EmoMAS: Emotion-Aware Multi-Agent System for High-Stakes Edge-Deployable Negotiation with Bayesian Orchestration

Abstract: LLMs has been widely used for automated negotiation, but their high computational cost and privacy risks limit deployment in privacy-sensitive, on-device settings such as mobile assistants or rescue robots. Small LLMs (SLMs) offer a viable alternative, yet struggle with the complex emotional dynamics of high-stakes negotiation. We introduces EmoMAS, a Bayesian multi-agent framework that transforms emotional decision-making from reactive to strategic. EmoMAS leverages a Bayesian orchestrator to coordinate three specialized agents: game-theoretic, reinforcement learning, and psychological coherence models. The system fuses their real-time insights to optimize emotional state transitions while continuously updating agent reliability based on negotiation feedback. This mixture-of-agents architecture enables online strategy learning without pre-training. We further introduce four high-stakes, edge-deployable negotiation benchmarks across debt, healthcare, emergency response, and educational domains. Through extensive agent-to-agent simulations across all benchmarks, both SLMs and LLMs equipped with EmoMAS consistently surpass all baseline models in negotiation performance while balancing ethical behavior. These results show that strategic emotional intelligence is also the key driver of negotiation success. By treating emotional expression as a strategic variable within a Bayesian multi-agent optimization framework, EmoMAS establishes a new paradigm for effective, private, and adaptive negotiation AI suitable for high-stakes edge deployment.

• The paper introduces a Bayesian orchestrated multi-agent framework that integrates game theory, reinforcement learning, and psychological reasoning for strategic, emotion-aware negotiation.
• It demonstrates robust performance across high-stakes domains such as debt collection, medical scheduling, disaster rescue, and student sleep alerting, outperforming established baselines.
• The study offers actionable insights for deploying privacy-preserving edge negotiation systems that maintain ethical constraints and dynamic emotional calibration.

Bayesian Emotion-Aware Multi-Agent Negotiation for Edge Deployment

Motivation and Context

Automated negotiation has become a pivotal application domain for LLMs, especially in scenarios requiring emotional intelligence and strategic reasoning under privacy constraints. However, cloud-based LLMs expose sensitive negotiation content—such as financial, medical, and educational information—to external risks, impeding deployment in data-critical edge environments. SLMs, optimized for local and real-time applications, display pronounced deficits in emotional sophistication and resilience to adversarial emotional manipulation. Static agent architectures, brittle persona embeddings, and pre-trained strategy reliance collectively hinder adaptive negotiation, especially in high-stakes domains with volatile emotional dynamics and sparse interaction feedback.

Framework: Architecture and Bayesian Orchestration

EmoMAS introduces a MoA framework with a Bayesian orchestrator for strategic emotional negotiation. The system comprises three specialized agents:

• Game Theory Agent: Implements WSLS emotional strategies, leveraging a payoff matrix reflecting interaction valence. This agent maintains cooperation for positive emotional exchanges, shifts responses for negative, and avoids escalation risk.
• Reinforcement Learning Agent: Employs tabular Q-learning to adaptively optimize emotional states via contextual features (current emotions, negotiation phase, gap category). This enables in situ online trajectory adjustment without requiring neural pre-training.
• Emotional Coherence Agent: Uses LLM-based psychological reasoning for emotion selection. The agent incorporates plausibility, appropriateness, strategic value, and rationale scores using structured context vectors and LLM-guided weighting.

The Bayesian Orchestrator operates as a meta-reasoner, continuously updating agent reliability weights via Bayesian inference based on negotiation feedback (macro-level: post-trajectory; micro-level: turn-by-turn agreement). The orchestrator fuses agent predictions, enforcing interpretability and domain expertise constraints, and selecting emotional states that optimize both trajectory coherence and final rewards.

Figure 1: The workflow of the EmoMAS framework, depicting agent specializations and Bayesian orchestration for emotion trajectory selection.

Scenario-Driven Evaluation and Dataset Design

EmoMAS was evaluated on four synthetic negotiation benchmarks, each representing distinct high-stakes, emotionally charged domains:

• Debt Collection (CRAD): Negotiator must distinguish genuine emotional appeals from fraud and optimize financial collection.
• Medical Scheduling (SSD): Agent balances patient urgency and resource allocation under surgical stress.
• Disaster Rescue (DESRD): Robot negotiators provide real-time psychological support and resource allocation, constrained by edge deployment.
• Student Sleep Alerting (SSAD): Educational robots negotiate with adolescents on bedtime anxiety and deadline pressures.

These datasets were structured to vary emotional intensity, adversarial tactics, and negotiation objectives, providing robust cross-domain evaluation.

Quantitative Results and Robustness Analysis

EmoMAS consistently outperformed all baselines and agent components in success rates and negotiation outcome metrics across domains. Notably:

• Debt Collection: EmoMAS-Bayes achieved 100% success (both LLM and SLM settings), with superior negotiation outcomes, outperforming game-theory, RL, and coherence agents.
• Medical and Emergency Scenarios: EmoMAS-Bayes maintained the highest success under adversarial strategies (pressuring, victim-playing, threatening), with robust outcomes despite opponent manipulation. Under severe emotional pressure, EmoMAS-Bayes showed only moderate performance decline, demonstrating resilience.
• Edge Deployment: SLMs equipped with EmoMAS matched or exceeded LLM performance in disaster rescue scenarios, achieving 100% success with authentic emotional calibration. Single-agent baselines exhibited architecture-dependent bias, whereas EmoMAS delivered stable performance across model types.

Ethical metrics indicated that EmoMAS-Bayes struck a favorable balance between instruction following, emotional consistency, and low manipulation rates; SLMs tended to be more manipulative than LLMs, reinforcing the orchestrator’s role in ethical constraint.

Prompt Engineering and Modular Adaptation

The framework integrates modular prompts for emotion detection, negotiator/opponent strategies, and value extraction, enabling flexible adaptation across scenarios (Figure 2, Figure 3). Scenario-specific prompts facilitate domain-relevant emotional guidance, adversarial tactic simulation, and consistent value assessment, enhancing framework generality and practical deployment.

Figure 2: Real-time LLM prompt for emotion detection during negotiation.

Figure 3: Scenario-centric prompt for edge-deployed debt negotiation agent.

Practical and Theoretical Implications

EmoMAS demonstrates that strategic emotional intelligence in negotiation is optimally achieved via dynamic, Bayesian multi-agent orchestration rather than static persona aggregation or reward-driven adaptation alone. The framework establishes that agent reliability and emotional trajectory optimization must be continually learned, rather than fixed at pre-training or averaged across experts. Practically, EmoMAS enables privacy-preserving, robust, and adaptive negotiation for mobile assistants, institutional edge systems, and embodied robotics in bandwidth-constrained environments. Theoretically, it affirms the necessity of treating emotional expression as a strategic variable in sequential optimization, directly targeting dialogue-level reward and coherence.

Future Outlook

Further extensions should incorporate continuous and blended emotion representations, multimodal negotiation inputs, and empirical evaluation in cross-cultural and human-in-the-loop settings. The orchestrator’s interpretability and transparency must be augmented for deployment in regulated domains, and its mechanism generalized for real-world data scarcity and adversarial robustness. The modular design and prompt flexibility suggest applicability to hierarchical multi-agent coordination and broader emotional planning tasks.

Conclusion

EmoMAS offers a principled architecture for strategic, emotion-aware negotiation in high-stakes, privacy-sensitive edge environments. The Bayesian orchestrator enables real-time fusion of game theoretic, RL, and psychological models, achieving stable, robust performance across domains, agents, and adversarial strategies. The results establish EmoMAS as a paradigm for effective agentic negotiation with dynamic emotional adaptation and ethical constraint, serving as a modular foundation for future embodied, multimodal, and cross-cultural negotiation AI (2604.07003).