Normatively guided graph diffusion for efficient and rational utility generation in bilateral bargaining

Autonomous AI agents operating in negotiation and resource-allocation environments must generate utility allocations that are individually rational, Pareto-efficient, and axiomatically fair, while remaining robust to the biased training data that characterizes real-world human negotiation corpora. Existing generative approaches reproduce empirical bargaining behavior without enforcing cooperative game-theoretic principles during inference, whereas classical optimization methods require exact analytical Pareto frontier specification unavailable for empirical data. This study introduces a normatively guided graph diffusion framework for efficient and individually rational utility generation in bilateral negotiation. Negotiation states are encoded by a Graph Attention Network (GATv2) conditioning a Denoising Diffusion Implicit Model (DDIM). A differentiable guidance mechanism penalizes Nash-product shortfalls and individual-rationality violations during a late-stage activation window, preserving generative diversity while enforcing axiomatic compliance. A theoretical approximation bound establishes how the activation window, guidance strength, and denoiser accuracy jointly control a worst-case distance from the Nash Bargaining Solution. Evaluation across a synthetic non-transferable utility corpus and two human negotiation datasets demonstrates that Nash Efficiency reaches 0.9913 with perfect individual-rationality compliance on the synthetic corpus, and improves from 0.9249 to 0.9921 on Deal-or-No-Deal. On CaSiNo, surplus recovery improves, though a geometric limitation of spherical frontier approximations on discrete allocation manifolds is identified and diagnosed mechanistically.