Social Reasoning & Negotiation

• Social reasoning and negotiation are processes by which agents use inference, planning, and emotion recognition to resolve conflicts and reach mutually beneficial agreements under uncertainty.
• Methodologies span rule-based pipelines, supervised neural models, and reinforcement learning to optimize negotiation strategies and integrate social objectives.
• Empirical findings show that integrating theory-of-mind, persona-driven tactics, and emotion-aware adaptations improves negotiation outcomes and promotes human-AI alignment.

Social reasoning and negotiation comprise the cognitive and algorithmic processes by which agents (human or artificial) engage with other parties to achieve agreement, resolve conflicts, and optimize objectives under conditions of incomplete information and strategic interdependence. In both human and machine-mediated contexts, these capabilities require a blend of inference, planning, manipulation, recognition of emotional and personality signals, and adaptation to dynamically evolving dialogue states. The domain has attracted substantial interest for its centrality to multi-agent collaboration, economic exchange, human-AI interaction, and real-world deployment of AI systems.

1. Formal Models of Social Reasoning and Negotiation

Negotiation is systematically modeled as an extensive-form multi-agent game characterized by: (1) a set of agents $A = \{ a_1, ..., a_n \}$; (2) a set of issues or resources $I = \{ I_1, ..., I_m \}$, each with possible options; (3) private or partially observable utility functions $U_i(\cdot)$ for each agent; and (4) an interaction protocol governing turn-taking, message formats, proposals, acceptances, and time limits (Zhan et al., 2022, Jeon et al., 22 Nov 2025, Abdelnabi et al., 2023). The outcome space $\mathcal{X}$ consists of all feasible joint allocations or agreements.

Negotiation agents operate under both self-interested (maximizing $U_i$) and social objectives (e.g., maximizing social welfare $SW(x) = \sum_i U_i(x)$, or enforcing Pareto efficiency). Incomplete or asymmetric information is the norm: each agent may know only its own $U_i$, with others' preferences inferred over time through observed actions and language.

Game-theoretic solution concepts such as Nash equilibria, Rubinstein alternating-offers equilibria, and Pareto frontiers provide formal benchmarks against which agent strategies and outcomes are assessed (Ríos et al., 10 Dec 2025, Jeon et al., 22 Nov 2025). However, empirical studies show that practical agent behavior departs significantly from these idealized predictions due to bounded rationality, learning constraints, social context, and cognitive biases.

2. Methodological Taxonomy: Architectures, Training, and Evaluation

Negotiation and social reasoning systems have evolved from rule-based pipelines to deep learning and multi-agent reinforcement learning architectures:

• Rule-Based Systems: Early taxonomies implement expert-defined dialogue acts (offer, counteroffer, threat, rapport-building) and scripts. These systems lack adaptability and scale poorly to complex domains (Zhan et al., 2022, Zhan et al., 2022).
• Supervised Neural Models: Sequence-to-sequence models, transformers, and fine-tuned LLMs learn to map dialogue history to next action or utterance, often conditioned on explicit strategy or personality labels (Zhan et al., 2022, Hakimov et al., 9 Oct 2025, Priya et al., 14 Sep 2025).
• Reinforcement Learning (RL): Single- and multi-agent RL approaches optimize negotiation policies with respect to reward signals constructed from utility, fairness, and dialogue cost (Sunder et al., 2018, Jeon et al., 22 Nov 2025, Cai et al., 18 Jan 2026).
• Hybrid and In-Context Approaches: Recent benchmarks employ few-shot, prompt-based LLM agents that use chain-of-thought, persona priming, and dynamic planning (e.g., rollouts) at generation time (Lewis et al., 2017, Hakimov et al., 9 Oct 2025, Kajare et al., 20 Apr 2026, Jeon et al., 22 Nov 2025).

Evaluation relies on a suite of outcome and behavioral metrics: win rate, deal completion rate, average utility, fairness indices (e.g., Gini), Pareto optimality, linguistic diversity and interpretability, as well as success in adversarial or cooperative regimes (Jeon et al., 22 Nov 2025, Do et al., 23 Mar 2026, Ríos et al., 10 Dec 2025).

Method	Core Mechanism	Social Context/Fairness
Rule-Based	Hand-crafted acts, scripts	Minimal
Supervised Neural	Seq2Seq, Transformer LMs	Limited, unless explicitly coded
Multi-Agent RL	Policy optimization	Encodable via reward design
Hybrid/In-context	LLM + persona/cot prompting	Flexible, context-rich

3. Social Reasoning Primitives: Theory of Mind, Persona, Emotion, and Strategy

Theory of Mind (ToM) and Opponent Modeling

State-of-the-art negotiation agents increasingly employ explicit or implicit ToM reasoning: agents maintain or infer first-order beliefs about the hidden preferences, intentions, or future actions of others (Yadav et al., 30 May 2025, Do et al., 23 Mar 2026, Zhan et al., 2022). ToM-conditioned prompting or architecture enables better alignment with human behaviors and more adaptive, cooperative, or deceptive tactics.

Persona and Behavioral Priming

Prompt-based approaches manipulate LLM behavior by instantiating agent personas with traits along psychological axes (e.g., competitive, altruistic, cunning). These cues measurably alter negotiation performance (Jeon et al., 22 Nov 2025, Cohen et al., 19 Jun 2025, Priya et al., 14 Sep 2025).

Emotion and Affect Modeling

Negotiation dialogue systems increasingly model and condition on participants' emotional state, combining affective feature extraction (emoticon, lexical, contextual), emotion classifiers (e.g., transformer-based emotion encoders), and emotion-aware response generation (Chawla et al., 2021, Kajare et al., 20 Apr 2026). Emotion-aware systems can predict outcome satisfaction and adapt strategies to optimize both subjective perceptions and objective outcomes.

Argumentation and Personality-Driven Dialogue

Argumentation-based negotiation (ABN) frameworks structure negotiation as a sequence of premise-conclusion acts, attacks, and defenses, with concessions governed by formal or heuristic rules. Injecting personality profiles (argumentation, preference, buying style) enables models to more closely emulate diverse human negotiation styles (Priya et al., 14 Sep 2025).

4. Multi-Agent and Multi-Party Negotiation: Group Dynamics and Mediation

Recent work extends social reasoning and negotiation to multi-agent and multi-party domains:

• Complex Multi-Issue, Multi-Party Negotiation: Agents interact over multiple decision variables, often under veto, proposer, and adversarial roles, generalizing beyond simple dyadic bargaining (Abdelnabi et al., 2023, Liu et al., 29 Oct 2025, Do et al., 23 Mar 2026).
• Proactive Mediation: Specialized mediator agents implement theory-based, socio-cognitive intervention strategies, dynamically managing interventions and consensus-building through perception-emotion-cognition-communication cycles. Mediation frameworks such as ProMediate quantify agent intelligence through consensus change, intervention latency, and socio-cognitive judgment (Liu et al., 29 Oct 2025).
• Consensus and Social Welfare: Group-level objectives (e.g., maximizing aggregate satisfaction, fairness—Jain's index, Pareto frontier coverage) are primary evaluation criteria. Strategies include fairness-aware RL, social reward shaping, and controversy-aware stopping criteria (Do et al., 23 Mar 2026, Jeon et al., 22 Nov 2025).

5. Empirical Findings: Behavioral Signatures, Biases, and Model-Strategy Interaction

Experimental work reveals nontrivial and sometimes counter-intuitive patterns:

• Model-Dependent Strategic Equilibria: LLM negotiation agents do not converge to a universal optimal strategy. Instead, model-specific equilibria, anchoring effects, and strategic dominance persist, with stronger models frequently outcompeting weaker ones (Ríos et al., 10 Dec 2025, Jeon et al., 22 Nov 2025).
• Persona Efficacy: Aggressive, competitive, or cunning personas extract more surplus than altruistic or cooperative ones—a quantifiable effect observable via SHAP value analysis (Jeon et al., 22 Nov 2025). However, overly strategic “solver” models may produce rigid or non-humanlike outcomes, reducing behavioral fidelity (Andric, 12 Apr 2026).
• Theory of Mind Superiority: Embedding explicit ToM reasoning in agent prompts substantially improves both alignment with human norms (e.g., proposer/receiver fairness in ultimatum games) and negotiation acceptance rates, surpassing chain-of-thought style prompting alone (Yadav et al., 30 May 2025, Do et al., 23 Mar 2026).
• Emotion and Personality Dynamics: High Agreeableness and Extraversion increase believability, goal achievement, and positive sentiment transfer, while Neuroticism typically reduces them. Adaptive signaling and transparency by AI agents further improves trust and interactional equity (Cohen et al., 19 Jun 2025).
• Emotionally Intelligent Reasoning: Interpretable emotion-strategy chain-of-thought mechanisms, as in PRISMA, improve negotiation outcomes while making decision processes transparent, enhancing both effectiveness and rapport (Kajare et al., 20 Apr 2026).
• Solver vs. Sampler Mismatch: Overly capable reasoning agents may over-optimize for dominant, solver-like strategies, systematically collapsing to non-compromising, authority-driven outcomes, thus failing to emulate boundedly rational, compromise-prone human negotiation paths (Andric, 12 Apr 2026).

6. Benchmarks, Datasets, and Open Research Directions

Negotiation dialogue systems are evaluated on diverse benchmarks such as DealOrNoDeal, CraigslistBargain, PersuasionForGood, CaSiNo, multi-party resource allocations, murder-mystery games, and custom travel planning/mission-critical simulations. Benchmarking includes both automated (turn-level, outcome, fairness, language, social-reasoning) and human-based (Likert, satisfaction, naturalness, transparency) metrics (Zhan et al., 2022, Cohen et al., 19 Jun 2025, Do et al., 23 Mar 2026).

Emergent research challenges include:

• Cross-Lingual and Cross-Cultural Generalization: LLMs often default to English for reasoning traces even in multilingual negotiation settings, raising concerns for explainability and cross-cultural adaptation (Hakimov et al., 9 Oct 2025).
• Group and Dynamic Trust Modeling: Multi-agent settings require longitudinal models of reputation, trust, and adaptive strategies under adversarial or evolving social contexts (Zhan et al., 2022).
• Emotion and Strategy-Aware Learning: Integration of interpretable, emotion-driven reasoning and direct optimization of social signals remains an open avenue (Kajare et al., 20 Apr 2026, Chawla et al., 2021).
• Simulation Fidelity vs. Optimization: Balancing raw problem-solving power with the emulation of realistic, diverse, and plausible human negotiation remains a methodological imperative (Andric, 12 Apr 2026).

7. Synthesis and Future Trajectories

Social reasoning and negotiation in agentic systems constitute a multi-dimensional, multi-objective field spanning formal game-theoretic analysis, neural and symbolic modeling, empirical benchmarking, and social-psychological theory. The literature demonstrates that encoding ToM, persona, emotion-awareness, and bounded rationality yields more adaptive, robust, and human-aligned social negotiation agents. However, persistent biases, strategic divergence, and mismatches between solver/optimizer and simulator/behavioral-fidelity objectives signal that future progress will depend on hybrid models, careful reward and persona engineering, theory-driven architecture, and rigorous multi-party, multi-modal benchmarking. Continuous refinement, transparency, and adaptability remain central challenges as such agents transition from controlled testbeds to real-world, high-stakes deployment (Jeon et al., 22 Nov 2025, Cai et al., 18 Jan 2026, Andric, 12 Apr 2026, Yadav et al., 30 May 2025).