A Dynamic Multi-Agent Game Framework for Live Streaming E-Commerce Real-Time Comment Driven Virtual Anchor Discourse and Strategy Generation

Virtual anchors in live-streaming e-commerce often suffer from delayed responses to real-time comments, rigid preset scripts, and insufficient dynamic strategic reasoning. To address these limitations, this paper proposes a dynamic multi-agent game framework for real-time comment-driven virtual anchor discourse and strategy generation. First, a viewer state space is constructed by encoding bullet-screen comments through fine-grained sentiment and intent analysis. Second, a three-party dynamic Bayesian game is introduced among the virtual anchor, audience group, and competitor environment, and a deep Q-network is used to solve equilibrium policy labels in real time. Third, these strategy labels constrain a large language model to generate controllable, context-aware, and persona-consistent scripts. Finally, a dual-model verification mechanism enforces compliance and virtual-anchor identity consistency. Experimental results show that the proposed method increases conversion rate from 3.82% to 4.53% and extends average dwell time from 187.4 s to 238.3 s. The game-win rate reaches 64.9%, and script adaptation accuracy reaches 87.2%, confirming the effectiveness of game-theoretic modeling in real-time e-commerce interaction.