Pith. sign in

REVIEW 4 major objections 5 minor 6 cited by

MASCOT claims a two-level optimization—per-agent persona rewards plus a director-level group reward—keeps AI companions in distinct roles and produces complementary, non-redundant dialogue.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 09:15 UTC pith:7P5JZYJL

load-bearing objection A plausible training pipeline for multi-agent social companions, undercut by a closed-loop LLM evaluation and an unfulfilled human-evaluation claim. the 4 major comments →

arxiv 2601.14230 v2 pith:7P5JZYJL submitted 2026-01-20 cs.CL cs.AIcs.HC

MASCOT: Towards Multi-Agent Socio-Collaborative Companion Systems

Yiyang Wang , Yiqiao Jin , Alex Cabral , Josiah Hester This is my paper
classification cs.CL cs.AIcs.HC
keywords multi-agent systemspersona collapsesocial sycophancyRLAIFGRPOLLM-as-a-judgeempathetic dialoguecollaborative dialogue optimization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes MASCOT, a framework for building multi-agent 'socio-collaborative' companions: several AI agents with distinct personas converse with a user and each other to give emotional or workplace support. Its central claim is that the two failure modes that plague such systems—persona collapse, where all agents drift into a generic assistant voice, and social sycophancy, where agents pile on agreeable, redundant comments—can be fixed by a two-level training scheme. At the individual level, each agent is fine-tuned via reinforcement learning from AI feedback to adhere to its persona; at the group level, a meta-agent 'director' is trained to decide who speaks and how, guided by rewards for coherent, non-redundant conversation. On empathetic-dialogue and meeting-summarization benchmarks, MASCOT reports up to +14.1 in persona consistency and +10.6 in social contribution over strong baselines. If true, this offers a practical, low-cost recipe for building companions that behave as an ensemble of distinct perspectives instead of a single echo-chambered assistant.

Core claim

MASCOT's central claim is that a bi-level optimization strategy harmonizes individual persona fidelity with collective dialogue synergy. The first level, Persona-Aware Behavioral Alignment, uses an RLAIF pipeline: candidate responses from a supervised policy are scored by an LLM judge on persona-specific rubrics; a reward model is trained on those pairwise preferences; and GRPO fine-tunes each agent to stay in character while keeping outputs concise. The second level, Collaborative Dialogue Optimization, introduces a meta-agent (the director) that observes the conversation and issues per-turn directives—who should speak next and with what intent—while being optimized by GRPO against a group

What carries the argument

Bi-level optimization with a director–speaker decoupling. Level 1 (per-agent): a learned persona reward model, derived from LLM-judge preferences, drives GRPO fine-tuning of each speaker agent toward persona fidelity and concise, reasoning-tagged outputs. Level 2 (group): a meta-agent 'director' generates per-turn directives selecting the next speaker and the intended contribution; the director itself is trained via GRPO on a group reward Rgroup = coherence + η·diversity-penalty, rewarding logical flow and penalizing repeated or redundant speaker turns. The key mechanism is the separation of 'who coordinates the conversation' (director) from 'who generates the line' (speaker), with each leve

Load-bearing premise

The reported gains depend on GPT-4o's rubric-based scores being a valid and unbiased proxy for human judgments of empathy, consistency, and contribution—a premise made fragile by the fact that the same model family generated the preference data for training the reward models and no human evaluation results are actually provided.

What would settle it

Run a blinded human evaluation on a random sample of MASCOT and the strongest baseline (MultiAgentESC) outputs from the same test sets (e.g., ED-Negative and QMSum-Product), scoring the identical rubrics. If human raters do not place MASCOT ahead of the baseline by roughly the same margins as the GPT-4o judge, the central performance claim is not supported.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the reported gains hold, multi-agent companions can avoid the 'model mean' phenomenon and sustain distinct voices across long, multi-turn conversations.
  • Group-level directive training with redundancy penalties can suppress echo-chamber behavior, making multi-agent dialogue more informative to the user.
  • The pipeline is parameter-efficient (LoRA on ~0.187% of parameters) and compatible with open-source models, so it can be reproduced at modest cost.
  • Both emotional support and workplace collaboration benefit, suggesting the director–speaker architecture is domain-general.
  • Ablations pin the division of labor: persona alignment drives individual quality, while collaborative optimization primarily shapes turn-taking and global coherence.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The evaluation is closed-loop: GPT-4o generates the preferences that train the reward models and also supplies the final rubric scores, with no human ratings reported despite the abstract's mention of human evaluation; if those scores are biased, the margin over baselines may shrink under independent human assessment.
  • The director–speaker decoupling is a transferable design for other multi-agent social tasks—e.g., tutoring or group decision support—where role separation and structured turn-taking matter.
  • The reported 'prudence trade-off' on high-arousal negative emotions implies that emotional calibration should be state-dependent; a testable extension is an adaptive persona that modulates emotional intensity based on inferred user arousal.
  • Since the paper limits itself to English and two domains, a natural extension is to evaluate whether persona modeling transfers cross-lingually and cross-culturally, or whether persona definitions must be re-grounded in local norms.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. MASCOT proposes a multi-agent socio-collaborative companion framework that addresses persona collapse and social sycophancy through a two-stage optimization pipeline: (1) Persona-Aware Behavioral Alignment, which uses RLAIF to fine-tune individual speaker agents against a reward model trained from LLM-judge preferences, and (2) Collaborative Dialogue Optimization, which trains a director meta-agent to issue turn-taking directives that maximize group-level coherence and diversity. The paper reports evaluations on Empathetic Dialogues and QMSum across agent-specific and collective-interaction metrics, claiming gains up to +14.1 in persona consistency and +10.6 in social contribution over prompting, SFT, and multi-agent baselines. The abstract states the evaluation suite includes human evaluation, but the full text describes only LLM-as-a-judge evaluation with GPT-4o, which is also the teacher used to construct the preference data for training the reward model.

Significance. If the reported improvements were independently validated, the framework would be a useful design pattern for multi-agent systems: it decomposes social quality into persona-level fidelity and group-level synergy, and the ablations and sensitivity analyses in Section 3.4 and Appendix B.1 give some insight into the contribution of each component. The paper also makes a practical choice by using LoRA and a small director model, addressing deployment concerns. However, the central empirical claim rests on a closed-loop evaluation: GPT-4o supplies the training preferences, the reward model is trained on those preferences, and GPT-4o is the sole evaluator of the final system. Without human validation or a held-out judge, the reported gains may reflect self-preference rather than genuine social quality. This is a load-bearing threat to the paper's main conclusion.

major comments (4)
  1. [Abstract / Appendix B.2] The abstract explicitly states that the evaluation suite includes 'human evaluation,' but Appendix B.2 says flatly 'we adopt LLM-as-a-Judge for evaluation,' and Appendix B.3 describes only GPT-4o as the evaluator. I could not find any human evaluation results in the manuscript. This is not a stylistic discrepancy: an abstract claim of human validation is contradicted by the full text. If human evaluation is not available, the claim must be removed; if it exists, the results, protocol, and inter-annotator agreement must be reported.
  2. [Section 2.1 and Appendix B.2/B.3] The evaluation is closed-loop. Section 2.1 builds the preference dataset by having an LLM judge score K=8 candidates (Eq. 1–2); Appendix B.2 identifies GPT-4o as the teacher. This dataset trains the reward model r_phi (Eq. 3), which is then optimized by GRPO. The final evaluation in Appendix B.3 uses GPT-4o as the judge with rubrics that overlap with the training criteria. Any systematic stylistic self-preference of GPT-4o will inflate MASCOT's scores relative to prompting and SFT baselines that were not optimized against that judge. The paper needs a human evaluation, or at minimum a different, independently validated judge, to demonstrate that the gains are not an artifact of reward hacking.
  3. [Section 3.2, Table 1] The SFT baseline is anomalously low: on ED-Negative it scores 44.1 Overall versus 55.1 for zero-shot, and the pattern repeats across all subsets. Since the SFT baseline is trained on GPT-4o-generated trajectories, one would expect it to be competitive with or better than zero-shot. The 'mean-seeking' explanation in Section 3.2 is speculative and not supported by any analysis (e.g., output length, diversity, or loss curves). Moreover, Section 2.1 uses pi_SFT as the base for MASCOT; if this SFT run is degenerate, the relative gains of MASCOT may be inflated. The authors should verify the SFT baseline's training setup, report training details, and consider whether the comparison is a fair distillation baseline.
  4. [Appendix B.3, Tables 5–6] The paper provides no evidence that the GPT-4o rubric-based scores are valid proxies for human judgments of empathy, consistency, or contribution. There is no human correlation study, no inter-annotator reliability, and no calibration of the 1–5 Likert scale before linear rescaling to [0,100]. Given that the same model family was used in training, this is not a minor psychometric concern. The headline improvements (+14.1 consistency, +10.6 contribution) are only as meaningful as the judge itself. A small human evaluation on a random subset, with agreement statistics, would be necessary to support the claims.
minor comments (5)
  1. [Section 3.2] Typo: 'static training dat with expert demonstrations' should be 'static training data'.
  2. [Appendix B.2] The reference to Qwen3-8B cites 'Qiu et al., 2025' and 'Yang et al., 2025.' The Qiu et al. reference appears to be a different paper on gated attention, not a Qwen3 technical report. Please correct the citation.
  3. [Section 3.1 and Appendix 7] The text refers to 'Appendix 7' for persona descriptions, but the content appears in Table 7. Please fix the cross-reference.
  4. [Table 1] The header 'PositiveEmotion' lacks a space; similar formatting issues appear in Table 2.
  5. [Figure 2] The label 'E xpert D ata Collection' has unintended spaces; please reformat.

Circularity Check

1 steps flagged

Closed-loop evaluation: GPT-4o supplies the RL preference signal and then serves as the only reported judge; the abstract's promised human evaluation is absent.

specific steps
  1. fitted input called prediction [Section 2.1 (Eqs. 2–5), Appendix B.2, Appendix B.3, Tables 5–6]
    "Each candidate response yk is evaluated by an LLM judge against a set of fine-grained criteria C (Table 5) ... We then train a reward model rϕ ... During sampling, we use GPT-4o as the teacher model ... we adopt LLM-as-a-Judge for evaluation ... We use GPT-4o as the evaluator with the temperature set to 0.0."

    The RL policy is optimized by GRPO against R = rϕ + λRformat (Eqs. 4–5), and rϕ is trained on GPT-4o rubric scores using the same Table 5 criteria that later define the final evaluation. The final judge is GPT-4o at temperature 0.0 (B.3). Thus MASCOT is optimized to maximize the very model and rubric that then produce the reported +14.1 / +10.6 gains, while prompting baselines are not optimized against that judge. The gain is therefore, to an unbroken extent, a measure of fit to the judge rather than an independent measure of social quality; no human evaluation is reported to break the loop.

full rationale

The central evaluation loop is the load-bearing circularity. Section 2.1 constructs the persona reward model from GPT-4o's rubric scores (Eqs. 1–3), and GRPO maximizes that reward (Eqs. 4–5). Appendix B.3 then evaluates with 'GPT-4o as the evaluator' using the same agent-specific and collective rubrics (Tables 5–6) that generated the preference data. The abstract says the evidence includes 'human evaluation,' but no human evaluation results appear anywhere in the manuscript; Appendix B.2 states the evaluation is LLM-as-a-Judge. This missing external anchor is material: without human ratings, the reported improvements may reflect reward-model/judge alignment rather than genuine persona fidelity or social contribution. I do not score this as an 8 because the paper is not relying on a self-citation chain or on a definitional identity; it is a valid RL pipeline whose validation is closed-loop. I also note the SFT baseline, also trained on GPT-4o trajectories, scores far below zero-shot (44.1 vs 55.1 on ED-Negative), which complicates a pure stylistic self-preference story; nevertheless the structural loop remains. The abstract's 'multiple LLM judges' is also unsupported: Appendix B.3 names only GPT-4o. No code or data are released, so the loop cannot be checked externally. Score 7.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The framework rests on a few hand-set hyperparameters (δ, format thresholds) and several strong domain assumptions. The most impactful is the closed-loop reliance on GPT-4o: it generates preference labels, trains the reward model, and then judges the final outputs. There are no human annotations or external benchmarks to break this loop, and the abstract's claim of human evaluation is not reflected in the methods.

free parameters (5)
  • Preference margin δ = 0.5
    Used to filter ambiguous preference pairs in the reward model training; chosen by hand.
  • Format reward weight λ = Not specified
    Weights the rule-based format reward in the composite reward R = r_phi + λ R_format; no value or tuning procedure is given.
  • Format reward constraints = reasoning trace >448 tokens, final response <64 tokens
    Hand-set thresholds to enforce chain-of-thought and brevity; arbitrary and not justified relative to human preferences.
  • GRPO group size G = Not specified
    Number of sampled outputs per group in GRPO; not stated in the paper.
  • LoRA rank/alpha = r=16, α=32
    Standard LoRA hyperparameters, hand-chosen.
axioms (4)
  • domain assumption GPT-4o judge scores are a valid proxy for human dialogue quality.
    The entire evaluation and preference labeling rely on GPT-4o scoring a 1–5 rubric; no correlation with human ratings is provided.
  • domain assumption Persona descriptions and rubrics are sufficient to capture meaningful persona consistency and social contribution.
    The criteria in Tables 5 and 6 are defined by the authors; there is no independent validation that these criteria align with human intuition.
  • domain assumption RLAIF training on GPT-4o preferences transfers to better real-world support behavior.
    The policy is optimized against a reward model learned from GPT-4o's preferences, assuming this improves actual persona fidelity in deployment, which is not tested with humans.
  • domain assumption Qwen3-8B retains enough base competence after LoRA fine-tuning to function as a coherent social agent.
    Catastrophic forgetting is not measured; the paper only notes the SFT baseline suffers from 'mean-seeking', but the same risk applies to RL fine-tuning.
invented entities (1)
  • Director meta-agent no independent evidence
    purpose: Issues high-level directives at each turn (e.g., which speaker, what angle) to coordinate group dialogue for coherence and diversity.
    This is an architectural component introduced by the paper, but its effectiveness is only measured through the same LLM judge used for training; no external evidence (e.g., human preference, downstream task success) is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 243 in / 6827 out tokens · 134357 ms · 2026-08-03T09:15:31.230727+00:00 · methodology

0 comments
read the original abstract

Multi-agent systems (MAS) are emerging as promising socio-collaborative companions for emotional and cognitive support. However, existing systems frequently suffer from persona collapse, where agents revert to generic, homogenized assistant behaviors, and social sycophancy, where agents produce redundant, non-constructive dialogue. We propose MASCOT, a multi-agent framework for multi-perspective socio-collaborative companions. MASCOT introduces a novel bi-level optimization strategy to harmonize individual and collective behaviors: 1) Persona-Aware Behavioral Alignment, an RLAIF-driven pipeline that fine-tunes individual agents for agent-specific identities; and 2) Collaborative Dialogue Optimization, a group-level adaptation process that promotes complementary, diverse, and productive discourse. We evaluate MASCOT using human-grounded contexts drawn across both in-domain and out-of-domain (OOD) settings against state-of-the-art baselines. MASCOT improves persona consistency by up to +14.1 and social contribution by up to +10.6. A broad evaluation suite, including human evaluation, multiple LLM judges, three-way comparisons, and automatic metrics, further shows that MASCOT produces more role-consistent and less redundant multi-agent dialogue.

Figures

Figures reproduced from arXiv: 2601.14230 by Alex Cabral, Josiah Hester, Yiqiao Jin, Yiyang Wang.

Figure 1
Figure 1. Figure 1: Moving beyond dyadic interactions to multi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of MASCOT for multi-agent socio-collaborative companions. MASCOT produces synergistic multi-agent dialogues that maintain distinct personas while collectively supporting the user. It leverages a two-phase optimization pipeline. In Persona-Aware Behavioral Alignment, individual agents are fine-tuned via RLAIF with a learned reward model to ensure stable persona fidelity and high-quality responses. … view at source ↗
Figure 3
Figure 3. Figure 3: Performance of MASCOT on the Neutral sub￾set of Empathetic Dialogues. approaches. Challenges in Negative Valence Most models exhibit a performance decline in the ED-Negative subset ( [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of MASCOT variants. MASCOT-P disables the Persona-Aware Behavioral Alignment (Section 2.1); MASCOT-C removes the Collaborative Optimization (Section 2.2). Emotion Empathy Relevance Consistency Contribution 60 65 70 75 80 Performance of Diverse Agents Anchor Beacon Catalyst [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance of MASCOT when interacting with users with different MBTI types on Empathetic Dialogues. 0.6B 1.7B 8B (a) Director Agent Sizes 40 50 60 70 80 Overall Performance Collective Persona 0.6B 1.7B 8B (b) Speaker Agent Sizes MASCOT Performance across Model Sizes [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance of MASCOT on Empathetic Di￾alogues dataset with varying model sizes for directive (left) and speaker agents (right). performance (agent-specific: 68.4 → 67.2; col￾lective: 71.5 → 69.1). However, at 0.6B, Collec￾tive performance plummets to 51.7, while agent￾specific scores remain relatively resilient at 63.9. Although a small director can still select the appro￾priate persona, it lacks the reas… view at source ↗
Figure 8
Figure 8. Figure 8: Performance of MASCOT on QMSum dataset in terms of Agent-specific metrics. 50 60 70 80 Score Fidelity Originality Relevance Engagement Overall Academic 50 60 70 80 Score Committee 50 60 70 80 Score Product Performance on QMSum (Collaborative Interaction Metrics) Zero-Shot Zero-Shot CoT Few-Shot Few-Shot + CoT SFT Self-Consistency MultiAgentESC MASCOT [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Performance of MASCOT on QMSum dataset in terms of Collaborative Interaction Metrics. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. CollabSim: A CSCW-Grounded Methodology for Investigating Collaborative Competence of LLM Agents through Controlled Multi-Agent Experiments

    cs.CL 2026-06 unverdicted novelty 7.0

    CollabSim is a new CSCW-grounded simulation framework that enables controlled multi-agent experiments to measure collaborative competence in LLM agents.

  2. TextReg: Mitigating Prompt Distributional Overfitting via Regularized Text-Space Optimization

    cs.CL 2026-05 unverdicted novelty 7.0

    TextReg mitigates prompt distributional overfitting via regularized text-space optimization, reporting up to +16.5% OOD accuracy gains over prior methods on reasoning benchmarks.

  3. Easier to Mislead Than to Correct: Harmful and Beneficial Revision in LLM Conformity

    cs.CL 2026-06 unverdicted novelty 6.0

    Peer agreement misleads initially correct LLMs more than it corrects initially wrong ones, with authority labels biasing choices independently of accuracy and reasoning prompts failing to mitigate the asymmetry.

  4. UniSD: Towards a Unified Self-Distillation Framework for Large Language Models

    cs.CL 2026-05 unverdicted novelty 6.0

    UniSD unifies self-distillation components for autoregressive LLMs and its full integrated version improves base models by 5.4 points and baselines by 2.8 points across six benchmarks.

  5. CultivAgents: Cultivating Relationship-Centered Multi-Agent Systems for Personalized Gardening

    cs.HC 2026-05 unverdicted novelty 5.0

    Presents CultivAgents, a relationship-centered multi-agent system for socio-culturally grounded gardening support, with a mixed-methods evaluation showing modest gains in gardener confidence and motivation.

  6. UniSD: Towards a Unified Self-Distillation Framework for Large Language Models

    cs.CL 2026-05 unverdicted novelty 5.0

    UniSD unifies complementary self-distillation mechanisms for autoregressive LLMs and achieves up to +5.4 point gains over base models and +2.8 over baselines across six benchmarks and six models.

Reference graph

Works this paper leans on

51 extracted references · 4 canonical work pages · cited by 5 Pith papers · 1 internal anchor

  1. [1]

    Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, and 1 others. 2022. Constitutional ai: Harmlessness from ai feedback. arXiv:2212.08073

    Pith/arXiv arXiv 2022

  2. [2]

    Razan Baltaji, Babak Hemmatian, and Lav Varshney. 2024. https://doi.org/10.18653/v1/2024.c3nlp-1.2 Conformity, Confabulation , and Impersonation : Persona Inconstancy in Multi - Agent LLM Collaboration . In Proceedings of the 2nd Workshop on Cross - Cultural Considerations in NLP , pages 17--31, Bangkok, Thailand. ACL

    doi:10.18653/v1/2024.c3nlp-1.2 2024

  3. [3]

    Ralph Allan Bradley and Milton E Terry. 1952. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324--345

    1952

  4. [4]

    By Katharine C Briggs and Isabel Briggs Myers. 1976. Myers-briggs type indicator

    1976

  5. [5]

    CDC. 2024 a . H ealth E ffects of S ocial I solation and L oneliness. https://www.cdc.gov/social-connectedness/risk-factors/index.html

    2024

  6. [6]

    CDC. 2024 b . S ocial C onnection. https://www.cdc.gov/social-connectedness/about/index.html

    2024

  7. [7]

    Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan Liu. 2024. Chateval: Towards better llm-based evaluators through multi-agent debate. In ICLR

    2024

  8. [8]

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, and 1 others. 2024. A survey on evaluation of large language models. TIST, 15(3):1--45

    2024

  9. [9]

    Jung, Nicola Dell, Deborah Estrin, and James A

    Andrea Cuadra, Maria Wang, Lynn Andrea Stein, Malte F. Jung, Nicola Dell, Deborah Estrin, and James A. Landay. 2024. https://doi.org/10.1145/3613904.3642336 The Illusion of Empathy ? Notes on Displays of Emotion in Human - Computer Interaction . In CHI, pages 1--18, Honolulu HI USA. ACM

    arXiv 2024

  10. [10]

    Daniel Goleman. 2006. Social intelligence: the new science of human relationships. Bantam Books

    2006

  11. [11]

    Ryuichiro Higashinaka, Masahiro Mizukami, Hidetoshi Kawabata, Emi Yamaguchi, Noritake Adachi, and Junji Tomita. 2018. https://doi.org/10.18653/v1/W18-5031 Role play-based question-answering by real users for building chatbots with consistent personalities . In Proceedings of the 19th Annual SIGdial Meeting on Discourse and Dialogue , pages 264--272, Melbo...

    work page doi:10.18653/v1/w18-5031 2018

  12. [12]

    Arlie Russel Hochschild. 2012. The managed heart: commercialization of human feeling, updated ed edition. University of California Press, Berkeley, Calif. London

    2012

  13. [13]

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, and 1 others. 2023. Metagpt: Meta programming for a multi-agent collaborative framework. In ICLT

    2023

  14. [14]

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, Jinlin Wang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, and Chenglin Wu. 2024. Metagpt: Meta programming for a multi-agent collaborative framework. ICLR

    2024

  15. [15]

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2021. Lora: Low-rank adaptation of large language models. In ICLR

    2021

  16. [16]

    Chicago Psychoanalytic Institute. 2025. https://chicagoanalysis.org/blog/elements-of-psychoanalytic-technique/therapeutic-neutrality/ Understanding therapeutic neutrality: Balancing compassion and objectivity

    2025

  17. [17]

    Yiqiao Jin, Qinlin Zhao, Yiyang Wang, Hao Chen, Kaijie Zhu, Yijia Xiao, and Jindong Wang. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.70 AgentReview : Exploring Peer Review Dynamics with LLM Agents . In EMNLP, pages 1208--1226, Miami, Florida, USA. Association for Computational Linguistics

    doi:10.18653/v1/2024.emnlp-main.70 2024

  18. [18]

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, and 1 others. 2017. Overcoming catastrophic forgetting in neural networks. PNAS, 114(13):3521--3526

    2017

  19. [19]

    Marek Korczynski. 2003. https://doi.org/10.1177/1350508403010001479 Communities of coping: Collective emotional labour in service work . Organization, 10(1):55--79. Tex.eprint: https://doi.org/10.1177/1350508403010001479

    work page doi:10.1177/1350508403010001479 2003

  20. [20]

    Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and Sushant Prakash. 2024. Rlaif vs. rlhf: scaling reinforcement learning from human feedback with ai feedback. In ICML. JMLR

    2024

  21. [21]

    Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, and 1 others. 2025. From generation to judgment: Opportunities and challenges of llm-as-a-judge. In ACL, pages 2757--2791

    2025

  22. [22]

    Chia-Wei Liu, Ryan Lowe, Iulian Serban, Mike Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. https://doi.org/10.18653/v1/D16-1230 How NOT To Evaluate Your Dialogue System : An Empirical Study of Unsupervised Evaluation Metrics for Dialogue Response Generation . In EMNLP, pages 2122--2132, Austin, Texas. ACL

    doi:10.18653/v1/d16-1230 2016

  23. [23]

    Siyang Liu, Chujie Zheng, Orianna Demasi, Sahand Sabour, Yu Li, Zhou Yu, Yong Jiang, and Minlie Huang. 2021. https://doi.org/10.18653/v1/2021.acl-long.269 Towards Emotional Support Dialog Systems . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Proc...

    doi:10.18653/v1/2021.acl-long.269 2021

  24. [24]

    Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. Simpo: Simple preference optimization with a reference-free reward. NeurIPS, 37:124198--124235

    2024

  25. [25]

    OpenAI. 2025. https://chat.openai.com/ Gpt-4o

    2025

  26. [26]

    Ana Rita Ortigoso, Gabriel Vieira, Daniel Fuentes, Luis Frazão, Nuno Costa, and António Pereira. 2025. https://doi.org/10.48550/arXiv.2505.20521 Project Riley : Multimodal Multi - Agent LLM Collaboration with Emotional Reasoning and Voting . arXiv preprint

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.20521 2025

  27. [27]

    Patil, Ion Stoica, and Joseph E

    Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. 2024. https://doi.org/10.48550/arXiv.2310.08560 MemGPT : Towards LLMs as Operating Systems . CoRR

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.08560 2024

  28. [28]

    Bernstein

    Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. https://doi.org/10.1145/3586183.3606763 Generative Agents : Interactive Simulacra of Human Behavior . In UIST, pages 1--22, New York, NY, USA. ACM

    arXiv 2023

  29. [29]

    Bernstein

    Joon Sung Park, Lindsay Popowski, Carrie Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2022. https://doi.org/10.1145/3526113.3545616 Social Simulacra : Creating Populated Prototypes for Social Computing Systems . In UIST, pages 1--18, New York, NY, USA. ACM

    arXiv 2022

  30. [30]

    Zihan Qiu, Zekun Wang, Bo Zheng, Zeyu Huang, Kaiyue Wen, Songlin Yang, Rui Men, Le Yu, Fei Huang, Suozhi Huang, and 1 others. 2025. Gated attention for large language models: Non-linearity, sparsity, and attention-sink-free. In NeurIPS

    2025

  31. [31]

    Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. 2019. Towards empathetic open-domain conversation models: A new benchmark and dataset. In ACL, pages 5370--5381

    2019

  32. [32]

    Sahakian and Christelle Langley

    Barbara J. Sahakian and Christelle Langley. 2025. Why Is Social Interaction So Important?, page 47–56. Cambridge University Press

    2025

  33. [33]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv:2402.03300

    Pith/arXiv arXiv 2024

  34. [34]

    Ashish Sharma, Adam Miner, David Atkins, and Tim Althoff. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.425 A Computational Approach to Understanding Empathy Expressed in Text - Based Mental Health Support . In EMNLP, pages 5263--5276, Online. ACL

    doi:10.18653/v1/2020.emnlp-main.425 2020

  35. [35]

    Daniel N Stern, Lynne Hofer, Wendy Haft, John Dore, and 1 others. 1985. Affect attunement: The sharing of feeling states between mother and infant by means of inter-modal fluency. Social perception in infants, pages 249--268

    1985

  36. [36]

    Olga Sutskova, Atsushi Senju, and Tim J. Smith. 2023. https://doi.org/10.1155/2023/6677789 Cognitive Impact of Social Virtual Reality : Audience and Mere Presence Effect of Virtual Companions . Human Behavior and Emerging Technologies, 2023:1--19

    work page doi:10.1155/2023/6677789 2023

  37. [37]

    Noah Wang, Zy Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, and 1 others. 2024. Rolellm: Benchmarking, eliciting, and enhancing role-playing abilities of large language models. In ACL, pages 14743--14777

    2024

  38. [38]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. In ICLR

    2023

  39. [39]

    Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. 2020. Generalizing from a few examples: A survey on few-shot learning. CSUR, 53(3):1--34

    2020

  40. [40]

    Yiyang Wang, Chen Chen, Tica Lin, Vishnu Raj, Josh Kimball, Alex Cabral, and Josiah Hester. 2025. Companioncast: A multi-agent conversational ai framework with spatial audio for social co-viewing experiences. arXiv:2512.10918

    arXiv 2025

  41. [41]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824--24837

    2022

  42. [42]

    Bosi Wen, Pei Ke, Xiaotao Gu, Lindong Wu, Hao Huang, Jinfeng Zhou, Wenchuang Li, Binxin Hu, Wendy Gao, Jiaxing Xu, and 1 others. 2024. Benchmarking complex instruction-following with multiple constraints composition. NeurIPS, 37:137610--137645

    2024

  43. [43]

    World Health Organization . 2025. S ocial I solation and L oneliness. https://www.who.int/teams/social-determinants-of-health/demographic-change-and-healthy-ageing/social-isolation-and-loneliness

    2025

  44. [44]

    Yangyang Xu, Jinpeng Hu, Zhuoer Zhao, Zhangling Duan, Xiao Sun, and Xun Yang. 2025. Multiagentesc: A llm-based multi-agent collaboration framework for emotional support conversation. In EMNLP, pages 4665--4681

    2025

  45. [45]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv:2505.09388

    Pith/arXiv arXiv 2025

  46. [46]

    Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. https://doi.org/10.18653/v1/P18-1205 Personalizing Dialogue Agents : I have a dog, do you have pets too? In ACL, pages 2204--2213, Melbourne, Australia. ACL

    doi:10.18653/v1/p18-1205 2018

  47. [47]

    Ming Zhong, Da Yin, Tao Yu, Ahmad Zaidi, Mutethia Mutuma, Rahul Jha, Ahmed Hassan, Asli Celikyilmaz, Yang Liu, Xipeng Qiu, and 1 others. 2021. Qmsum: A new benchmark for query-based multi-domain meeting summarization. In NAACL, pages 5905--5921

    2021

  48. [48]

    Li Zhou, Jianfeng Gao, Di Li, and Heung-Yeung Shum. 2020. https://doi.org/10.1162/coli_a_00368 The Design and Implementation of XiaoIce , an Empathetic Social Chatbot . Computational Linguistics, 46(1):53--93. Place: Cambridge, MA Publisher: MIT Press

    doi:10.1162/coli_a_00368 2020

  49. [49]

    Yiming Zhu, Yupeng He, Ehsan-Ul Haq, Gareth Tyson, and Pan Hui. 2025. https://doi.org/10.48550/arXiv.2504.10286 Characterizing LLM -driven Social Network : The Chirper .ai Case . arXiv preprint

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.10286 2025

  50. [50]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  51. [51]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...