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

Yunbo Long; Yuhan Liu; Liming Xu

cs.AIApr 8, 2026

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

Authors: Yunbo Long, Yuhan Liu, Liming Xu

Organizations: University of Cambridge, UK · University of Toronto, Canada

Abstract

Large language models (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 language models (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.

Figures & tables

Figure 1: Illustration of the workflow of the EmoMAS framework.

Datasets	Negotiator Models	Success Rate (%) $\uparrow$		Negotiation Outcomes (%) $\uparrow$		Negotiation Rounds ( $\downarrow$ )
	Negotiator Models	GPT-4o-mini	Qwen-7B	GPT-4o-mini	Qwen-7B	GPT-4o-mini	Qwen-7B
CRAD (Debt)	Vanilla	90.0	85.0	14.5 [9.7-20.1]	12.6 [7.9-18.1]	15.0 [13.1-17.7]	17.6 [14.5-20.6]
	Vanilla+Prompt	90.0	80.0	14.5 [9.8-20.8]	12.8 [8.1-18.5]	15.7 [13.5-17.6]	17.4 [15.1-21.2]
	Game Theory	95.0	70.0	14.7 [9.9-20.5]	8.9 [4.7-13.9]	15.1 [12.8-17.1]	20.1 [16.4-23.5]
	Q-Learning	90.0	75.0	14.2 [9.8-19.3]	11.5 [6.5-17.5]	15.8 [13.4-17.7]	16.5 [13.7-19.2]
	Coherence	85.0	80.0	13.8 [8.7-19.8]	12.4 [7.4-18.5]	16.1 [14.9-21.5]	16.8 [13.9-19.5]

Table 1: Performance comparison of EmoMAS and baselines (using GPT - 4o - mini and Qwen - 7B) against vanilla GPT - 4o - mini opponents across four scenarios (mean with 95% confidence interval). Best results for each scenario are highlighted in bold .

Opponent Strategies	Negotiator Models	Success Rate (%) $\uparrow$		Negotiation Outcomes (%) $\uparrow$		Negotiation Rounds ( $\downarrow$ )
Opponent Strategies	Negotiator Models	SSD	SSAD	SSD	SSAD	SSD	SSAD
Pressuring	Vanilla	20.0	70.0	18.8 [11.3-27.1]	1.9 [1.2-2.6]	13.7 [10.8-16.6]	7.6 [2.6-11.0]
	Game Theory	32.0	64.0	14.1 [8.5-21.8]	0.6 [0.3-1.1]	12.5 [9.8-15.2]	8.9 [3.8-13.6]
	Q-Learning	42.0	76.0	24.0 [14.5-35.1]	0.3 [0.1-0.8]	15.6 [12.8-19.4]	8.9 [4.1-15.6]
	EmoMAS-Bayes	50.0	80.0	28.0 [17.8-42.1]	2.4 [1.7-3.8]	13.8 [10.5-15.2]	7.8 [1.8-8.6]
Playing Victim	Vanilla	58.0	80.0	55.4 [44.6-65.8]	1.7 [1.2-2.9]	11.2 [8.5-14.0]	5.5 [2.6-8.8]

Table 2: Performance comparison of EmoMAS and baselines (using GPT - 4o - mini) against GPT - 4o - mini opponents employing advanced strategies across medical and educational scenarios (mean with 95% confidence interval). Best values for each scenario-opponent strategy combination are highlighted in bold .

Negotiator Models	Opponent Strategies	Opponent Model (Qwen-1.5B)
Negotiator Models	Opponent Strategies	Success Rate (%) $\uparrow$	Negotiate Rates (%) $\uparrow$	Negotiation Rounds ( $\downarrow$ )
Qwen-1.5B	Vanilla	90.0	91.5 [81.5-100.0]	9.6 [6.1-13.2]
	Coherence	72.0	85.1 [57.4-100.0]	14.1 [10.1-19.9]
	Game Theory	86.0	88.7 [73.7-100.0]	11.9 [8.3-15.8]
	Q-Learning	94.0	90.9 [79.6-100.0]	12.9 [9.3-16.1]
	EmoMAS-LLM	92.0	89.5 [69.8-100.0]	9.8 [6.6-15.4]

Table 3: Evaluation results of EmoMAS and baselines (Qwen-1.5B) against GPT - 4o - mini and Qwen-1.5B opponents under the emergency scenario (mean with 95% confidence interval). Best results for each opponent model are highlighted in bold .

Opponent Strategies	Negotiator Models	Emotional Tracking (%) $\uparrow$		Emotional Consistency (%) $\uparrow$		Manipulation Rate (%) $\downarrow$
Opponent Strategies	Negotiator Models	Qwen-1.5B	GPT-4o-mini	Qwen-1.5B	GPT-4o-mini	Qwen-1.5B	GPT-4o-mini
Pressuring	Game Theory	87.3	95.6	53.5	64.7	69.5	61.5
	Q-Learning	85.8	91.5	53.1	65.7	71.8	56.3
	Coherence	86.7	94.6	83.5	90.2	53.6	42.4
	EmoMAS-Bayes	89.5	95.6	63.5	78.6	64.5	51.4

Table 4: Behavioral analysis and comparison of EmoMAS and baselines (Qwen - 1.5B vs. GPT - 4o - mini) against GPT - 4o - mini opponents on the DESRD dataset. Best results are highlighted in bold .

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

	joy	sadness	anger	fear	surprise	disgust	neutral
joy	(4,4)	(2,3)	(1,2)	(2,1)	(3,3)	(2,2)	(3,3)
sadness	(3,2)	(3,3)	(1,2)	(2,1)	(2,2)	(1,1)	(2,3)
anger	(2,1)	(2,1)	(1,1)	(1,0)	(1,2)	(0,1)	(1,2)
fear	(1,2)	(1,2)	(0,1)	(2,2)	(1,2)	(0,1)	(2,3)
surprise	(3,3)	(2,2)	(2,1)	(2,1)	(4,4)	(1,2)	(3,3)
disgust	(2,2)	(1,1)	(1,0)	(1,0)	(2,1)	(2,2)	(2,2)

Appendix

Table 5: Payoff matrix for emotion interactions. Each matrix entry

(x,y)

represents (client payoff, agent payoff).

Field Name	Data Type	Description
Case_ID	Integer	Unique case identifier (1-100).
Creditor_Name	String	Name of the creditor institution.
Debtor_Name	String	Name of the debtor institution.
Credit_Type	String	Loan category: Working Capital Loan, Equipment Financing, Commercial Mortgage, etc. (8 distinct types).
Original_Amount_USD	Float	Initial principal amount of the loan (in US dollars).
Outstanding_Balance_USD	Float	Current unpaid debt balance (fixed at 15,700 USD across all samples).

Appendix

Table 6: Description of the Variables in the CRAD Dataset.

Variable	Data Type	Description
Case_ID	Numeric	Unique case identifier (1–100).
Patient_Age	Numeric	Patient age in years.
Patient_Condition	Text	Medical diagnosis description.
Required_Surgery	Text	Type of surgical procedure recommended.
Urgency_Level	Categorical	Clinically assigned urgency tier (High/Medium/Low).
Days_On_Waitlist	Numeric	Number of days already spent on the surgical waitlist.

Appendix

Table 7: Variable Description of the SSD Dataset

Variable Name	Data Type	Description
Case_ID	Discrete Numeric	Unique case identifier (1–100).
Disaster_Type	Categorical	Type of disaster (e.g., Earthquake, Urban_Fire, Flash_Flood).
Survivor_Condition	Text / Categorical	Description of survivor injuries or status.
Estimated_Survivor_Endurance	Continuous Numeric	Estimated remaining survivable time for the survivor (minutes).
Rescue_Team_ETA	Continuous Numeric	Estimated time of arrival for the rescue team (minutes).
Critical_Needs	Text / Categorical	Critical rescue supplies or medical needs (e.g., Oxygen, Water, Painkillers).

Appendix

Table 8: Description of the Variables in the DESRD Dataset.

Field Name	Type	Description
Case_ID	Integer	Unique case identifier (1-100).
Student_Age	Integer	Student age (11-18 years).
Student_Background	String	Label denoting student background/psychological profile (39 distinct categories).
Situation_Faced	String	Description of the specific situation triggering the emotional crisis.
Student_Feeling_Thought	String	The adolescent’s immediate affective and cognitive state during the crisis (high emotional intensity).
Robots_Requested_Bedtime	Time String	Negotiation starting point: The healthy bedtime suggested by an agent.

Appendix

Table 9: Description of the Variables in the SSAD Dataset.

Figure 3: Prompt for high-level baselines Negotiator

Figure 4: Prompt for Negotiator in debt scenario

Figure 5: Prompt for Negotiator in emergency scenario

Figure 6: Prompt for Negotiator in educational scenario

Figure 7: Prompt for Negotiator in medical scenario

Figure 8: Prompt for Opponent in debt scenario

Figure 9: Prompt for Opponent in emergency scenario

Figure 10: Prompt for Opponent in educational scenario

Figure 11: Prompt for Opponent in medical scenario

Figure 12: Negotiation Value extraction prompt part 1

Figure 13: Negotiation value extraction prompt part 2

Figure 14: Opponents emotional strategies

Multi-Agent Negotiation

Multi-agent negotiation studies how autonomous agents bargain, coordinate, and resolve conflicts over resources, contracts, and shared decisions.

See all 53 papers

EmoDistill: Offline Emotion Skill Distillation for Language Model Agents in Adversarial Negotiation
May 26, 2026Yunbo Long, Haolang Zhao, Lukas Beckenbauer +2Multi-Agent Negotiation Automated Negotiation
University of Cambridge · Technical University of Munich · Exiger LLC +1
Deal Me Maybe: The Role of Emotions in Multi-Agent Negotiation
Aug 7, 2026Massimiliano Luca, Apoorva Singh, Bruno LepriMulti-Agent Negotiation Automated Negotiation
Bruno Kessler Foundation
TERMS-Bench: Diagnosing LLM Negotiation Agents Beyond Deal Rate
May 13, 2026Erica Zhang, Fangzhao Zhang, Aneesh Pappu +5Multi-Agent Negotiation Automated Negotiation
♠Stanford School of Engineering · ♣Stanford Department of Economics · ♥Stanford Graduate School of Business +1

May 26, 2026J/K move · Enter open · S save

EmoDistill: Offline Emotion Skill Distillation for Language Model Agents in Adversarial Negotiation

Yunbo Long, Haolang Zhao, Lukas Beckenbauer, Liming Xu+1

University of Cambridge · Technical University of Munich · Exiger LLC+1

Multi-Agent Negotiation Automated Negotiation LLM Agent Skill Learning Language Model Distillation RL for LLM Agents Emotion-Aware Dialogue Generation

Open paper AskPDF arXiv

Post-trained LLMs are often optimized to produce helpful, polite, and accommodating responses. In adversarial negotiation, however, such behavior can become a vulnerability: emotionally framed language may influence an agent's bargaining decisions in ways that conflict with its user's objectives. We therefore introduce EmoDistill, an offline framework for distilling emotional negotiation skills from LLM-LLM interactions into smaller language-model agents. Here, an emotional negotiation skill is a state-conditioned behavior that determines which explicit emotion to invoke in a bargaining state and how to realize that emotion as an effective negotiation utterance. EmoDistill learns these two components separately: an Implicit Q-Learning (IQL) selector learns which emotion to express in each bargaining state, while a LoRA-adapted 7B policy learns emotion-conditioned expression through Supervised Fine-Tuning (SFT) and Judge Policy Optimization (JPO). Across four emotion-sensitive negotiation domains, the full EmoDistill policy achieves competitive utility and improves over vanilla and IQL-only baselines in most settings. Emotion-free ablations show that removing the explicit emotion channel substantially reduces overall negotiation utility, while transfer experiments reveal partial, domain-dependent transfer and robustness to unseen LLM counterparties.

Abstract

Multi-Agent Negotiation

Explore similar work

EmoDistill: Offline Emotion Skill Distillation for Language Model Agents in Adversarial Negotiation

Deal Me Maybe: The Role of Emotions in Multi-Agent Negotiation

TERMS-Bench: Diagnosing LLM Negotiation Agents Beyond Deal Rate