Organizations: School of Information Sciences, University of Illinois Urbana-Champaign, Champaign, Illinois, United States · College of Information, University of Maryland, College Park, Maryland, United States · University of Illinois Urbana-Champaign, Champaign, Illinois, United States · Department of Computer Science, Florida State University, Tallahassee, Florida, United States
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.
Figures & tables
Figure 1. PAIR supports users across everyday emotional moments over 14 days. (1) Users can check in whenever something happens and share the experience through speech or text. (2) The companion draws on what it has learned about the user, including persona, calibration, and prior conversations, to understand how they are feeling and choose a fitting regulation strategy. (3) It then guides the user through the moment in either voice or chat, helping them move toward a more manageable next step with more support. A three-stage illustration of a user's experience with \sysname{} over a 14-day deployment. A user checks in after an emotional event and shares what happened through speech or text. The companion uses information about the user and prior interactions to interpret the emotion and choose a regulation strategy. It then provides guidance through voice or chat, helping the user identify a concrete next step and feel more supported.
Inference
Guidance
Continuity
System
Emo.
State.
Supp.
Cond.
Mem.
Adapt.
Schedule
What the system does
MindfulDiary (CHI’24) ( Kim et al., 2024a )
○
○
28 d, on demand
Conversational journaling with emotion summaries for patients and clinicians.
Socrates 2.0 (JMIR MH’25) ( Held et al., 2025 )
✓
28 d, on demand
Socratic cognitive reappraisal for depression and anxiety.
Therabot (NEJM AI’25) ( Heinz et al., 2025 )
○
○
○
4 wk daily; 4 wk on demand
Generative mental health support using expert-trained dialogue and conversation history.
ComPeer (UIST’24) ( Liu et al., 2024 )
✓
✓
○
7 + 7 d, proactive
Proactive peer support from a persona and conversation history.
VITA (THRI’25) ( Spitale et al., 2025 )
✓
✓
○
✓
○
✓
28 d × 1/week
Robotic well-being coaching through positive-psychology exercises.
Table 1. Comparison of selected longitudinal language-model-based emotional support systems. Icons: ✓ yes, ○ partial, blank = not reported. Emo. = emotion regulation as the target; State. = explicit state representation; Supp. = multiple support strategies; Cond. = state-conditioned strategy selection; Mem. = cross-session user memory; Adapt. = memory updated from subsequent user feedback; Schedule = deployment duration and interaction frequency.
Figure 2. Prompt evaluation pipeline. (a) Four emotion-inference prompts with incremental reasoning scaffolding are compared on (b) 30 scenarios pairing personas with lived-experience events, through (c, d) blind pairwise bracket tournaments by three human raters, (e) yielding two top performers with complementary strengths, both deployed in PAIR. A five-panel diagram showing four prompt designs, the construction of 30 persona and event scenarios, a blind pairwise prompt-evaluation bracket, the raters' scoring interface with three rating dimensions, and the two winning prompts selected for deployment.
Figure 3. Event structure and emotional mixing were associated with prompt preference in the 30-scenario formative set, while preferences were more evenly divided for personal-history use and goal clarity. Event structure and emotional mixing largely overlapped, and the deployed router uses event structure alone. Bars give the number of scenarios in which each prompt was preferred by majority vote based on the raters’ rankings; the right column reports the winning share. A butterfly chart with eight rows grouped into four scenario features. Scenarios won by Cog extend left in green, those won by Combined extend right in orange. For event structure and emotional mixing the long bar switches sides between the feature's two levels; for personal-history use and goal clarity both levels show short, near-equal bars on either side.
Cog (appraisal-based)
Combined (integrative)
Best suited for
Single-component events (14/17): discrete incidents, isolated conflicts, and single anticipatory states
Focused, person-centered reasoning. Follows one clear causal chain.
Preserves multiple emotional and causal components. Captures bodily and layered experiences.
Failure mode when mismatched
Omits relevant components, especially bodily signals and layered emotions.
Introduces unsupported links to the persona or past experience.
Table 2. Complementary strengths and failure modes of the two selected prompts.
Figure 4. User interaction flow and guidance logic in PAIR. (a) The user recounts an everyday event through speech or text. (b) The system perceives the user’s current emotional state using one of two theory-grounded reasoning scaffolds, selects a regulation strategy based on the event’s intensity and controllability, and generates an internal guidance plan. (c) The plan is delivered through a brief guided conversation with coordinated multimodal cues. After guidance, the system produces a post-guidance estimate from the original event, the initial estimate, and the conversation transcript. A three-panel diagram showing the user interaction flow and core guidance logic in \sysname{}. Panel (a) shows a user describing an everyday event through speech or text. Panel (b) shows the system perceiving the user's current emotional state, selecting among four regulation-strategy families based on intensity and controllability, and generating an internal guidance plan. Panel (c) shows the plan delivered through a short guided conversation with multimodal cues, followed by a post-guidance estimate produced from the original event, the initial emotion estimate, and the conversation transcript.
Figure 5. Study procedure. Day 0: researcher-led onboarding with consent and app setup, baseline surveys, the personalization interview, and a calibration and practice exercise. Days 1–14: three types of daily touchpoints (morning check-in, just-in-time logging, evening reflection), each pairing the system’s affective estimates with SAM self-ratings before and after the guided conversation. Day 14+: post-study questionnaire and semi-structured interview. A timeline diagram showing the three study phases: Day 0 onboarding, the 14-day daily companion phase with its per-session step sequence, and the post-study session.
Figure 6. Exit ratings of perceived emotional understanding. In-the-moment understanding item had the highest mean rating, while emotional nuance recognition had the lowest. Boxes show the distribution of participant ratings, grey points are individual participants, and diamonds with error bars show means with 95% confidence intervals. Items were rated on a 7-point scale (1 = strongly disagree, 7 = strongly agree). Horizontal box plot with three rows, one per questionnaire item, on a one-to-seven scale. All three boxes sit between five and six. The in-the-moment item has the highest mean and the tightest spread.
Dimension / Estimator
MAE
r
Signed err.
Initial
Valence
1.196
.669***
+0.213
Dominance
1.299
.538***
−0.061
Arousal
1.651
.236***
−0.129
Post-guidance
Valence
1.115
.614***
−0.128
Dominance
1.189
.495***
−0.393
Arousal
1.963
.175***
−1.097
Table 3. Estimation accuracy against unanchored self-report on the 9-point SAM scale. Initial estimates are evaluated against before-guidance ratings and post-guidance estimates against after-guidance ratings. The bottom three rows compare initial valence estimators across all narrations, with Claude Opus 4.8 used for both LLM conditions. Signed error is computed as estimate minus self-report, so positive values indicate overestimation. All correlations are significant at p<.001 (***).
Figure 7. Exit ratings of perceived emotion-regulation support. Relevant emotional guidance and support for emotional coping received the highest mean ratings (both M=5.26 ), while improved emotional self-understanding received the lowest ( M=4.63 ). Boxes show participant rating distributions, grey points show individual participants, and diamonds with error bars show means with 95% confidence intervals. Items were rated on a 7-point scale (1 = strongly disagree, 7 = strongly agree). A horizontal box plot with six emotion-regulation support items on a 1–7 self-rating scale. From top to bottom, the mean ratings are 4.63 for improved emotional self-understanding, 5.26 for relevant emotional guidance, 4.84 for actionable emotion-management ideas, 4.79 for a new perspective, 5.26 for emotional coping, and 4.89 for everyday emotion management. Grey points show individual participant ratings, boxes show their distributions, and black diamonds with error bars show means with 95% confidence intervals.
Figure 8. Guided conversations were followed by state-dependent affective movement. (A) Arrows connect clustered mean self-reported valence and arousal before guidance to the corresponding means after guidance; labels show the number of sessions in each cluster. Anxious and dejected trajectories generally moved toward higher valence, with arousal decreasing for anxious states and increasing for dejected states. Calm trajectories also became more activated, while excited trajectories changed least. (B) Example session from each starting-state quadrant illustrates these directions by connecting the narrated event and guidance with the self-reported state after guidance. A two-part figure plotting valence horizontally and arousal vertically. Panel A divides the rating space into anxious or agitated, excited or energized, dejected or flat, and calm or content quadrants. Fifteen arrows connect the mean state before guidance, shown by a circle, to the mean state after guidance for clusters of nearby starting coordinates; each arrow is labelled with its session count. Most anxious arrows point right and downward, dejected and calm arrows point right and upward, and excited arrows are shorter and vary in direction. Panel B contains four example trajectories, one per quadrant. Each example shows a dashed circle for the state before guidance, a solid circle for the state after guidance, an arrow between them, and text boxes summarizing the event, guidance, and post-guidance experience.
Figure 9. Perceived understanding was associated with within-session valence change, and participant corrections identified recurring system mismatches. (A) Mean valence change (after minus before) by understanding rating. Bars show means, error bars show 95% confidence intervals, and labels show session counts; higher ratings corresponded to greater valence gains ( r=.134 , 95% CI [.061,.208] ). (B) The 44 correction texts by coded category, each paired with a representative correction. The four system-attributable categories contained 37 corrections, representing 3.4% of all sessions; the remaining seven indicated that the participant was unable to evaluate the system. A two-panel figure. Panel A is a bar chart of mean valence change after versus before guidance across understanding ratings from 1 to 7. Bars for ratings 1 to 3 extend below zero, while bars for ratings 4 to 7 extend above zero. Error bars show 95% confidence intervals, and labels above the bars show 2, 6, 15, 21, 575, 377, and 97 sessions. Panel B is a horizontal bar chart of 44 correction texts grouped into emotion misread (15, 34%), context confusion (8, 18%), memory failure (8, 18%), strategy mismatch (6, 14%), and unable to evaluate (7, 16%). Each category includes a representative participant correction. A bracket groups the first four categories as 37 corrections attributable to the system, or 3.4% of all sessions.
Figure 10. Exit ratings of perceived personalization and ease of use over the 14-day deployment. Feeling increasingly known and ease of use received higher mean ratings than personalized emotion estimates. Items were rated on a 7-point scale (1 = strongly disagree, 7 = strongly agree). Horizontal box plot with three rows on a one-to-seven scale. Learning and ease of use have means near five and a half. Personalization has the lowest mean.
Figure 11. Four design implications for sustained emotional support, organized around evolving support goals, guidance continuity, and user-controlled memory. Each implication connects its rationale to a system requirement. A four-column table presents design focuses, design implications, rationales, and system requirements. Evolving support goals motivate aligning listening, advice, and session completion with current needs. Guidance continuity motivates adapting recommendations to prior attempts, outcomes, and constraints. User-controlled memory motivates two implications: making memory sources visible and corrections persistent, and giving users control over stored details, retention periods, and contextual reuse.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 12. Runtime orchestration for a single session. An event narration and the persisted user context enter a router that counts interacting emotional components and sends the narration to the cognitive-appraisal scaffold or, where several components interact, to the combined scaffold. Both produce one unified output: a dimensional emotion estimate with a matching SAM reading, emotion labels, appraised intensity, controllability, and situation phase, from which the regulation strategy and all delivery parameters are derived. The guided conversation then advances one plan step per turn along the exploration, comforting, and action arc until it closes. The scaffold selected by the router then re-runs over the event, the estimate, and the transcript to produce a post-guidance estimate. A nightly pass folds themes, calibration trend, corrections, and technique outcomes back into the user context. A left-to-right flow diagram of one session. Event narration and a persistent user context block feed a router, which branches on whether the narration contains multiple interacting emotional components and selects one of two appraisal scaffolds. Both scaffolds converge on a unified output box holding the emotion estimate and a strategy and delivery box. A guided conversation loop follows, cycling companion turns until a close condition is met, and its transcript feeds a post-guidance estimate. A dashed return path labelled nightly memory update runs from the end of the pipeline back to the user context.
Prompt
#1
#2
#3
#4
Mean rank
Cog
39
26
19
6
1.91
Combined
34
31
8
17
2.09
Neuro
13
12
51
14
2.73
Baseline
4
21
12
53
3.27
Appendix
Table 4. Ranking distribution of the four emotion-inference prompts across 90 scenario-rater rankings (30 scenarios × 3 raters, overall-preference dimension). Each cell counts how often a prompt placed at that rank in the blind bracket tournament; lower mean rank is better. Results were consistent across the outcome-plausibility and reasoning-fidelity dimensions, which are omitted for brevity.
Characteristic
n (%) or M ( SD )
Demographics
Age in years
25.1 (5.1)
Age range
18–33
Woman
13 (68%)
Man
6 (32%)
Asian
15 (79%)
Appendix
Table 5. Participant characteristics from the onboarding questionnaires ( N=19 ). AI familiarity was rated on a 5-point scale from not familiar at all to extremely familiar. The three ERQ-S expressive-suppression items were rated on a 7-point scale from strongly disagree to strongly agree; the composite is their mean (Cronbach’s α=.74 ).
Source
Data
Onboarding (Day 0)
Demographics; AI background; ERQ suppression; persona interview responses; calibration narration + SAM ratings
Self-report (per session)
Before/after SAM ratings (valence, arousal, dominance); three feedback items; understanding rating + correction text if ≤ 4; rotating-question responses
System logs (per session)
Narration text and modality; initial and post-guidance affective estimates (PAD, SAM, emotion labels, intensity); selected regulation strategy; per-turn conversational stage and support-strategy labels; full transcript with timestamps; delivery-channel state (color, pulse speed, expression sequence, speech on/off, language); prompt version and persona snapshot
Engagement telemetry
App opens; just-in-time logging frequency and timing; missed sessions; voice vs. text ratio; slider interaction dynamics
Table 6. Data collected per session and per study phase.
Figure 13. Self-reported affect before and after guided conversation. Boxes show session-level distributions, grey points show each participant’s mean, and diamonds with error bars show the session-weighted mean across sessions with 95% participant-bootstrap confidence intervals. Δ denotes mean within-session change and rrb the rank-biserial correlation from participant-level Wilcoxon signed-rank tests. All three contrasts remained significant after Holm correction ( pHolm<.001 ). Six horizontal box plots on a 1–9 rating scale show valence, arousal, and dominance before and after guidance. For each dimension, the after-guidance distribution is shifted to the right of the before-guidance distribution. Grey points show participant means, and black diamonds with confidence intervals show session-level means. Labels report mean increases of 0.53 for valence, 0.33 for arousal, and 0.47 for dominance, with rank-biserial correlations of .99, .87, and 1.00, respectively.
Rating
SM
CC
RM
AD
χ2(3)
p
pFDR
Helpfulness
5.03
5.16
5.01
5.00
3.00
.391
.395
Feasibility
5.28
5.18
5.20
5.04
2.98
.395
.395
Willingness
5.24
5.09
5.09
5.10
3.72
.293
.395
Appendix
Table 7. Mean immediate feedback on 7-point scales by selected regulation strategy. SM denotes situation modification ( n=317 ), CC cognitive change ( n=246 ), RM response modulation ( n=256 ), and AD attentional deployment ( n=274 ). Means are unadjusted session averages. Test statistics refer to the strategy term in the mixed-effects models.
M (SD)
Day-continuous model
Week-binned model
Measure
Overall
Week 1
Week 2
bday×100 [95% CI]
pFDR
Δweek [95% CI]
pFDR
Counts (negative binomial)
Sessions per day
4.13 (1.06)
4.16 (1.13)
4.10 (1.00)
−0.04 [ −1.52 , 1.45 ]
.963
−0.019 [ −0.137 , 0.098 ]
.747
Conversation turns
10.02 (2.80)
10.23 (2.93)
9.87 (2.66)
−0.50 [ −0.98 , −0.02 ]
.091
−0.037 [ −0.075 , 0.001 ]
.109
Log-transformed (lengths in characters, ∣eV∣ in SAM points with log1p)
Narration length
59.5 (105.9)
69.9 (140.7)
48.6 (45.0)
−3.43 [ −4.75 , −2.10 ]
<.001
−0.234 [ −0.341 , −0.127 ]
<.001
Appendix
Table 8. Longitudinal models for the nine session measures across the 14-day deployment. M (SD) are unadjusted means over all sessions and within Week 1 (days 1–7) and Week 2 (days 8–14). bday is the per-day slope from the day-continuous mixed-effects model (reported ×100 ), and Δweek is the Week 2 minus Week 1 contrast from the week-binned model. Estimates for count and log-transformed outcomes are on the link scale.
Emotional support conversation studies how people and AI systems provide empathetic, safe help through dialogue, including counseling, peer support, and companionship.
Recent advances in foundation models have enabled conversational agents that aim for sustained companionship rather than mere task completion. Yet most still remain unable to support natural, long-term companion-like interactions, resulting in experiences that feel episodic and inauthentic. We argue that current agents overlooked cross-temporal modeling of agents' social behaviors and internal emotions: generated behaviors rarely influence an agent's emotional state, and emotional states seldom shape subsequent behaviors. We present Cross-Temporal Emotion Modeling (CTEM), a framework that links long-term behavioral history to moment-to-moment emotional expression. CTEM establishes a closed loop where past experiences update an evolving emotional state; this state conditions immediate interactions; and user feedback continually revises both memory and emotional state, enabling reflection and anticipation. We instantiate CTEM as Auri, a companion agent on an instant-messaging platform, and report a 21-day in-the-wild study showing that CTEM shows improvements in perceived naturalness, coherence, and emotional harmony.
Feier Qin, Xiao Li, Yi Zheng +5
Communication University of China BeiJing, China · Microsoft Research Asia BeiJing, China · Institute of Artificial Intelligence, China Telecom BeiJing, China
Lifelong digital companions must integrate cross-session cues, continually update their understanding of users, and adapt to shifting privacy boundaries. Existing evaluations fail to capture this, testing memory recall and short-term empathy in isolation. To bridge this gap, we introduce \benchmark, a benchmark centered on multi-session Memory-Emotion-Environment loops. By modeling users as persistent worlds with layered profiles and event trajectories, \benchmark uses multi-agent simulation to project environmental dynamics into dialogue, preserving the critical gap between latent thoughts and observable expressions. Evaluating 2,000 personas and 111K tasks across memory tracking, user understanding, privacy control, and emotional companionship, our experiment results reveal a stark reality: even models that saturate current memory benchmarks fail to sustain accurate user understanding and true companionship over long horizons.
Yuqian Wu, Zhijie Deng, Wei Chen +8
Hong Kong University of Science and Technology (Guangzhou) · Tencent · Hong Kong University of Science and Technology
Public discourse and emerging policy typically assume that AI emotional support is a deliberate act: a lonely user consciously seeking comfort from a dedicated companion chatbot. In this paper, we draw on emerging empirical evidence and argue that this picture is inaccurate on two accounts, both in how AI emotional support arises and how it shapes future behavior. First, AI emotional support commonly emerges incidentally within task-oriented interactions on general-purpose platforms, much as workplace friendships deepen through collaboration. Second, these incidental encounters are path-dependent: positive experiences of AI emotional support update people's beliefs about AI's emotional capabilities and redirect their choices for future emotional support, increasing preference for AI and decreasing preference for humans. We review recent evidence, including a large-scale longitudinal study conducted in collaboration with OpenAI, showing that daily five-minute conversations with an AI about personal issues over 28 days led to a 10.3% decrease in the preference for seeking support from humans and an 11.6% increase in the preference for AI. These findings suggest that current policy, focused on companion apps and isolated interactions, cannot adequately protect human connection. Instead, effective regulations should extend to general-purpose AI systems and address cumulative, trajectory-level changes in how people seek support. Recognizing how people stumble into AI emotional support and how those encounters redirect human connections over time is essential to safeguarding human well-being.
Yaoxi Shi, Cathy Mengying Fang, Pattie Maez +1
Imperial College Business School, Imperial College London; London, UK. · Harvard Business School AI Institute, Harvard University; Cambridge, USA. · MIT Media Lab, Massachusetts Institute of Technology; Cambridge, USA. +2