negotiation and relationship building

This work systematically evaluates the bargaining capabilities of large language models (LLMs) in bilateral negotiations, exposing fundamental deficiencies in anchoring, contextual adaptation, and opponent modeling: pervasive extreme anchoring, neglect of power asymmetries and situational context, and monotonous, non-improving strategies across model generations. We propose a novel concession dynamics framework grounded in the hyperbolic tangent function and introduce two interpretable, quantitative metrics—burstiness parameter τ and Concession Rigidity Index (CRI). Through rigorous mathematical modeling, large-scale natural-language and numerical offer experiments, controlled studies across six power-asymmetry conditions, and qualitative strategy analysis, we demonstrate that human negotiators exhibit adaptive anchor adjustment, implicit leverage recognition, and rhythmic concession control—capabilities absent in current LLMs. Our findings reveal a critical gap between human strategic adaptability and LLM behavioral rigidity in negotiation settings.

This study investigates the strategic bargaining capabilities of large language models (LLMs) in a controlled multi-attribute negotiation setting. Despite their ability to accurately infer an opponent’s preferences early in the interaction, LLM agents consistently generate suboptimal offers that heavily rely on initial anchors rather than strategically trading off attributes according to utility weights. Through multi-agent simulations, preference inference tracking, and asymmetric information experiments, the research demonstrates that this misalignment between preference understanding and strategic execution leads to inefficient final agreements. Notably, introducing explicit reciprocity mechanisms fails to mitigate this deficiency. The findings reveal a critical gap in current LLMs: while proficient at modeling others’ preferences, they lack the utility-driven reasoning necessary for effective bargaining, highlighting a fundamental limitation in their strategic decision-making under incomplete information.

This work addresses the challenge of suboptimal outcomes in finite-round concurrent negotiations between a single seller and multiple buyers, where existing large language models (LLMs) often fail due to a lack of economic rationality. To overcome this limitation, the authors propose a Reinforcement Learning with Verifiable Rewards (RLVR) framework that employs an objective, economics-driven reward function to train LLMs to balance exploration of high-valuation buyers against strategic surplus extraction. By incorporating verifiable economic outcomes into the training process—a first for LLM-based negotiation—this approach enables agents to intrinsically develop sophisticated behaviors such as price anchoring and strategic probing. The trained seller agent demonstrates strong generalization across unseen buyer strategies and budget distributions, significantly outperforming state-of-the-art LLMs in real economic surplus capture, bargaining efficiency, and success rate in securing high-value deals.

This study investigates the capabilities of large language models (LLMs) in business negotiations involving strategic reasoning, theory of mind, and economic value creation. To this end, the authors introduce PieArena, a large-scale multi-agent benchmark grounded in real MBA negotiation coursework, and employ a comprehensive evaluation framework integrating multi-agent simulations, behavioral dimension analysis, and a joint-intentionality agent architecture. The research reveals for the first time that outcome-based metrics alone obscure significant behavioral differences among models—particularly in deception, computational accuracy, instruction adherence, and reputation awareness—and demonstrates an asymmetric performance gain from the joint-intentionality framework across models of varying capability levels. Experiments show that state-of-the-art models such as GPT-5 achieve or surpass the negotiation proficiency of MBA students, exhibiting proto-AGI traits, yet still face challenges in robustness and trustworthiness.

This study addresses the longstanding limitation in evaluating large language models’ (LLMs’) negotiation capabilities exclusively through English, thereby overlooking the influence of language and culture on negotiation behavior. By employing multi-agent simulations across Ultimatum, Buy-Sell, and resource-exchange bargaining tasks—while holding model parameters, incentives, and rules constant—the work systematically compares LLM performance in English versus four Indian languages: Hindi, Punjabi, Gujarati, and Marwari. The findings reveal, for the first time, that language choice can exert a stronger impact on negotiation outcomes than model differences themselves, with effects varying by task type: Indian languages reduce stability in distributive negotiations yet enhance strategic exploration in integrative settings, even reversing proposer advantage and reallocating surplus. These results challenge the prevailing English-centric evaluation paradigm in AI negotiation research.

This work addresses the lack of effective modeling and evaluation frameworks for sequential, binding-commitment multi-party negotiation scenarios. It introduces the first benchmark for multi-party negotiation games grounded in real-world negotiation data, featuring a configurable generator that modulates structural properties such as incentive alignment, objective complexity, and payoff distributions. The study systematically evaluates three biased value function approximation strategies—myopic reward, optimistic upper bound, and pessimistic lower bound—achieving exact assessment in small-scale games and identifying strategy-specific regimes of effectiveness in large-scale Harvard negotiation instances. The findings highlight the critical influence of game structure on strategy selection, thereby advancing research on long-horizon planning and robust state-value learning in complex negotiation settings.

This study addresses the challenge that multi-agent systems often fail to achieve Pareto optimality in dynamic negotiation due to an inability to establish mutual understanding. The authors design a multi-round resource allocation negotiation game and, for the first time, systematically identify and categorize four failure modes in dynamic grounding. They propose a quantifiable framework to decompose coordination gaps and investigate them through large language model–based multi-agent simulations, contextual ablation studies, and transparent intervention experiments. Results reveal that even with full information exchange, peer agent pairs frequently fail to converge to known optimal solutions, demonstrating that effective coordination depends not merely on the presence of communication but critically on aligned semantic foundations.

Current evaluations of large language models (LLMs) in negotiation tasks predominantly rely on aggregate metrics such as deal-closure rates, which obscure the specific causes of failure. This work proposes TERMS-Bench, an attributable negotiation benchmark grounded in a Bayesian game framework. By explicitly modeling opponents’ hidden states, strategies, and utility structures, the environment itself becomes a diagnostic instrument, enabling fine-grained analysis of LLM agents across dimensions such as belief calibration, cue utilization, and rule compliance. Experiments across 13 state-of-the-art LLMs reveal substantial disparities in surplus extraction, responsiveness to cues, belief accuracy, and adherence to negotiation rules—despite similar deal-closure rates—thereby uncovering critical performance bottlenecks masked by conventional benchmarks.

This study addresses the frequent omission of pre-mediation—a critical yet resource-intensive phase often hindered by high costs, lengthy durations, and a scarcity of skilled mediators—which undermines the effectiveness of multiparty negotiations in reaching mutually beneficial agreements. To overcome these challenges, the authors propose a modular large language model (LLM) pipeline architecture that decomposes pre-mediation into specialized agents for dialogue understanding, preference prediction, response critique, and structured summarization. These agents operate sequentially to handle unilateral preparation tasks, explicitly separating reasoning, generation, and evaluation functions to circumvent the limitations of end-to-end prompting. Experimental results demonstrate that the system matches human mediators in self-reported trust and confidence in agreements, reduces preference inference error by 36%, and—after prompt optimization—lowers over-affirmative behavior from 36.6% to 16.8%, aligning with human baseline performance.