Personalized Affective Perception via Continual Learning in Human-Agent Negotiation
Authors: Mehmet Onur Keskin, Berk Buzcu, Nikhil Churamani, Reyhan Aydogan, Hatice Gunes, Catholijn Jonker
Organizations: Dept. of Computer Science and Artificial Intelligence & Data Engineering, Özyeğin University, Istanbul, Türkiye. · Dept. of Computer Science, HES-SO Valais-Wallis, Sierre, Switzerland. · Dept. of Computer Science and Technology, University of Cambridge, Cambridge, UK. · Interactive Intelligence, Delft University of Technology, Delft, Netherlands.
Self-interested intelligent agents that negotiate with humans should interpret nonverbal expressions and adapt their behavior accordingly. However, generic facial-affect models degrade across individuals and contexts, creating a generalization gap. We introduce an affect-adaptive negotiation strategy that incorporates Continual Learning with Imagination for Facial Expression Recognition (CLIFER) to personalize valence-arousal perception in real time. Differential feedback converts changes in valence-arousal direction and intensity into adjustments of the agent's target utility. An asynchronous architecture updates CLIFER in the background during live interaction. In a counterbalanced within-subject study in which 66 participants negotiated with a humanoid robot, the CL-enabled agent achieved significantly higher individual utility and elicited higher participant concession rates than the FaceChannel baseline. The CL condition yielded significantly higher mean arousal estimates, alongside a broader output range. Participant utility and Nash distance did not change significantly between conditions. These outcomes support the practical value of coupling continual affective personalization with differential feedback for real-time negotiation.
Figures & tables
Fig. 1: Stepwise Move Classification [ 24 ] : Illustrated move types are defined by both parties’ utility changes from the same actor’s previous offer: Fortunate (both gain), Nice (opponent gains, own unchanged), Concession (own loss, opponent gain), Selfish (own gain, opponent loss), Unfortunate (both lose).
Fig. 2: System Architecture of the CL-enabled Solver Agent: The robot’s audiovisual input is converted to offers (Speech-to-Offer) and valence–arousal estimates (CLIFER: episodic memory adapted to the participant, CAAE providing generative replay). The Solver uses these to compute counter-offers, which the Interaction Manager voices with gestures. θ denotes the GWR memory parameters; the FaceChannel encoder and regression head remain fixed.
Fig. 3: Differential affective feedback between the previous ( E~t−1 , blue) and current ( E~t , orange) normalized VA vectors. Equation 7 defines angular ( Prot ) and radial ( Prad ) features; the dashed line shows displacement ( ΔE~ ).
Fig. 4: Human-Robot Negotiation Experimental Setup: (a) Participant’s view of the Nao robot and negotiation interface. (b) Side view of the room.
Fig. 5: Experiment Procedure: A 6-min Grocery demo initializes CLIFER. Participants then complete two Holiday negotiations (10-min deadline), one with each perception model in counterbalanced order, each preceded by preference elicitation and followed by a questionnaire.
Fig. 6: Utility Space of the Holiday Domain: Each point represents a possible agreement; stars mark the Pareto frontier and the triangle the Nash solution (0.80, 0.80). Both sessions use the same space.
Issue (Weight)
1
2
3
4
Destin. (0.48)
Barcelona (1) Venice (2)
Rome (1) Lisbon (2)
London (1) Sydney (2)
Boston (1) Miami (2)
Accom. (0.32)
AirBnB
Hotel
Hostel
Boat
Events (0.16)
Shopping
Museum
Sports
Show
Season (0.04)
Summer
Winter
Spring
Fall
TABLE I: Holiday Domain Preferences: The Issue Weights & Value Orders
Fig. 7: Affect Perception Analysis: (a-b) Affective Signal: CL produces significantly higher arousal estimates at both the subject ( dz=0.68 ) and round levels ( d=0.88 ). (c) Individual Differences: Most subjects ( 75.8% ) showed higher arousal values with the CL-enabled Solver Agent. (d) Temporal Profile: The offer-indexed means include initial zero entries and fewer contributing sessions at later positions. (e-g) Correlations: CL produces a positive arousal-valence correlation ( r=0.45 at round level; r=0.31 at subject level) compared with the negative association seen with FaceChannel at round and subject levels.
Metric
FC
CL
p
d
Subject-Level ( N=66 ; dz )
Arousal
0.02±0.07
0.14±0.16
<0.001
0.68
Valence
0.03±0.13
0.05±0.13
0.317
0.12
Round-Level ( nFC=1,134 , nCL=980 ; d )
Arousal
0.01±0.09
0.15±0.22
<0.001
0.88
Valence
0.03±0.18
0.07±0.20
<0.001
0.18
TABLE II: Estimated Affect Statistics
Fig. 8: Estimated Affect by Move Type: Round-level VA estimates for (a) FC and (b) CL, pooling human and robot moves; colored regions show density and large circles denote per-move centroids. CL centroids are higher in arousal across all move types.
Fig. 9: Participant Offer Moves: Per-participant proportions under FC and CL; lines connect each participant’s two sessions. Participants made a higher proportion of concession moves under CL (unadjusted * p<0.05 ).
Metric
FC
CL
p
dz
Concession
0.28±0.20
0.37±0.27
0.013
0.31
Nice
0.04±0.08
0.04±0.09
0.821
0.03
Fortunate
0.18±0.16
0.15±0.14
0.272
-0.14
Unfortunate
0.21±0.17
0.21±0.18
0.803
-0.03
Selfish
0.23±0.19
0.19±0.17
0.172
-0.17
Silent
0.06±0.14
0.04±0.11
0.390
-0.11
TABLE III: Participant Offer Moves
Move Type
FC Arousal
CL Arousal
p
d
Concession
0.011
0.173
<0.001
1.05
Selfish
0.011
0.187
<0.001
1.10
Nice
0.048
0.125
0.001
0.89
Fortunate
0.039
0.142
<0.001
0.58
Unfortunate
0.038
0.107
0.018
0.34
TABLE IV: Estimated Arousal by Move Type
Fig. 10: Negotiation Outcomes: Paired raincloud plots compare agent utility, participant utility, agreement rounds, and Nash distance. CL yielded significantly higher agent utility (unadjusted * p<0.05 ); the other differences were not statistically significant.
Metric
FC
CL
p
dz
Agent Utility
0.73±0.08
0.75±0.06
0.017
0.30
Participant Utility
0.76±0.09
0.75±0.10
0.495
-0.08
Nash Distance
0.12±0.08
0.11±0.08
0.378
-0.11
Agree. Rounds
8.56±5.05
7.61±3.85
0.134
-0.19
TABLE V: Negotiation Outcomes
Metric
Session 1: CL
Session 2: FC
Session 1: FC
Session 2: CL
Est. Arousal
0.17±0.16
0.00±0.05
0.03±0.07
0.11±0.15
Concession Rate
38.8±27.1
22.1±18.7
34.5±18.9
36.0±27.8
Agent Utility
0.75±0.07
0.72±0.09
0.74±0.06
0.75±0.06
Rounds
7.56±3.85
10.09±5.47
6.94±4.06
7.66±3.92
TABLE VI: Outcomes by Session and Sequence
Fig. 11: Questionnaire Responses: Ratings on a 5-point Likert scale for Q1–Q5 (Table VII ); white diamonds mark means. CL received higher ratings on Q1–Q3 (unadjusted * p<0.05 , exploratory).
Questions
FC
CL
p
dz
Q1) Preference Modeling: Solver modelled my preferences accurately.
3.33±0.77
3.65±0.71
0.013
0.32
Q2) Affective Awareness: Solver considered my emotional state.
3.09±1.06
3.44±0.88
0.015
0.30
Q3) Persuasiveness: Solver was convincing during the negotiation.
3.32±0.93
3.65±0.79
0.043
0.26
Q4) Participant Decisions: Solver’s gestures affected my offer behavior.
3.29±1.30
3.50±1.23
0.058
0.21
Q5) Time Awareness: While Solver was making its offers, it considered the time.
Affective computing develops systems that recognize, model, and respond to human emotions and affective states from behavior and physiological signals.
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.
Yunbo Long, Haolang Zhao, Lukas Beckenbauer +2
University of Cambridge · Technical University of Munich · Exiger LLC +1
According to canonical negotiation theory, people's success in a negotiation depends on how well they balance competing demands--empathizing and asserting, demonstrating concern for other and concern for self, being soft on the people and hard on the problem. Yet people struggle to manage these tensions, so researchers have lacked the ability to rigorously test the field's prescriptions under controlled conditions. AI agents do not face the same limitations, and their precision, repertoire, consistency, and scalability enable a new class of experiments to contribute to negotiation theory. In this article, we introduce personality engineering: a methodology that uses AI agents to precisely parameterize, manipulate, and evaluate negotiator personality. We propose using the interpersonal circumplex--and its two core dimensions of warmth and dominance--as a foundational coordinate system for the field. This approach offers both a rigorous methodology for testing classic negotiation theories and a practical guide for designing the personalities of AI negotiation agents.
Michelle A. Vaccaro, Jared R. Curhan
MIT Institute for Data, Systems, and Society · MIT Sloan School of Management
Negotiation is a demanding social task for LLM agents, requiring strategic reasoning, persuasion, and interpersonal adaptation. Yet existing benchmarks often treat agents as emotionally neutral, overlooking a key driver of human bargaining behavior. We study how prompt-conditioned emotions affect LLM-based price negotiation. In a controlled framework, buyer and seller agents are independently assigned one of six emotional states and negotiate over 350 real consumer products under two budget conditions. Across 36 emotion-pair settings and five widely used LLMs, we find that emotions strongly shape outcomes. Angry buyers almost never reach agreement (0.39% deal rate), while happy buyers agree most often (28.91%), but obtain worse prices than fearful buyers. Emotion effects are role-dependent: buyer emotion mainly drives acceptance and rejection, whereas seller emotion shapes concession dynamics. These effects influence not only language, but also termination behavior and price trajectories, raising concerns for emotion-conditioned agents in commerce.