Andrii Bidochko

• Updated: July 14, 2026
• 7 min read

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

Direct Answer

The paper introduces CPM‑MultiAgent, a multi‑agent framework grounded in the Component Process Model (CPM) that enables a dialogue persona’s emotional state to evolve dynamically in response to conversational triggers. This matters because it moves emotional AI from static, trait‑based representations to a process‑oriented, continuously adapting system, unlocking more authentic role‑play in healthcare, education, and customer‑service bots.

Background: Why This Problem Is Hard

Persona‑based dialogue agents have become a cornerstone of modern AI products, from virtual therapists to interactive storytellers. Yet most existing systems treat emotion as a fixed attribute—either a preset mood label or a superficial stylistic cue such as “cheerful language.” This simplification creates two critical bottlenecks:

• Static emotional profiles: When a user’s narrative shifts—e.g., a patient discloses worsening symptoms—the agent’s emotional response often remains unchanged, breaking immersion and reducing trust.
• Lack of self‑modeling: Current affective dialogue research focuses on empathetic responses toward the user, ignoring how the agent’s own affective state should evolve as the story unfolds.

These gaps are especially problematic in high‑stakes domains where emotional consistency influences outcomes: a counseling bot that fails to show growing concern may appear uncaring; an educational tutor that never shows frustration may seem unrealistic. Traditional approaches—rule‑based sentiment toggles, static emotion embeddings, or single‑turn affect classifiers—cannot capture the nuanced, multi‑turn appraisal process that humans experience.

What the Researchers Propose

To bridge the gap, the authors propose CPM‑MultiAgent, a framework that operationalizes the Component Process Model—a psychological theory describing emotion as a sequence of appraisals of external events. The framework consists of three cooperating agents:

1. Trigger Extraction Agent: Scans each user utterance for affective cues (e.g., “I’m scared,” “That’s wonderful”) and extracts structured “triggers” that can influence the persona’s appraisal.
2. Collaborative Appraisal Agent: Implements CPM’s appraisal dimensions (novelty, goal relevance, coping potential, etc.) to evaluate how each trigger should shift the latent emotional state.
3. Emotion State Update Agent: Maintains a continuous latent representation of the persona’s emotion and updates it after every turn based on the appraisal outcomes.

By treating emotion as a mutable latent variable rather than a static label, CPM‑MultiAgent enables a persona to exhibit realistic emotional trajectories—e.g., escalating anxiety after repeated negative news or gradual relief after successful problem solving.

How It Works in Practice

The operational workflow can be visualized as a loop that repeats for each dialogue turn:

Step 1 – User Input → Trigger Extraction
The Trigger Extraction Agent parses the user’s message, identifies affect‑laden phrases, and encodes them as a set of trigger vectors.

Step 2 – Triggers + Current State → Collaborative Appraisal
The appraisal agent receives both the trigger set and the persona’s current latent emotion vector. It runs CPM‑inspired computations across appraisal dimensions, producing an “appraisal delta” that quantifies how each dimension should shift.

Step 3 – Delta → Emotion State Update
The update agent integrates the appraisal delta with the existing latent state using a gated recurrent mechanism, yielding a refreshed emotion vector that reflects the cumulative impact of the conversation.

Step 4 – Generation → Response
The refreshed emotion vector conditions the downstream language model (e.g., GPT‑4, LLaMA) that generates the persona’s next utterance, ensuring the response aligns with the updated affective stance.

What distinguishes this approach from prior work is the explicit, theory‑driven appraisal step. Instead of a black‑box sentiment classifier that directly flips a mood flag, CPM‑MultiAgent decomposes the emotional impact into interpretable dimensions, allowing developers to inspect, debug, and even manually adjust appraisal weights for domain‑specific tuning.

{{IMAGE_URL}}

Evaluation & Results

The authors validated CPM‑MultiAgent across three benchmark settings that mimic real‑world role‑play scenarios:

• Healthcare counseling simulation: A virtual therapist interacts with patients describing varying symptom trajectories.
• Educational tutoring dialogue: A tutor guides a student through a problem set while reacting to frustration or confidence cues.
• Interactive storytelling: A narrative NPC responds to player choices that trigger emotional highs and lows.

For each setting, they compared CPM‑MultiAgent against two baselines:

1. Static‑Emotion Baseline: Persona emotion is fixed at the start of the conversation.
2. Sentiment‑Flip Baseline: A simple sentiment classifier toggles a “happy/sad” flag based on the latest user sentiment.

Key findings include:

• Human Likert Scores: In blind evaluations, participants rated CPM‑MultiAgent responses as significantly more emotionally coherent (average 4.3/5) compared to the static baseline (2.9/5) and sentiment‑flip baseline (3.5/5).
• Consistency Over Turns: CPM‑MultiAgent maintained a logical emotional arc across 10‑turn dialogues, whereas baselines often produced abrupt or contradictory mood shifts.
• Ablation Insights: Removing the collaborative appraisal component dropped coherence scores by 0.8 points, confirming its central role.

These results demonstrate that a CPM‑grounded multi‑agent architecture can reliably model dynamic emotional evolution, outperforming simpler sentiment‑based methods both quantitatively and qualitatively.

Why This Matters for AI Systems and Agents

Dynamic emotion modeling unlocks several practical advantages for developers and product teams:

• Richer User Experience: Agents that adapt their affective stance create deeper engagement, essential for therapeutic bots, learning companions, and immersive games.
• Fine‑Grained Control: Because appraisal dimensions are explicit, system architects can calibrate how strongly a trigger influences novelty, coping potential, or goal relevance, tailoring agents to industry‑specific emotional policies.
• Improved Evaluation Metrics: Traditional chatbot metrics (BLEU, ROUGE) ignore affective consistency. CPM‑MultiAgent provides a built‑in diagnostic (appraisal delta) that can be logged and monitored, facilitating continuous quality assurance.
• Seamless Integration with Existing Pipelines: The framework is modular; each agent can be swapped for a custom implementation (e.g., using UBOS platform overview for orchestration) without redesigning the entire dialogue stack.

For enterprises building large‑scale conversational products, the ability to simulate authentic emotional trajectories reduces the need for manual scriptwriting and enables more scalable personalization.

What Comes Next

While CPM‑MultiAgent marks a substantial step forward, several open challenges remain:

• Cross‑Cultural Appraisal: CPM’s appraisal dimensions were derived from Western psychological literature. Extending the model to respect cultural variations in emotional expression will require multilingual trigger extraction and culturally aware appraisal weights.
• Long‑Term Memory Integration: Current experiments span up to ten turns. Incorporating episodic memory so that emotions persist across sessions could further enhance realism for long‑term tutoring or therapy bots.
• Real‑World Deployment Constraints: Latency introduced by multi‑agent coordination must be optimized for production environments. Leveraging edge‑compute or model quantization (e.g., via Ollama) could mitigate performance bottlenecks.
• Ethical Guardrails: Dynamically evolving emotions raise questions about manipulation and user consent. Future work should embed transparency mechanisms that disclose an agent’s affective state to end‑users.

Potential application avenues include:

• Embedding CPM‑MultiAgent in AI marketing agents to simulate authentic customer sentiment during sales simulations.
• Integrating with Enterprise AI platform by UBOS for large‑scale employee wellness chatbots that adapt to workplace stress signals.
• Coupling with voice synthesis tools like ElevenLabs AI voice integration to deliver emotionally congruent spoken responses.

Overall, the framework opens a research frontier where affective psychology meets scalable AI engineering, promising agents that not only understand users but also evolve emotionally in lockstep with the conversation.

References

Cai, J., Liu, S., Rehman, A., Guo, Y., Tian, Q., Dolby, T., Green, S., Cox, C., & Yang, X. (2026). From Triggers to Emotions: A CPM‑Grounded Appraisal Multi‑Agent for Dynamic Emotional Evolution in Persona‑Based Dialogue. arXiv preprint arXiv:2607.07824.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.