. 2, affective PI? · p. 2

  • Results

    The synthesis found recurring challenges in pipeline construction, interaction modality, and participant recruitment in current LLM-based affective PI tools.

    Source: affective PI? · p. 2, §6 · p. 23

  • Takeaways & Limitations

    The review offers actionable recommendations for developing more robust, low-burden, and inclusive affective PI systems.

    Source: §6 · p. 23

    The review excludes commercial closed-source applications, emerging preprints, and non-English systems, and captures a time-bounded view of rapidly changing LLMs.

    Source: §5.3 · p. 23

Abstract

from arXiv · show

Affective personal informatics (PI) involves capturing, tracking, analyzing, and reflecting on personal data to support emotional self-awareness and mental well-being. Recent advances in large language models (LLMs) are reshaping affective PI, yet a comprehensive survey of LLM-based affective PI systems is lacking. To bridge this gap, we systematically analyzed 45 papers published between January 2020 and June 2026 across four major computing and health databases, examining their macro trends, system taxonomies, and research methodologies. Our synthesis reveals an overreliance on multi-stage, text-centric pipelines, a predominance of active, screen-bound interactions, and a severe demographic skew toward young, university-campus-based cohorts in empirical evaluations. Based on these findings, we critically reflect on prevailing architectural and methodological shortcomings and outline actionable future directions for building end-to-end multimodal, low-burden, and inclusive affective PI systems.

1 Introduction

Affective PI supports emotional self-awareness and well-being through personal-data tracking, reflection, and action, while LLMs enable contextual interpretation of fragmented self-reports. This review addresses the lack of a dedicated synthesis and identifies recurring design and evaluation limitations alongside future directions.

1 Introduction

  • Affective PI has expanded from recording physiological signals and self-reported moods toward diverse interaction, contextualized reflection, and personalized intervention.It aims to support emotional self-awareness and mental well-being through personal data and reflection.Source: §1 · p. 2
  • LLMs can interpret language in context and generate conversational responses from fragmented, subjective self-reports that traditional rule-based methods or classifiers struggle to process.These capabilities create opportunities for more flexible, contextualized, and personalized reflection and support.Source: §1 · p. 2
  • Existing surveys address emotion-related computational tasks, LLM empathy, or clinical and psychiatric diagnosis, but do not synthesize how LLMs are integrated into affective PI tools.The review examines macro trends, system taxonomies, and research methodologies across 45 papers.Source: §1 · p. 2, affective PI? · p. 2
  • The review identifies multi-stage text-centric pipelines, active screen-bound interactions, and evaluations skewed toward young, university-affiliated cohorts as recurring limitations.It outlines end-to-end multimodal modeling, low-burden interaction, and inclusive design as future directions.Source: affective PI? · pp. 2–3

2 Background and Related Work

Affective PI has broadened from retrospective tracking toward diverse, contextualized, and personalized support for everyday emotional well-being. Unlike adjacent LLM work on affective tasks, empathy, or clinical diagnosis, LLM-based affective PI has not yet received a focused synthesis.

2.1 Affective PI

  • Affective PI research increasingly combines diverse interactions, contextual data, and personalization with emotional tracking and reflection.Examples include conversational agents, tangible artifacts, ambient lighting, situational context, and tailored intervention timing or dialogue.Source: §2.1 · p. 3
  • Conventional classifiers can produce coarse, uninterpretable labels, and rule-based dialogue systems are constrained by predefined rules and limited flexibility with nuanced feelings.LLMs offer capabilities for integrating emotional data and generating personalized reflections and responses.Source: §2.1 · p. 3

2.2 LLMs in Affective Computing

  • LLMs are explored for affective algorithms, computational empathy, and clinical applications, while affective PI focuses on everyday, longitudinal emotional well-being.The literature lacks a focused synthesis of how LLMs are used in affective PI tools.Source: §2.2 · p. 4

3 Methodology

The review followed PRISMA-guided searching, screening, and analysis to assemble and synthesize empirical studies of LLM-based affective PI tools. Its process covered four databases, eligibility criteria for interactive tools and user evaluation, and extraction across three research dimensions.

3 Methodology

  • The review followed PRISMA guidelines and used Covidence to manage the review papers.The corpus comprised 45 papers after title/abstract and full-text screening.Source: §3 · p. 4, §3.2 · p. 6

3.1 Identification

  • The search covered ACM Digital Library, IEEE Xplore, Scopus, and PubMed for literature published from January 2020 through June 2026.The search combined LLM-related and affective-PI-related terms; 2,155 records were found and 248 duplicates removed, leaving 1,907 for screening.Source: §3.1 · pp. 4–5

3.2 Screening and Eligibility

  • Studies had to be empirical interactive-tool papers with LLMs as a core component, user studies, and an affective-PI purpose; 48 full-text papers were excluded, leaving 45.Three independent reviewers screened records, resolved disagreements, and audited initial exclusions; final eligibility disagreements were resolved by consensus.Source: §3.2 · p. 6

3.3 Data Analysis

  • The extraction protocol covered macro trends, system taxonomies, and research methodologies, with independent extraction followed by discussion to resolve discrepancies.The analysis categorized and cross-compared extracted variables to map architectures, theoretical foundations, and evaluation paradigms.Source: §3.3 · p. 6

4 Results

The reviewed literature expanded sharply after 2024, while systems remained concentrated in GPT models and text inputs. Research was regionally concentrated, and multimodal inputs were often transformed into textual representations before LLM processing.

4.1.1 Temporal Trend in the Context of LLM Advancements.

  • 13 papers appeared in January–June 2026 after 25 in 2025, following a rise from one paper in 2020–2022 to one in 2023 and five in 2024.The review describes a clear trajectory from early exploration (2020–2024) to a surge in 2025 and 2026.Source: §4.1.1 · p. 7, §4.1 · p. 7

4.1.2 Distribution of Utilized LLMs.

  • The GPT family powered 64.4% of reviewed systems (n=29), while five studies (11.1%) used multi-LLM architectures or cross-model comparisons.The corpus also included LLaMA (8.9%, n=4), Gemini (6.7%, n=3), and other alternatives (8.9%, n=4).Source: Table 1 · p. 8, §4.1.2 · p. 8

4.2 Geographical and Disciplinary Trend

  • East Asia was the dominant contributor, accounting for nearly half of the corpus (46.7%, n=21), with China (n=11), the Republic of Korea (n=6), and Japan (n=4) driving this concentration.Source: §4.2 · p. 8

4.3.1 System Input Modalities and Data Processing.

  • Across the reviewed systems, pure textual data was the most common input configuration (60.0%, n=27), followed by visual and textual data and speech dialogue (13.3%, n=6 each).Other configurations included physiological signals (6.7%, n=3) and passive behavioral data (4.4%, n=2).Source: §4.3.1 · p. 11
  • Non-text inputs were commonly converted into text for LLM processing, including transcribed speech, textual summaries of biometric or behavioral data, and captions or emotion labels for visual data.AutoJournaling instead fed raw vision tokens to multimodal LLMs; RageSense combined acoustic sensing history with gameplay video frames.Source: §4.3.1 · p. 11

Studies

The examples describe varied ways systems turn personal records and reflections into visual or structured outputs, while also supporting recall and self-reflection.

  • MoodMonster extracts emotion keywords from text to synthesize monster images.Source: Studies · p. 13
  • CRAFTJAI transforms diary entries into gamified cartoon mood characters, while Personal Data Probe translates personal informatics logs into generative images.Source: Studies · p. 13
  • Reverba guides memory recall through dialogue and generates visual memory art; RemiGraph and MemoryLib structure lifelogs or egocentric traces into memory graphs.Source: Studies · p. 13
  • MindScape uses passive sensing, while AutoJournaling summarizes continuous smartphone screenshots to infer feelings.Source: Studies · p. 13
  • Other examples include reflective dialogue for emotion concepts, body-image cognitive restructuring, and adaptive prompts for future-oriented goals.Source: Studies · p. 13

4.3.2 Tool Categories and Functional Paradigms.

The review spans tool categories and research practices, from intervention and journaling systems to varied user-needs, design, and evaluation methods. Most studies relied on expert-led design, mixed-methods evaluation, within-subjects designs, and young, non-clinical participants.

4.3.2 Tool Categories and Functional Paradigms.

  • Journaling support systems comprised 28.9% (n=13), while conversational chatbots accounted for 20.0% (n=9); embodied AI represented 13.3% (n=6), and VR environments 6.7% (n=3).Source: §4.3.2 · pp. 14–15

4.4.1 Identifying Users’ Needs and Requirements.

  • Empirical formative research appeared in 42.2% of studies (n=19), while 51.1% (n=23) derived requirements from prior literature, theories, or problem-driven rationales without pre-deployment user studies.Interviews were the most common empirical method (17.8%, n=8).Source: §4.4.1 · p. 16

4.4.2 Designing Methodologies.

  • Expert- and developer-led design was used by 68.9% of the corpus (n=31), compared with user-centered design in 15.6% (n=7).Three studies used autobiographical design, and one used co-design.Source: §4.4.2 · pp. 16–17

4.4.3 Evaluation Methods and Study Design.

  • Mixed-methods evaluation predominated (68.9%, n=31), combining quantitative measures with qualitative methods; 24.4% used purely quantitative and 6.7% purely qualitative evaluations.Source: §4.4.3 · p. 17
  • Within-subjects designs were used in 77.8% of studies (n=35), and 62.2% (n=28) evaluated systems for at least one week.Short-term evaluations under one week accounted for 37.8% (n=17).Source: §4.4.3 · p. 17, §2.0 · p. 18

4.4.4 Participant Demographics.

  • Participants were generally young adults recruited from universities, with 37.8% of studies focused on ages 18–25 and 35.6% on ages 18–40.Only one study each focused on adolescents and adults over 60; 22.2% did not report age demographics.Source: §4.4.4 · p. 19
  • Most studies evaluated general non-clinical participants (71.1%, n=32), while 6.7% (n=3) evaluated sub-clinical or at-risk populations.Another 17.8% (n=8) studied non-clinical sub-populations experiencing specific emotional burdens or situational stressors.Source: §4.4.4 · p. 19

5 Discussion

The discussion identifies weaknesses in multimodal processing, interaction burden, and participant sampling, then proposes directions for system design and evaluation. It also bounds the review’s scope and evidential certainty.

5.1 Critical Reflection

  • 60.0% (n=27) of systems used pure text; non-text inputs in many others were textualized, losing affective detail and compounding errors.Speech transcription can discard vocal cues, while static labels can erase physiological trajectories; upstream errors can then mislead downstream reasoning and recommendations.Source: §5.1 · pp. 19–20
  • Active, screen-bound tools can burden users during distress, while embodied and VR systems together represent only 19.0% (n=8) of the corpus.Reviewed systems commonly involve journaling or multi-turn interaction; the discussion presents embodied and spatial approaches as possibilities for reducing active operational demands.Source: §5.1 · pp. 20–21
  • Young cohorts aged 18–25 (37.8%, n=17) and 18–40 (35.6%, n=16) dominate reported age samples; adolescents and adults over 60 each account for 2.2% (n=1).Recruitment was overwhelmingly campus-based, and the review cautions that positive results in digitally literate cohorts may not generalize to broader populations.Source: §5.1 · p. 21

5.2 Future Directions

  • Future systems should process visual, audio, and time-series inputs natively, while connecting affective interaction research with domain-specific sensor-model development.The discussion points to cross-modal alignment that maps sensor patches into LLM-compatible representations, enabling continuous sensor dynamics to accompany dialogue.Source: §5.2 · pp. 21–22
  • Low-burden interaction should balance reduced workload with active user expression, using micro-, peripheral, or fatigue- and state-adaptive engagement.Examples include one-tap check-ins, interactions through objects or environments, and adjusting dialogue pacing and verbosity to users’ capacity.Source: §5.2 · p. 22
  • Evaluations should recruit beyond university settings and include underrepresented, vulnerable populations across life stages and varied socioeconomic and occupational backgrounds.The discussion calls for longitudinal in-the-wild studies involving diverse stressors to identify demographic-specific failure modes, including clinical misinterpretations and cultural misalignments.Source: §5.2 · pp. 22–23

5.3 Limitations

  • The review excludes non-English literature, preprints, and commercial closed-source applications, and cannot independently audit claims or compare systems quantitatively.Its search covers peer-reviewed English literature from January 2020 to June 2026 across four digital libraries; the authors attribute audit and comparison limits to scarce open artifacts, datasets, and standardized protocols.Source: §5.3 · p. 23

6 Conclusion

The review charts LLM-based affective PI research across system design and research methods. It identifies challenges in pipelines, interactions, and recruitment, and offers recommendations for future work.

6 Conclusion

  • The review of 45 articles maps trends, system taxonomies, and lifecycle methods, identifying design and recruitment challenges and offering future research recommendations.Source: §6 · p. 23

Related papers

Loading 2610.20084v1…