PaperlayerView exact revision on arXiv

Source-linked AI summary

From Momentary Emotion Inference to Sustained Emotion Support: Evaluating a Companion Agent in a Longitudinal Study

Kexin Quan, Zijian Ding, Jiaye Yong, Qinshi Zhang, Dong Wang, Jessie Chin

arXiv:2609.16344v2cs.HCcs.AI

TL;DR

How generative agents can sustain emotional support over time remains unclear, despite demonstrated momentary emotional-support capabilities. PAIR combines appraisal-guided emotion inference, regulation guidance, and cross-session memory, evaluated with 19 participants over 14 days using self-reports, logs, and interviews. Estimates aligned more closely with self-reported valence and dominance than arousal, and guided conversations were followed by higher valence and state-dependent arousal changes.

  • Problem

    How generative agents can sustain emotional support over time remains unclear, despite demonstrated momentary emotional-support capabilities.

    Source: §1 · p. 2

  • Method

    PAIR combines appraisal-guided emotion inference, regulation guidance, and cross-session memory, evaluated with 19 participants over 14 days using self-reports, logs, and interviews.

    Source: §1 · p. 2

  • Results

    Estimates aligned more closely with self-reported valence and dominance than arousal, and guided conversations were followed by higher valence and state-dependent arousal changes.

    Source: §1 · p. 3, §7.1 · p. 21

  • Takeaways & Limitations

    Remembered experiences and retained corrections relate to cross-session personalization, while forgotten corrections and repetitive advice reveal difficulties adapting support to changing circumstances.

    Source: §1 · p. 3

    The 14-day study included 19 adults aged 18–33, generally familiar with AI, and its fixed check-in schedule constrained naturalistic use.

    Source: §7.2.1 · p. 21

Abstract

from arXiv · show

Sustained emotional support is a long-horizon interaction task closely tied to human well-being. Recent research demonstrates generative agents' capacity for momentary emotional support, yet how these capabilities sustain support over time remains unclear. To examine this challenge, we deployed PAIR, a theory-based emotion-regulation companion, with 19 participants for 14 days. Across 1,093 sessions, we paired emotion estimates with self-reports before and after guidance and analyzed logs and interviews. Estimates corresponded more closely to self-reported valence and dominance than arousal. Guided conversations were followed by higher valence and state-dependent arousal changes. Participants felt understood through contextual exploration and emotional acknowledgment, acting on guidance suited to their needs and constraints. Perceived helpfulness of guided conversation significantly increased over time. Our findings link memory updates and retained corrections to cross-session personalization, informing future emotional support tools that adapt to evolving needs, learn from prior outcomes, and preserve user control over memory.

1 INTRODUCTION

The paper asks how generative agents can turn momentary emotion understanding into support that adapts as users’ circumstances and needs change. It introduces PAIR and evaluates its inference, guidance, and cross-session continuity in a 14-day deployment.

1 INTRODUCTION

  • Sustained support remains an open challenge because earlier deployments report repetitive questioning and unmet expectations of recall, while needs and circumstances change across events.The paper asks how generative agents can sustain emotional support over time and examines emotion understanding, regulation guidance, and cross-session continuity.Source: §1 · p. 2
  • Estimates aligned more closely with self-reported valence and dominance than arousal; guided conversations were followed by state-dependent emotional changes and guidance uptake shaped by practical constraints.Participants reported higher valence, lower arousal in anxious sessions, and higher arousal in dejected sessions.Source: §1 · p. 3
  • PAIR combines appraisal-based emotion inference, state-conditioned regulation strategies, and memory of recurring concerns, user corrections, and guidance outcomes.It routes event narratives to reasoning scaffolds and selects strategies using emotional intensity, controllability, and event timing.Source: §1 · p. 3
  • The study evaluates PAIR over 14 days with 19 participants and 1,093 sessions, pairing emotion estimates with self-reports and analyzing feedback, logs, and interviews.The analysis examines perceived understanding, guidance uptake, and repeated use.Source: §1 · p. 3
  • The paper proposes design implications for aligning dialogue with evolving goals and adapting guidance to prior attempts and outcomes.It highlights inspectable, revisable memory and user control over retention and reuse as supports for sustained personalization.Source: §1 · p. 3

2 RELATED WORK

Prior work studies emotional support through disclosure, reflective dialogue, coping assistance, and companionship, but the evidence is dominated by single-session evaluations. Cross-session work highlights memory and evolving relationships as challenges for sustained support.

2.1 Emotional Support Systems and Companion Agents

  • Support agents combine disclosure, reflection, coping assistance, and companionship with personalized or expressive interaction; selected deployments span journaling, coaching, and adaptation across encounters.PAIR integrates emotional understanding, coping guidance, and expressive interaction within this landscape.Source: §2.1 · pp. 3–4

2.2 Single-Session Capabilities and Limits in AI Emotional Support

  • Single-session studies find promising emotion inference and supportive responses, but cannot establish whether support persists as user trajectories and circumstances evolve.Reviews also report short-term symptom effects without significant long-term effects or overall well-being improvement.Source: §2.2 · p. 4

2.3 Cross-Session Continuity in AI Support Systems

  • Cross-session systems use memory and proactive contact to build continuity, while longitudinal studies document changing expectations, disclosures, trust, and reasons for returning.Memory benchmarks identify difficulties retrieving relevant information, reasoning across sessions, updating knowledge, and tracking changing preferences.Source: §2.3 · p. 5

3 TECHNICAL EVALUATION OF AI EMOTIONAL UNDERSTANDING

The technical evaluation compares theory-grounded emotion-inference prompts through human judgments of plausibility and reasoning fidelity. Appraisal scaffolds outperform alternatives, while the two selected prompts have complementary strengths across event structures.

3 TECHNICAL EVALUATION OF AI EMOTIONAL UNDERSTANDING

  • The paper frames theory-grounded prompting as a way to test how structured reasoning affects emotion inference and select prompts for PAIR.Human raters with first-hand or close-person knowledge judged interpretation plausibility and fidelity.Source: §3 · p. 5

3.1.1 Four Theory-based Prompt Designs.

  • The evaluation tested four prompts with incremental reasoning scaffolding, including direct inference, a five-stage cognitive-appraisal chain, and a combined neurocognitive-appraisal scaffold.The prompts shared an output format returning emotion estimates, a guidance message, and predicted post-guidance emotion.Source: §3.1.1 · pp. 5–6

3.1.2 Procedure.

  • Three researchers contributed ten real-life events each, paired with narrator personas, and rated anonymized outputs in a blinded bracket tournament using the same model.Three raters completed rankings for every scenario, yielding 90 scenario-rater rankings and 1,080 dimension-level pairwise judgments.Source: §3.1.2 · p. 6

3.1.3 Measures and Analysis.

  • Raters assessed outcome plausibility, reasoning fidelity, and overall preference; the analysis used majority votes and coded scenarios on four binary event features.Overall preference integrated plausibility and fidelity, and the feature coding was completed before examining prompt preferences.Source: §3.1.3 · pp. 6–7

3.2.1 Appraisal-grounded prompts outperformed the alternatives.

  • Appraisal-grounded prompts ranked above Baseline and Neuro overall; Cog had the best mean rank and most consistent performance, while Combined received the second-most first-place rankings.Both beat Baseline in most direct comparisons, and Cog beat Neuro in most comparisons.Source: §3.2.1 · p. 7

3.2.2 Cog and Combined Scaffolds performed best on different event structures.

  • Cog and Combined split scenario-level preference 15:15, but matching prompts to event structure aligned with the majority preference in 26 of 30 scenarios (87%).Cog was preferred for single-trigger events with one dominant emotion; Combined was preferred for multiple interacting events or emotional components.Source: §3.2.2 · p. 7, §3.2.1 · p. 7

3.2.3 Scaffold complexity should match event structure.

  • Raters’ overall judgments matched reasoning fidelity in all 30 scenarios, compared with 27 for predicted post-guidance emotion.Cog often omitted bodily influences and layered emotions in multi-component events, while Combined introduced unsupported connections for simpler events; raters penalized complexity when integration was unnecessary.Source: §3.2.3 · pp. 7–8, Table 2 · p. 8

4 SYSTEM DESIGN

PAIR is designed to provide momentary emotion understanding and guidance, carry user context across sessions, and sustain engagement. Its implementation combines event-sensitive inference, memory updates, lightweight interactions, and coordinated multimodal feedback.

4.1 Design Goals

  • PAIR routes narrations to an event-matched reasoning scaffold, estimates pleasure-valence, arousal, and dominance, and selects guidance based on intensity, controllability, and event phase.The design also aims to carry user knowledge forward and sustain engagement as needs and situations change.Source: §4.1 · p. 8

4.2 System Implementation

  • The system infers affect and context, selects among four regulation-strategy families, and delivers an actionable plan through an Exploration–Comforting–Action conversation.Afterward, it re-estimates affect using the original event, initial estimate, and conversation transcript without presuming improvement.Source: §4.2 · pp. 9–10, Figure 4 · p. 9

4.2.2 Continuity through Memory and Personalization (DG2).

  • Onboarding persona and calibration establish personal context, while rolling memory summarizes concerns, estimate–self-report differences, corrections, and prior strategy outcomes for later guidance.The memory is included in subsequent appraisal and conversation calls so prior errors, corrections, and outcomes can inform adaptation.Source: §4.2.2 · p. 10

4.2.3 Sustained Engagement through Low-Burden and Multimodal Interaction (DG3).

  • Daily touchpoints include morning check-ins, anytime event logs, and evening reflections; speech or text and coordinated color, pulse, prosody, and avatar cues support interaction.PAIR presents consistent visual and auditory feedback based on the same emotional representation.Source: §4.1 · p. 9, §4.2.3 · p. 10

5 STUDY DESIGN

The study paired PAIR’s emotion estimates with participant self-reports across a 14-day field deployment, then combined quantitative measures and qualitative accounts to examine inference, regulation, and continued use.

5 STUDY DESIGN

  • The 14-day diary-style deployment paired PAIR’s estimates with participants’ self-reports of the same moments.The procedure included researcher-led onboarding and post-study questionnaires and interviews.Source: §5 · p. 10, Figure 5 · p. 11

5.1 Participants

  • The final sample comprised 19 adults aged 18–33 who were not receiving psychological or psychiatric treatment.Five of 24 recruits withdrew during the study.Source: §5.1 · p. 11

5.2.1 Day 0: Onboarding.

  • On Day 0, participants completed a persona interview and calibrated emotion ratings using SAM before a practice log.The interview covered daily life, coping habits, support sources, and current concerns.Source: §5.2.1 · p. 11

5.2.2 Days 1–14: Everyday Companion Use.

  • During Days 1–14, participants narrated events, rated valence, arousal, and dominance before and after guidance, and could correct low-rated interpretations.Estimates were withheld until self-ratings were submitted to avoid anchoring.Source: §5.2.2 · p. 11, Figure 5 · p. 11

5.2.3 Day 14+: Post-Study Session.

  • The Day 14+ session used a 12-item questionnaire and a semi-structured interview to assess understanding, guidance, adoption, and experience.The interview lasted 20–25 minutes.Source: §5.2.3 · p. 12

5.3 Data Collection and Analysis

  • The mixed-methods analysis combined paired estimates and self-reports with logs, feedback, questionnaires, and interviews.Quantitative analyses covered inference accuracy, within-session affect change, engagement, and perceived value.Source: §5.3 · p. 12

5.3.1 RQ1: Emotion Inference and Perceived Understanding.

  • Emotion-inference evaluation compared estimates with paired SAM reports using error, correlation, signed-error, and categorical-agreement measures.Initial valence estimates were also compared with a same-model baseline without appraisal reasoning and a lexicon-based estimator.Source: §5.3.1 · p. 12

5.3.2 RQ2: Regulation Outcomes and Guidance Fit.

  • Regulation analyses tested participant-level pre/post affect changes and examined trajectories by initial state, understanding, and conversation stage.The study also related perceived understanding to valence change and characterized conversation-stage transitions.Source: §5.3.2 · p. 12

5.3.3 RQ3: Continued Use and Personalization.

The RQ3 analyses tracked use and perceived value across two weeks, then examined whether remembered context and corrections shaped personalization.

5.3.3 RQ3: Continued Use and Personalization.

  • RQ3 analyses examined daily engagement, helpfulness, feasibility, willingness to act, and perceived personalization over two weeks.Interviews contextualized trends and explored how earlier interactions and corrections appeared in later responses.Source: §5.3.3 · p. 13

5.3.4 Qualitative Analysis.

  • Two researchers independently coded interviews and rotating-question responses, reconciled differences, and iteratively refined the codebook.Themes followed the research questions, with subthemes developed through discussion.Source: §5.3.4 · p. 13

6 RESULTS

Across the deployment, PAIR’s estimates aligned more closely with self-reported valence and dominance than arousal, while guidance was followed by state-dependent affect changes. Participants described understanding and useful personalization as depending on contextual exploration, fit, and continuity across sessions.

6.2.1 Appraisal-Guided Inference from Daily Narration Yielded Strong Valence and Moderate Dominance Correlations

  • Across all narrations, appraisal-guided inference from everyday narration correlated with self-reported valence at r=.669 (MAE 1.196), outperforming same-model baseline without the appraisal scaffold (r=.640, MAE 1.334) and cross-validated NRC-VAD (r=.318, MAE 1.679).Source: §6.2.1 · p. 14

6.2 RQ1: Text-Based Appraisal Inference Supports Valence and Dominance but Remains Limited for Arousal

  • Arousal estimates showed weak correspondence with self-reports, with mismatches often involving emotional intensity when text lacked expressive cues.Binary arousal agreement was slight (κ= .154), and changing the cut point did not raise it above κ= .21.Source: §6.2.2 · p. 15

6.3.1 State-Dependent Emotional Adjustment Following Guided Conversations Suggested Momentary Regulation.

  • Guided conversations were followed by state-dependent changes: anxious and dejected sessions gained valence, while arousal decreased for anxious and increased for dejected sessions.Low-valence sessions gained 1.344 points; high-valence sessions changed by −0.108 points.Source: §6.3.1 · p. 16, Figure 8 · p. 17

6.3.2 Contextual Exploration and Emotional Acknowledgment Before Advice Supported High Perceived Understanding.

  • Participants described feeling understood through contextual exploration and emotional acknowledgment, although perceived understanding had little association with estimation error.Exploration began 99.9% of sessions, and Comforting transitioned to Action in 72.2% of adjacent stage pairs.Source: §6.3.2 · pp. 16–17

Outcomes.

  • No strategy family showed a clear advantage in immediate outcomes, and participants said suggestion usefulness depended on fit to the situation.Immediate valence-gain differences across the four strategy families were not detectable (permutation p= .897).Source: §6.3.3 · pp. 18–19

Uptake.

  • Participants reported acting on guidance when it fit their available time, resources, and immediate support needs.Of 44 categorized responses, 24 described trying guidance with a positive result, while five described partial or alternative enactment or mixed results.Source: §6.3.4 · p. 19

6.4.1 Familiarity with Response Patterns Guided Narration Practices and Support Expectations.

  • Participants kept using PAIR at a similar frequency while writing less and rating helpfulness more highly over the two weeks.Helpfulness rose from 4.94 in Week 1 to 5.14 in Week 2; initial narration length and per-turn input length declined.Source: §6.4.1 · pp. 19–20

6.4.2 Relevant Recall Supported Personalization, While Outdated Context and Forgotten Corrections Disrupted Con-

  • Personalization depended on relevant context being interpreted and updated across sessions; missing, misplaced, or persistent outdated context and forgotten corrections disrupted continuity.Twelve participants reported forgotten context, and nine described corrections that did not persist or stopped correcting the system.Source: §6.4.1 · pp. 20–21

7 DISCUSSION

PAIR’s fourteen-day deployment connected momentary emotion understanding and regulation with cross-session continuity; results and participant accounts point to support that adapts to changing needs, prior attempts, and user corrections.

7.1 Summary of Results

  • Appraisal-guided estimates aligned more closely with self-reported valence and dominance than arousal; guided conversations were followed by state-dependent changes and participants valued contextual exploration and acknowledgment.Participants adopted guidance suited to immediate needs and practical constraints; no strategy family showed a clear advantage in immediate outcomes under contextual selection.Source: §7.1 · p. 21

7.2.1 Study Limitations.

  • The 14-day deployment included 19 adults aged 18–33 who were generally familiar with AI, and fixed check-ins did not establish voluntary-use frequency.Affect ratings relied on participants’ scale use, post-guidance reports may reflect conversation framing, and strategy comparisons were observational; longer follow-up could test whether momentary effects carry into everyday well-being.Source: §7.2.1 · pp. 21–22

7.2.2 Open Challenges for Sustained Emotional Support.

  • Sustained support must interpret affect from selective disclosure and interaction-shaped accounts, while deciding which prior experiences remain relevant to later recommendations.Text-based estimates agreed less for arousal than valence and dominance; recalling personal information alone does not ensure adapted guidance.Source: §7.2.2 · p. 22

7.3 Toward Sustained Emotional Support

  • Sustained support adds tracking changing preferences, incorporating prior guidance outcomes, and updating personal context to momentary acknowledgment and contextual guidance.The paper derives implications for evolving goals, longitudinal feedback, and user-controlled memory.Source: §7.3 · pp. 22–23

7.3.1 Changing Support Goals Require Flexible Transitions Between Listening and Guidance.

  • Support goals vary across encounters, so systems should elicit current intent and flexibly move between listening and guidance.Advice may be premature even when a strategy fits the inferred emotion; users should be able to revise preferences, decline advice, and end interactions when needs are met.Source: §7.3.1 · p. 23

7.3.2 Prior Attempts and Outcomes Should Inform the Next Step in Regulation Guidance.

  • Later recommendations should account for prior attempts, reported outcomes, and barriers, distinguishing immediate reassurance from subsequent usefulness.A support history and brief follow-up could help repeat, modify, or replace suggestions based on practical outcomes and retrospective judgments.Source: §7.3.2 · pp. 23–24

7.3.3 Inspectable and Revisable Memory Should Give Users Control over Subsequent Personalization.

  • Personalization depends on memories remaining accurate and acceptable to retain; users should be able to inspect, revise, selectively reuse, and delete stored context.The proposed implications also call for distinguishing user statements from system interpretations and checking permissions before retrieving memories.Source: §7.3.3 · p. 24

A RUNTIME ORCHESTRATION

  • PAIR routes event narrations through an appraisal scaffold, derives a strategy and delivery parameters, guides a conversation, re-estimates emotion, and updates context nightly.The conclusion frames sustained support as repeatedly interpreting experiences, learning from prior interactions, and adjusting responses over time.Source: §7.4 · pp. 24–25, Figure 12 · p. 30

B PROMPT EVALUATION DETAILS

The appendix reports formative prompt-ranking results and lists the study’s onboarding, questionnaire, and interview materials. The prompt-selection rule matched majority preferences in most formative scenarios.

B PROMPT EVALUATION DETAILS

  • Cog placed in the top two in 72% of rankings and won 84% of direct matchups against Baseline in the 90 scenario-rater rankings.Combined earned 34 first-place finishes versus Cog’s 39 and won 78% of its matchups against Baseline; Table 4 gives the ranking distribution.Source: B PROMPT EVALUATION DETAILS · p. 30, Table 4 · p. 31
  • 26 of 30 formative scenarios (87%) matched majority preference when Cog was selected for single-component events and Combined for multi-component events.Preferences split 15:15 overall; Cog was preferred in 14 of 17 single-component scenarios and Combined in 12 of 13 multi-component scenarios.Source: B PROMPT EVALUATION DETAILS · pp. 30–31

C PERSONA INTERVIEW QUESTIONS

  • The onboarding persona interview used seven conversational questions about self-description, daily life, coping, support, concerns, good days, and mood contagion.Participants answered by voice with editable transcripts and optional hints.Source: C PERSONA INTERVIEW QUESTIONS · p. 31

D POST-STUDY QUESTIONNAIRE

Section 1: Perceived Emotional Understanding (RQ1)

  • The perceived-understanding items asked about momentary accuracy, anticipated feelings, emotional nuance, personalization, and getting to know participants over two weeks.Q1 and Q3 were adapted from PETS; Q2, Q4, and Q5 were developed for this study.Source: Section 1: Perceived Emotional… · p. 31

Section 2: Perceived Emotional Guidance (RQ2)

  • The perceived-guidance items covered emotional understanding, relevance, concrete ideas, perspective-taking, and coping support.Q7, Q9, and Q10 were adapted or derived from prior measures, while Q6 and Q8 were developed for this study.Source: Section 2: Perceived Emotional… · p. 32

Section 3: Technology Acceptance

E SEMI-STRUCTURED INTERVIEW PROTOCOL

  • The semi-structured interview asked about overall experience, surprises, emotion interpretations, suggestion adoption, emotional awareness, usefulness, frustrations, and other comments.The protocol included probes about daily check-in burden and desired changes.Source: E SEMI-STRUCTURED INTERVIEW PROTOCOL · p. 32

G DATA COLLECTED

The appendix identifies the onboarding measures and session-level data summaries presented in Tables 5 and 6.

G DATA COLLECTED

F PARTICIPANT CHARACTERISTICS

  • Table 5 reports demographic, AI-background, and baseline expressive-suppression measures for 19 participants.AI familiarity used a 5-point scale, and the three expressive-suppression items used a 7-point scale whose composite was their mean.Source: Table 5 · p. 33
  • The expressive-suppression composite had Cronbach’s α = .74.Source: Table 5 · p. 33

H WITHIN-SESSION CHANGES IN SELF-REPORTED AFFECT

The section describes how marginal self-reported affect changes were tested and how the regulation-field visualization summarizes session trajectories.

H WITHIN-SESSION CHANGES IN SELF-REPORTED AFFECT

  • Figure 13 pools sessions across entry-state quadrants and tests the 19 participants’ mean changes against zero using exact two-sided Wilcoxon signed-rank tests.The analysis used 10,000 whole-participant bootstrap resamples and Holm correction across dimensions.Source: H WITHIN-SESSION CHANGES IN… · p. 34, Figure 13 · p. 34
  • Figure 8 vectors run from each cluster’s mean pre-guidance coordinate to its mean post-guidance state within an entry-state quadrant.Three clusters were used for the smallest quadrant and four for each other quadrant to retain local variation without plotting every trajectory.Source: H WITHIN-SESSION CHANGES IN… · p. 34
  • All three marginal SAM-dimension changes remained significant after Holm correction (pHolm < .001).Figure 13 shows session-level distributions, participant means, and session-weighted means with 95% participant-bootstrap confidence intervals; Δ marks mean within-session change.Source: Figure 13 · p. 34, H WITHIN-SESSION CHANGES IN… · p. 34

I IMMEDIATE FEEDBACK BY REGULATION STRATEGY

This section describes immediate feedback ratings grouped by selected regulation strategy and the mixed-effects tests used to compare them.

I IMMEDIATE FEEDBACK BY REGULATION STRATEGY

  • Strategy was tested as a fixed effect in linear mixed-effects models with participant and participant-day random intercepts, using likelihood-ratio tests and Benjamini–Hochberg correction across three tests.Source: I IMMEDIATE FEEDBACK BY REGULATION… · p. 34
  • Table 7 reports unadjusted mean immediate feedback ratings on 7-point scales, grouped by selected regulation strategy.The table caption identifies the strategy groups and their session counts.Source: Table 7 · p. 34
  • The strategy groups were situation modification (n=317), cognitive change (n=246), response modulation (n=256), and attentional deployment (n=274).Source: Table 7 · p. 34

J SESSION MEASURES OVER THE DEPLOYMENT

This section reports longitudinal models for nine session measures across the 14-day deployment, using both per-day slopes and Week 2 versus Week 1 contrasts.

J SESSION MEASURES OVER THE DEPLOYMENT

Loading 2609.16344v2…