Skip to content

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

🕒 Published (v1): 2026-07-08 18:06 UTC · Source: Arxiv · link

Why this paper was selected

CPM-grounded multi-agent emotional appraisal; advances persona dialogue agents for healthcare/counseling

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

CPM-MultiAgent is a multi-agent LLM framework that models trigger-driven emotional state evolution in persona-based dialogue by operationalizing the Component Process Model (CPM) from psychology. Rather than treating a character's emotion as a static attribute, the framework continuously updates a latent multi-dimensional emotion state through structured appraisal of dialogue triggers. Experiments across three role-simulation scenarios show consistent gains over prompting and agent-based baselines.

Problem

Existing persona-based dialogue systems encode emotion as a fixed trait or prompt-level control signal, while affective dialogue research focuses on empathizing with the user rather than tracking the agent character's own evolving internal emotional state. This leaves trigger-driven, turn-level emotional evolution of a simulated persona largely unaddressed.

Method

CPM-MultiAgent decomposes emotion evolution into three sequential layers instantiated as a multi-agent pipeline (implemented in LangGraph):

  1. Trigger Analysis Layer — A Trigger Analyzer Agent parses the current dialogue turn \((P, S, M, H_t, E_{t-1}, T_t)\) into a structured analysis pack \(A_t = \{O_t, G_t, C_t\}\): an objective trigger structure, a persona-grounded context, and a CPM stimulus-evaluation-check (SEC) cue set organized into four groups (relevance, implication, coping potential, normative significance).

  2. CPM Emotion Appraisal Layer — Four specialized agents evaluate the trigger along CPM's four appraisal dimensions in a dynamic, recursive order where later checks condition on earlier ones. After all four produce initial appraisals \(X_t = \{R_t, I_t, K_t, N_t\}\), a Peer Review step allows agents to revise each other's outputs, yielding \(\hat{X}_t\). An Integration Agent then computes a per-emotion change vector \(\Delta_t\) and clips updated intensities: \(e_{t,k} = \text{clip}(e_{t-1,k} + \Delta_{t,k})\).

  3. Consistency Critique Layer — A Critic Agent audits CPM faithfulness, contextual grounding, and temporal coherence. If the verdict \(v_t \neq \text{approve}\), critique rationale \(b_t\) is fed back to the flagged layer for revision, bounded by a maximum revision budget \(R\).

The calibrated emotion state \(E_t\) is then used as an intermediate affective representation for response generation.

Key Contributions

  • CPM-MultiAgent framework: a novel architecture for character-internal, trigger-driven emotional state dynamics in persona-based dialogue, replacing static emotion control with latent state evolution.
  • Psychological grounding: first operationalization of CPM's four appraisal checks (relevance, implication, coping potential, normative significance) as distinct LLM agents with structured SEC cues, linking dialogue triggers to emotional transitions in a principled way.
  • Comprehensive evaluation: LLM-as-judge evaluation (GPT-5.4), blind pairwise human evaluation (103 annotators), ablation over 7 component variants, and robustness testing across 4 backbone models (GPT-5.4, GPT-5.4-mini, GPT-5.4-nano, Qwen3.6-35B-A3B).

Results

  • LLM-as-judge (6 Likert-5 metrics; Table 1): CPM-MultiAgent achieves the highest score on all metrics; largest gains on Appraisal Reasoning Quality (4.833 vs. best baseline 4.778) and Trigger Grounding (4.189 vs. 4.172), with Overall 4.322 vs. 4.372 for EQ-Negotiator being the closest competitor on Overall.
  • Human pairwise preference (Table 2): CPM-MultiAgent wins on Appraisal Reasoning in 67–87 out of 103 comparisons across all baselines; wins on Emotion Update in 41–55 comparisons.
  • Ablation (Table 3): Removing any single component degrades all metrics; removing Peer Review causes the largest drops (Overall 4.079 vs. 4.322 full).
  • Robustness (Table 4): CPM-MultiAgent outperforms zero-shot and monolithic CPM prompting on every backbone tested; gains persist even on weaker models (GPT-5.4-nano: 4.277 vs. 4.083 zero-shot).

Limitations

  • Evaluation is conducted on only 24 hand-constructed dialogue trials across three scenarios, which is a small sample with limited ecological validity.
  • No gold-standard trigger-to-emotion transition labels exist; evaluation relies on LLM-as-judge (GPT-5.4, the same model used as backbone) and human preference ratings, both of which may have systematic biases.
  • The framework adds significant inference overhead (7+ sequential agent calls per turn plus optional revision loops), making real-time deployment costly.
  • Scenarios are limited to healthcare training, educational communication, and customer service; generalizability to open-domain or adversarial settings is untested.
  • Backbone models (GPT-5.4, Qwen3.6) are very recent and not publicly reproducible, limiting independent benchmarking.

Relevance to Agentic AI / LLM Agents

This paper directly advances the design of multi-agent systems where specialized sub-agents divide a complex reasoning task—here, emotional appraisal—and coordinate through structured information passing and peer review, a pattern increasingly central to agentic LLM architectures. The Critic Agent with bounded revision loops is an instance of self-correction/reflection mechanisms studied broadly in agentic pipelines. For researchers tracking LLM agents in simulation and role-play contexts (e.g., social simulation, NPC behavior, training environments), CPM-MultiAgent offers a psychologically grounded template for endowing agents with internally consistent, dynamically evolving mental states rather than stateless response policies. The peer-review and critique mechanisms also contribute to the growing literature on agent self-consistency and inter-agent verification.