ST-GFN model cuts negotiation inequality by 43.8% with dual-stream fusion

Forecasting outcomes in mixed-motive negotiations remains challenging because AI models must integrate explicit linguistic cues—like offers and threats—with latent strategic constraints such as budgets and alternatives. Existing approaches often fail to adapt across different task structures and can perpetuate historical biases in utility distributions. To address this, Moirangthem Tiken Singh proposes ST-GFN (Semantic-Temporal Graph Fusion Network), a unified dual-stream architecture. One stream uses transformer encoders to process natural language dialogue; the other employs Graph Attention Networks to model economic states and constraints. A dynamic gated fusion mechanism adaptively weighs each modality based on the negotiation setting. The model also incorporates a fairness-regularized composite loss that penalizes deviations from ground-truth utility gaps, explicitly reducing inequality in predictions.

Evaluated on two contrasting benchmarks—the linguistically oriented DealOrNoDeal and the strategy-oriented CaSiNo—ST-GFN demonstrates strong adaptability. In free-form, language-heavy scenarios (DealOrNoDeal), the model assigns higher weight to linguistic cues (attention coefficient z ≈ 0.97), while switching to rely more heavily on strategic constraints in structured tasks (CaSiNo, z ≈ 0.73). The key results show a 43.8% reduction in Inequality Discrepancy in high-disparity environments, with minimal impact on overall forecast accuracy. Moreover, performance improved in high-variance domains, suggesting that reflective regularization can enhance both predictive reliability and equitable representation. These findings support the development of transparent Group Decision and Negotiation Support Systems (GDNSS), making AI-aided negotiation fairer and more adaptable.