Simulation of Language Evolution under Regulated Social Media Platforms: A Synergistic Approach of Large Language Models and Genetic Algorithms

A new multi-agent framework uses large language models to simulate how social media users evolve language strategies to evade platform regulations, according to research posted on arXiv. [1] The framework, detailed in a paper by Jinyu Cai and colleagues, pairs participant agents acting as social media users with supervisory agents that emulate platform-level content moderation. [1] The participant agents continuously adapt their language expression, while the supervisory agents assess policy violations. [1] To better mirror real-world dynamics, the researchers employed a dual design of language strategies, separating constraint and expression to differentiate conflicting goals. [1] The simulation leverages an LLM-driven Genetic Algorithm to directly manipulate textual language strategies through selection, mutation, and crossover. [4] Selection represents the adoption of effective expressions, mutation reflects the emergence of novel covert expressions under regulatory pressure, and crossover simulates the combination of different strategies to enhance adaptability. [4] The framework was evaluated across two distinct scenarios: an abstract password game and a realistic simulated illicit pet trade. [2] Experimental results showed that as the number of dialogue rounds increased, both the number of uninterrupted dialogue turns and the accuracy of information transmission improved significantly. [2] The researchers also conducted a user study with 40 participants, which validated the real-world relevance of the generated dialogues and strategies. [2] Ablation studies further underscored the importance of the Genetic Algorithm, emphasizing its contribution to long-term adaptability and improved overall results. [2] The initial version of the paper was submitted on 26 February 2025, with a revised version posted on 18 June 2026. [1]