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
cs.HC · cs.AI
Quick verdict
Compelling 14-day mixed-methods field study of an LLM emotion companion (19 users, 1,093 sessions); no control group, so causal claims stay weak.
Summary
PAIR is a theory-based LLM companion that estimates users' valence, arousal, and dominance from narrated everyday events, delivers brief regulation guidance through exploration-comforting-action dialogue, and carries persona, calibration, and daily memory summaries across sessions. The authors first ran a blinded prompt tournament (four scaffolds, 30 scenarios, three raters) and deployed the two winners, routed by event structure. A 14-day field deployment with 19 participants (1,093 sessions) paired system estimates with SAM self-reports before and after guidance, alongside logs, evening probes, exit questionnaires, and interviews. Estimates tracked self-reported valence and dominance far better than arousal (r = .236; kappa = .154 at any cut). Guided conversations were followed by valence gains concentrated in anxious (+1.875) and dejected (+1.208) sessions; no regulation-strategy family outperformed another (Friedman p = .897). Perceived understanding tracked conversational acknowledgment rather than numerical accuracy, and helpfulness ratings rose over two weeks while users wrote less. Limitations: no control condition, so gains may partly reflect regression to the mean and in-conversation framing; a young, AI-savvy, largely Asian sample; weak arousal inference; and no code or data release.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Text-based appraisal-guided inference corresponded more closely to self-reported valence and dominance than to arousal (arousal r=.236, κ=.154 at any cut point).
- Guided conversations were followed by state-dependent self-reported affect changes, with the largest valence gains in anxious (+1.875) and dejected (+1.208) sessions and arousal moving down when anxious and up when dejected.
- Perceived understanding emerged from contextual exploration and emotional acknowledgment and was only weakly related to numerical estimation error (ρ between -.014 and .101).
- Perceived helpfulness of guided conversation increased over 14 days (Week2-Week1 +0.216, pFDR=.011) while narration and per-turn input lengths declined (3.37% and 4.30% daily).
- Cross-session personalization depended on applying remembered experiences and on user corrections persisting in memory rather than being forgotten.
Strengths
Rare longitudinal field evidence with paired first-person ground truth
1,093 sessions from 19 participants over 14 days pair each system estimate with SAM self-reports taken before the estimate is revealed (Sec. 5.2.2 anchoring control), grounding RQ1 in the supported user's own ratings rather than third-party corpora.
Theory-grounded pipeline gated by a pre-deployment formative evaluation
Sec. 3's blinded four-prompt tournament on a fixed model (Table 4: mean ranks Cog 1.91, Combined 2.09, Neuro 2.73, Baseline 3.27 across 90 rankings) selected scaffolds before deployment and motivated the event-structure router, which matched majority preference in 26/30 scenarios (87%).
Above-norm statistical transparency including null results
Mixed-methods triangulation converging on mechanisms
Design implications are traceable to specific evidence
Weaknesses
No control or comparison condition
Sample homogeneity and attrition
Arousal near-failure undermines a core design premise
Formative evaluation circularity and thin validation base
Memory-personalization claims rest on anecdotes
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
PAIR (Perceptual Affective Inference and Regulation) is a web-based LLM companion combining (a) appraisal-guided PAD emotion inference, (b) state-conditioned regulation guidance drawn from Gross's process model and delivered through an Exploration-Comforting-Action (ECA) dialogue, and (c) cross-session memory (onboarding persona, calibration example, rolling 150-word daily summaries; Claude Opus 4.8 for inference and dialogue, GPT-5-mini for routing and memory). The paper asks whether momentary emotion understanding translates into sustained support via three RQs: inference accuracy (RQ1), regulation outcomes and guidance fit (RQ2), and cross-session continuity (RQ3). Central claims: text-only estimates track self-reported valence and dominance far better than arousal; guided conversations are followed by state-dependent affect changes; perceived understanding stems from conversational acknowledgment rather than numerical accuracy; helpfulness rises over two weeks; personalization hinges on remembered context and corrections that persist in memory.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is PAIR and what does it do?
PAIR is an LLM-based emotional support companion that infers a user's emotional state (valence, arousal, dominance) from their description of everyday events, holds a short guided regulation conversation, and remembers persona details, calibration examples, and daily summaries to personalize later support. It combines appraisal-theory prompts, strategy selection based on emotion intensity and controllability, and cross-session memory.
Did the companion actually improve users' emotions?
Participants' self-reported valence rose after guided conversations, most when they started anxious (+1.875 points on a 9-point scale) or dejected (+1.208). However, the study had no control group, so part of the change could reflect regression to the mean or the framing of the conversation rather than the agent's guidance.
How accurate was PAIR's emotion inference from text?
Against participants' self-ratings, estimates matched valence and dominance reasonably well (binary agreement kappa = .539 and .430) but arousal poorly (r = .236; kappa no higher than .21 at any cut point), largely because text transcripts lack intensity cues such as tone of voice.
How big was the study and who participated?
19 U.S. adults aged 18-33 (13 women, 6 men; 21 of 24 recruits finished) used PAIR for 14 days, completing 1,093 sessions (about 4 per day). Data included before/after emotion self-ratings, session feedback, correction texts, evening adoption probes, an exit questionnaire, and interviews.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Extraction omits several sections (opening of 6.1, part of 6.4.2, Section 7.1, parts of Appendices C-D); some numbers, e.g. the deployed scaffold's valence r, were truncated mid-sentence.
- Figures 1-13 could not be analyzed from text; cluster-level claims about Figure 8 rest on captions and in-text statistics.
- arXiv preprint formatted for CHI'27; peer-review status unknown.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.