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

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 harness that models dynamic emotional evolution in persona-based dialogue by decomposing the psychological Component Process Model (CPM) into a pipeline of specialized agents: trigger analysis, four sequential appraisal agents, an integration agent, and a critic. Emotion is represented as a latent intensity vector over Plutchik's eight primary emotions, updated each turn via \(e_{t,k} = \text{clip}(e_{t-1,k} + \Delta_{t,k})\). The framework outperforms prompting-based and agent-based baselines on all six evaluation dimensions in both LLM-as-judge and human pairwise evaluation.

Problem

Persona-based dialogue systems treat a character's emotion as a static attribute or prompt-level control signal, while affective dialogue research focuses on generating empathetic responses to the user's emotion rather than modeling how the agent character's own internal emotional state evolves in response to dialogue triggers. No existing approach provides psychologically grounded, trigger-driven, multi-turn emotional state updates for simulated personas across general role-playing scenarios.

Method

The framework implements a three-stage sequential pipeline executed once per dialogue turn \(t\):

Stage 1 — Trigger Analysis. A Trigger Analyzer Agent maps the current turn context \(\{P, S, M, H_t, E_{t-1}, T_t\}\) into a structured analysis pack \(A_t = \{O_t, G_t, C_t\}\), where \(O_t\) captures objective event structure, \(G_t\) grounds it in the persona, and \(C_t\) organizes Stimulus Evaluation Checks (SECs) into four CPM groups (relevance, implication, coping, normative).

Stage 2 — CPM Appraisal. Four specialized agents process the SEC groups in cascade order, each conditioning on prior results: Relevance → Implication → Coping Potential → Normative Significance, producing \(\{R_t, I_t, K_t, N_t\}\). A peer review step allows later agents to challenge earlier appraisals (\(\tilde{X}_t = \text{PeerReview}(X_t, O_t, G_t, E_{t-1})\)). An Integration Agent synthesizes \(\tilde{X}_t\) into emotion change vector \(\Delta_t\), updating each emotion label intensity by \(e_{t,k} = \text{clip}(e_{t-1,k} + \Delta_{t,k})\).

Stage 3 — Consistency Critique. A Critic Agent checks CPM faithfulness, contextual grounding, appraisal coherence, and emotion-update validity. If rejected, it issues a targeted revision directive \(b_t\) back to either the trigger or appraisal layer, bounded by \(R\) revision rounds (Algorithm 1). The calibrated \(E_t\) then conditions response generation.

Implementation uses LangGraph with GPT-5.4 as backbone (\(T=0.2\), top-p=1.0); evaluation judge uses \(T=0\) for stability.

Key Contributions

• CPM-MultiAgent framework: first multi-agent system grounding trigger-driven emotional evolution in the full CPM appraisal theory for persona-based dialogue
• Novel operationalization of CPM's four appraisal checks (relevance, implication, coping potential, normative significance) as distinct LLM agents with defined SEC-structured interfaces
• Six-dimensional evaluation protocol (Emotional Update Correctness, Trigger Grounding, Temporal Consistency, Persona Consistency, Appraisal Reasoning Quality, Overall) validated with 103 human annotators under blind pairwise comparison
• Robustness study across four backbone models (GPT-5.4, GPT-5.4-mini, GPT-5.4-nano, Qwen3.6-35B-A3B) showing structured decomposition benefits persist independent of model capacity

Results

• LLM-as-judge (Table 1): CPM-MultiAgent achieves best scores on all six metrics vs. seven baselines; ARQ 4.833 vs. best baseline EQ-Negotiator 4.778; Overall 4.322 vs. EQ-Negotiator 4.372 (note: EQ-Negotiator scores higher on TC 4.311 but lower overall)
• Human pairwise (Table 2): Wins 41–55 / 103 comparisons on emotion update quality against each baseline (losses: 14–26); wins 67–87 / 103 on appraisal reasoning quality (losses: 7–24); strongest advantage is on appraisal reasoning
• Ablation (Table 3): Removing peer review causes largest single-component drop (Overall 4.079 vs. 4.322); removing Trigger Analysis second (4.226); removing Critic Layer (4.262); all four individual appraisal check removals degrade performance with varying metric-specific impact
• Robustness (Table 4): CPM-MultiAgent consistently outperforms both zero-shot and monolithic CPM-aware prompting on every backbone; gains persist even for weaker models (GPT-5.4-nano: 4.277 vs. zero-shot 4.083)

Limitations

• Evaluation conducted on only 24 synthetic dialogue trials; no large-scale naturalistic dataset with ground-truth emotional trajectories
• No gold-standard emotion transition labels exist; evaluation relies entirely on Likert-scale rubrics judged by GPT-5.4 and human annotators, introducing subjectivity
• Framework incurs significant per-turn inference cost (7+ sequential agent calls plus revision loops), not benchmarked for latency or cost
• Peer review and critique mechanisms are bounded by fixed \(R\) without dynamic exit criteria
• Backbone GPT-5.4 is not a publicly released model at the time of the paper's writing, limiting immediate reproducibility

Relevance to Harnesses / Meta-Harnesses

CPM-MultiAgent is a domain-specific agent harness that implements a rigidly staged, psychology-grounded pipeline over specialized sub-agents with well-defined input/output schemas — a clear instance of harness design as task decomposition. The Consistency Critique Layer with bounded revision budget \(R\) is a harness-internal self-correction loop, analogous to critic-in-the-loop patterns seen in code-generation and reasoning harnesses. The clean separation between orchestration logic (LangGraph pipeline), backbone model, and domain knowledge (CPM theory) exemplifies how harnesses decouple task structure from model capacity, evidenced by the robustness experiments. The peer review mechanism between appraisal agents represents an intra-harness coordination pattern — an early example of structured multi-agent deliberation within a fixed pipeline rather than emergent tool use.