From Triggers to Emotions: A CPM-Grounded Appraisal Multi-Agent for Dynamic Emotional Evolution in Persona-Based Dialogue

Personas, identity, and agents2026arXivApproved editorial review

Authors: Jingyao Cai, Shuaijun Liu, Abdul Rehman, Yutong Guo, Qin Tian, Thomas Dolby, Sue Green, Chantel Cox, Xiaosong Yang

Keywords: Component Process Model (CPM), Dynamic emotion state, Multi-agent prompting, Persona-based dialogue, Affective appraisal, LLM-as-judge, Human preference evaluation, Synthetic role simulation, Construct validity, Reproducibility

Source: Open primary source (opens in a new tab)

9
Authors
15
Findings
32
Limitations
6
Evidence

Editorial summary

English

This preprint proposes CPM-MultiAgent, a prompt orchestrator that maintains an explicit representation of a fictional persona's emotion during dialogue. A trigger analyzer extracts the event and context; four agents assess relevance, implications, coping potential, and normative significance; peer review consolidates those appraisals; and an integrator changes each emotion by -1, 0, or +1 within a 1-5 range. A critic can request a redo. Although the paper calls the state latent, operationally it is structured JSON visible to the agents, not a learned latent variable. The taxonomy is configurable; case studies use Plutchik's eight emotions and an intensity scale the paper associates with PANAS.

The main evaluation contains only 24 author-constructed synthetic trials in three families: a simulated patient for healthcare training, simulated student for school communication, and simulated customer for customer service. Three complete five-turn sequences are shown, but the 24 inputs, initial states, outputs, and judgments are not released. Every agent uses GPT-5.4 by default at temperature 0.2 and top-p 1.0; GPT-5.4 also judges at temperature 0. Baselines are zero-shot, few-shot, two CoT variants, self-consistency, Self-Refine, and an adapted EQ-Negotiator. No official code or data repository was found. The TeX claims baseline prompts are supplied, but its disabled section contains seven literal “Insert the ... prompt” comments; few-shot examples, self-consistency sampling and aggregation, Self-Refine rounds, and the EQ-Negotiator adaptation are absent.

The authors acknowledge that no gold standard maps a trigger to a correct emotion transition. They substitute six GPT-5.4-judged Likert dimensions: update correctness, trigger grounding, temporal consistency, persona consistency, appraisal reasoning quality, and overall. CPM-MultiAgent scores 4.322/5 overall versus 4.311 for EQ-Negotiator, the strongest overall baseline: a 0.011-point difference without an interval, test, or clearly documented denominator. Persona consistency differs by 0.006. All means are near ceiling. Because GPT-5.4 both generates and judges headline results, anonymizing labels and shuffling order reduces identity leakage but does not remove self-preference, correlated rubric bias, or design-rubric alignment.

One reproducible inconsistency is material: the Appendix D.2 judge prompt requests EUC, grounding, temporal consistency, persona consistency, and overall, but omits Appraisal Reasoning Quality entirely. Table 1 and the ablations nevertheless report ARQ, including the method's 4.833. That column cannot be reproduced from the disclosed prompt. A second contradiction affects human evaluation: Section 4.3 says humans and the LLM use the same metrics and 1-5 scale, while the human form asks only two categorical pairwise preferences, emotion update and appraisal reasoning. These are neither the same metrics nor response scale.

The study uses 103 anonymous, voluntary, unpaid annotators and says it received IRB approval, without naming a board, protocol, demographics, or recruitment procedure. Against EQ-Negotiator, emotion update receives 41 votes for the method, 26 ties, and 36 for the baseline. Excluding ties and optimistically treating responses as independent, 41/67 yields an exact two-sided binomial p=0.086437: not distinguishable from chance even before multiplicity correction. Reasoning preferences are stronger, 67/24/12 against EQ, but the method is required to produce a long structured appraisal rationale; preference can measure clarity or presence of this structure rather than emotional correctness. No inter-rater reliability, intervals, power, assignment, scenario strata, or repeated-measures analysis is reported.

Ablations lower the same judge's overall score: 4.079 without peer review, 4.200 without normative significance, 4.226 without trigger analysis, and 4.262 without the critic. This shows that decomposition helps satisfy the benchmark's plausibility rubric, not that each CPM component has psychological construct validity. Robustness repeats the same 24 trials: with GPT-5.4, mini, nano, and Qwen3.6-35B-A3B, the multi-agent system scores above zero-shot and a monolithic variant, but without uncertainty or external domains. Reported parallel latency falls from 14,528 to 11,004 ms for GPT-5.4 text and 8,063 to 6,487 ms for mini; provider, region, hardware, repetitions, variance, and speech components are absent.

The defensible contribution is a software pattern for externalizing and reviewing a discrete emotional trajectory during role-play, with initial evidence that its structured explanation appears more plausible or preferable in synthetic scenarios. It does not validate the Component Process Model, an internal model emotion, or real human transitions. Integer steps and clipping mechanically favor smooth trajectories, the same property the judge rewards. Final persona responses, user outcomes, safety, clinical or educational benefit, and real deployment are not evaluated. The work therefore does not demonstrate psychologically correct or generalizable emotional simulation; it demonstrates rubric fit and presentation preference on a small, unreleased, difficult-to-reproduce artifact.

Español

Este preprint propone CPM-MultiAgent, un orquestador de prompts para mantener una representación explícita de la emoción de una persona ficticia durante un diálogo. Un analizador extrae el desencadenante y el contexto; cuatro agentes valoran relevancia, implicaciones, potencial de afrontamiento y significado normativo; una revisión entre agentes consolida esas valoraciones; y un integrador cambia cada emoción en -1, 0 o +1, limitado a una escala de 1 a 5. Un crítico puede pedir rehacer la actualización. Aunque el texto llama «latente» al estado, en la implementación es un vector JSON estructurado y visible para los agentes, no una variable latente aprendida. La taxonomía es configurable; los casos usan las ocho emociones de Plutchik y una escala que el artículo vincula a PANAS.

La evaluación principal consta de solo 24 pruebas sintéticas construidas por los autores en tres familias: paciente simulado para formación sanitaria, estudiante simulado para comunicación escolar y cliente simulado para atención al cliente. El artículo muestra tres secuencias completas de cinco turnos, pero no publica las 24 entradas, estados iniciales, salidas o juicios. Todos los agentes usan GPT-5.4 por defecto, con temperatura 0,2 y top-p 1,0; el juez usa GPT-5.4 a temperatura 0. Se comparan zero-shot, few-shot, dos variantes CoT, self-consistency, Self-Refine y una adaptación de EQ-Negotiator. No hay repositorio oficial de código o datos. El archivo TeX dice ofrecer los prompts de baseline, pero su sección desactivada contiene siete comentarios literales «Insert the ... prompt»; faltan ejemplos few-shot, número y agregación de self-consistency, rondas de Self-Refine y adaptación de EQ-Negotiator.

Los autores reconocen que no existe un gold estándar que asigne a cada desencadenante una transición emocional correcta. Sustituyen ese criterio por seis escalas Likert juzgadas por GPT-5.4: corrección de actualización, grounding del desencadenante, consistencia temporal, consistencia de persona, calidad del razonamiento de appraisal y global. CPM-MultiAgent obtiene 4,322/5 global frente a 4,311 de EQ-Negotiator, el baseline global más fuerte: una diferencia de 0,011 sin intervalo, test o denominador claramente documentado. En consistencia de persona la diferencia es 0,006. Todas las medias están cerca del techo. Como el mismo GPT-5.4 genera y juzga los resultados, ocultar nombres y barajar el orden reduce filtración de identidad, pero no elimina autopreferencia, sesgo correlacionado ni alineación entre el diseño y la rúbrica.

Hay una inconsistencia reproducible importante: el prompt de juez impreso en Appendix D.2 solo solicita EUC, grounding, consistencia temporal, consistencia de persona y global; omite por completo Appraisal Reasoning Quality. Sin embargo, la Tabla 1 y las ablaciones publican ARQ, incluido el 4,833 del método. Esa columna no puede derivarse del prompt divulgado. Otra contradicción afecta la evaluación humana: Section 4.3 afirma que humanos y LLM usan el mismo conjunto de métricas y escala 1-5, pero el formulario humano pide solo dos preferencias pareadas categóricas, actualización emocional y razonamiento. No es la misma métrica ni la misma escala.

Participan 103 anotadores anónimos, voluntarios y no remunerados. El artículo dice contar con aprobación IRB, sin identificar comité, protocolo, demografía o reclutamiento. Frente a EQ-Negotiator, la actualización emocional recibe 41 preferencias por el método, 26 empates y 36 por el baseline. Si se excluyen empates y, de forma optimista, se tratan las respuestas como independientes, 41/67 da p binomial exacta bilateral=0,086437: no se distingue del azar ni antes de corregir multiplicidad. Las preferencias de razonamiento son más fuertes, 67/24/12 frente a EQ, pero el método obliga a producir una justificación de appraisal extensa y estructurada; la preferencia puede medir claridad o presencia de esa estructura, no que el estado emocional sea correcto. No se reportan fiabilidad interjuez, intervalos, potencia, asignación, estratos por escenario o análisis de medidas repetidas.

Las ablaciones bajan la puntuación global del mismo juez: sin peer review, 4,079; sin significado normativo, 4,200; sin desencadenante, 4,226; sin crítico, 4,262. Esto indica que la descomposición ayuda a satisfacer la rúbrica de plausibilidad en este banco, no que cada componente CPM tenga validez psicológica. La robustez repite las mismas 24 pruebas: con GPT-5.4, mini, nano y Qwen3.6-35B-A3B, el multiagente puntúa por encima de zero-shot y de una variante monolítica, pero sin incertidumbre ni dominios externos. La latencia paralela reportada baja de 14.528 a 11.004 ms para texto con GPT-5.4 y de 8.063 a 6.487 ms con mini; no se publican proveedor, región, hardware, repeticiones, varianza ni componentes de voz.

La contribución defendible es un patrón de software para externalizar y revisar una trayectoria emocional discreta durante role-play, junto con evidencia inicial de que su explicación estructurada resulta más plausible o preferible en escenarios sintéticos. No valida el Component Process Model, una emoción interna del modelo ni transiciones humanas reales. La regla de pasos enteros y el clipping predisponen trayectorias suaves, precisamente una propiedad premiada por el juez. Tampoco se evalúan las respuestas finales del personaje, resultados de usuarios, seguridad, beneficio clínico o educativo ni despliegue real. Por ello el trabajo no demuestra una simulación emocional psicológicamente correcta o generalizable; demuestra ajuste a una rúbrica y preferencia de presentación bajo un artefacto pequeño, no publicado y difícil de reproducir.

Research question

Can a CPM-inspired multiagent architecture produce more plausible, persona-consistent, and temporally coherent emotional state updates than monolithic prompts in synthetic role-play dialogues?

Method

Three-stage LangGraph pipeline: trigger analysis; parallel appraisal of relevance, implications, coping, and normative meaning with peer review; integration of -1/0/+1 changes and critique. Compares seven baselines using GPT-5.4 judge, ablations, four backbones, 103 human preferences, and three five-turn cases.

Sample: Twenty-four synthetic tests in healthcare, school communication, and customer service; three visible five-turn sequences. Human evaluation with 103 anonymous, voluntary, unpaid annotators; each row sums 103 preferences.

Findings

  • CPM-MultiAgent obtains 4.322/5 global with the GPT-5.4 judge.
  • EQ-Negotiator obtains 4.311; the global difference is 0.011 points.
  • The persona consistency difference against EQ-Negotiator is 0.006.
  • Automatic means are close to the ceiling of the scale.
  • The disclosed judge prompt omits ARQ although the tables report that dimension.
  • The human evaluation does not use the same set of metrics or the 1-5 scale stated in the method.
  • Against EQ-Negotiator, emotional update receives 41 votes, 26 ties, and 36 votes for the baseline.
  • The exploratory binomial test without ties gives p=0.086437 for update against EQ-Negotiator.
  • Against EQ-Negotiator, appraisal reasoning receives 67 votes, 24 ties, and 12 for the baseline.
  • Removing peer review produces the largest ablation drop, from 4.322 to 4.079.
  • The multiagent outperforms zero-shot and the monolithic across the four reported backbones.
  • The robustness tests reuse the same 24 synthetic tests.
  • Parallelization reduces reported latency, but the protocol is not reproducible.
  • The state called latent is an explicit vector shared among agents.
  • The evidence supports judged plausibility and explanation preference, not human emotional correctness.

Limitations

  • Preprint v1 without verified peer review.
  • Only 24 synthetic tests constructed by the authors.
  • Only three scenario families and three complete sequences published.
  • The 24 inputs, initial states, outputs, or judgments are not published.
  • No official code or data repository was found.
  • Baseline prompts are placeholders in the TeX.
  • Few-shot examples and configuration of self-consistency, Self-Refine, and EQ-Negotiator are missing.
  • There is no gold standard of emotional transition.
  • Evaluation dimensions have no reported psychometric validation.
  • The expert or process that reviewed the metrics from psychology is not identified.
  • GPT-5.4 generates and judges the main results.
  • Automatic scores are close to the ceiling.
  • Intervals, tests, or exact denominator of judgments are not reported.
  • ARQ is missing from the published judge prompt.
  • The human evaluation description contradicts the actual form.
  • The 103 participants lack demographics and recruitment procedure.
  • IRB approval does not include identifiable committee or protocol.
  • There is no inter-rater reliability, intervals, power, or multiplicity adjustment.
  • The dependence of repeated measures by annotator is not modeled.
  • The assignment of examples and scenarios to annotators is not documented.
  • Preference for structured reasonings may reflect length and format.
  • The -1/0/+1 rule and clipping predefine temporal smoothness.
  • The judge rewards properties induced by the architecture.
  • Ablations use the same judge and do not test construct validity.
  • Robustness does not extend scenarios, domains, or population.
  • Exact versions of models, provider, or dependencies are not reported.
  • Seeds, token limits, retries, failures, and costs are missing.
  • Latency lacks hardware, region, repetitions, and variance.
  • ASR/TTS are not identified in the voice condition.
  • Final character responses are not evaluated.
  • Real users or clinical, educational, or commercial outcomes are not evaluated.
  • There is no specific safety or deployment harm evaluation.

What the study does not establish

  • That the LLM experiences emotions or has an internal affective state
  • That the explicit vector is a learned latent representation
  • That the four prompts validate the psychological components of CPM
  • That a judged plausible transition is emotionally correct
  • That GPT-5.4 scores constitute a gold standard
  • That the method improves final character responses
  • That the simulation represents real patients, students, or clients
  • That clinical, educational, or customer service benefit exists
  • That the update difference against EQ-Negotiator is conclusive
  • That the preferred explanations are causally faithful to the model
  • That the results generalize beyond 24 synthetic tests
  • That the system is reproducible without prompts, data, outputs, and code
  • That the system is validated for safe deployment

Traceability

Scope: Full text

Version: arXiv:2607.07824v1; 23-page preprint; PDF and TeX source, every PDF page, publication status, prompts, tables, human evaluation and numerical claims audited 2026-07-16

Consulted source: https://arxiv.org/abs/2607.07824v1

Review: Codex 23-page full-text visual, arXiv TeX, prompt-artifact, evaluation, statistical and claim-boundary audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • GPT-5.4 generator and headline judge; exact API snapshot not reported
  • GPT-5.4-mini
  • GPT-5.4-nano
  • Qwen3.6-35B-A3B

Instruments and metrics

  • Component Process Model-inspired prompt decomposition
  • Trigger Analyzer
  • Relevance, implications, coping-potential and normative-significance appraisal agents
  • Peer review and critic agents
  • Explicit 1-5 emotion-state vector with -1/0/+1 updates
  • Plutchik eight-emotion taxonomy in case studies
  • PANAS-associated intensity scale
  • Six-dimension GPT-5.4 LLM-as-judge rubric
  • Two-question pairwise human preference evaluation
  • LangGraph

Data used

  • 24 author-constructed synthetic persona-dialogue trials across three scenario families
  • Three published five-turn case trajectories
  • No released full trial set, outputs, judge logs, code or baseline prompts

Evidence and location

  • Method, results, appendices, forms, and inconsistencies: arXiv:2607.07824v1 PDF, 23 pages; every page rendered and visually inspected
  • Identity, version, date, and categories: Official arXiv record and API for 2607.07824v1
  • Absent baseline prompts and artifact content: arXiv TeX source archive sha256:36e2fac5f8acc6b92eaec8bf590196f70f229778e0c50e07ab80e2ef5a2af407
  • Editorial status as preprint and absence of official repo: Exact-title arXiv, OpenReview, ACL Anthology and GitHub searches checked 2026-07-16
  • Statistical audit of preferences: Exact two-sided binomial calculation from Table 2 counts, standard-library audit 2026-07-16
  • Consolidated audit: reports/verification/article-279-cpm-multiagent-emotion-state-gold-standard-human-evaluation-prompt-artifact-and-claim-audit.json