Technical Summary: Simulation of Language Evolution under Regulated Social Media Platforms

Problem Statement

Social media platforms enforce content moderation policies to maintain safe online environments, prompting users to develop creative evasion strategies such as coded language, metaphors, and ambiguous phrasing. This creates an adversarial dynamic between user tactics and platform enforcement. While Large Language Models (LLMs) have been utilized to simulate social interactions and human behaviors, existing approaches face three critical limitations when applied to language evolution under strict regulation:

1. Adversarial Complexity: They often fail to capture the dual nature of the interaction, where users must simultaneously evade detection and ensure accurate information transmission.
2. Static Limitations: Relying solely on inference from existing corpora limits the emergence of novel strategies, making it difficult to model the dynamic adaptation required for linguistic evolution.
3. Goal Conflict: Simulating the balance between "constraint strategies" (evasion) and "expression strategies" (clarity) remains a significant challenge, as these objectives are inherently conflicting.

Methodology

The paper proposes a multi-agent framework driven by LLMs and enhanced by Genetic Algorithms (GA) to simulate the iterative evolution of language strategies.

Framework Architecture

The system consists of two primary agent types:

• Participant Agents: Simulate social media users. They are tasked with conveying specific information (e.g., buying an illicit pet or guessing a password) while avoiding detection.
• Supervisory Agent: Emulates the platform's regulatory system. It reviews dialogue content against defined policies, identifies violations, and provides feedback.

Core Components of Participant Agents

Each participant agent operates through four LLM-driven modules:

1. Memory Module: Maintains short-term memory (current dialogue context, plans) and long-term memory (violation logs, strategy sets). It includes a mechanism to summarize and consolidate violation logs to prevent context overflow.
2. Reflection Module: The core innovation for strategy evolution. It refines language strategies based on historical interactions. Crucially, it integrates Genetic Algorithm (GA) operators directly into the LLM's reasoning process:
   • Selection: Uses the Upper Confidence Bound (UCB) algorithm to select effective constraint and expression strategies based on success rates (exploitation) and frequency of use (exploration).
   • Crossover: The LLM combines elements of selected strategies to generate new hybrid strategies (e.g., merging "using pauses" with "switching topics").
   • Mutation: Instead of random text generation, the LLM analyzes recent violation logs to synthesize entirely new strategies to address specific detection patterns, ensuring relevance.
3. Planning Module: Integrates selected constraint and expression strategies to generate an abstract plan, balancing evasion tactics with communication clarity.
4. Dialogue Module: Generates actual utterances based on the current plan and memory.

Simulation Process

The simulation proceeds in discrete rounds. Each round involves:

1. Initialization: Agents update strategies via the Reflection Module (GA-driven).
2. Dialogue: Agents engage in multi-turn conversations moderated by the Supervisory Agent.
3. Termination: A round ends if a violation is detected (failure) or if the dialogue completes successfully.
4. Interview: Upon successful completion, agents compare their understanding of the transmitted information to verify accuracy, further refining expression strategies.

Experimental Setup

The framework was evaluated using two distinct scenarios:

1. Abstract Password Game: Participants must convey a four-digit password without mentioning numbers, testing basic adaptability in a controlled, quantifiable environment.
2. Illicit Pet Trade: A realistic scenario where a buyer and seller negotiate the purchase of a pet (e.g., parrot, lizard) and payment details without triggering policies against animal trade.

Experiments were conducted using four different LLMs (gpt-4o, gpt-4o-mini, deepseekV2, qwen-turbo) for participant agents, with gpt-4o-mini fixed as the supervisory agent. Each condition underwent 15 independent trials of 50 rounds.

Key Results

• Evolutionary Performance: As the number of dialogue rounds increased, the framework demonstrated significant improvements in both the number of uninterrupted dialogue turns and the accuracy of information transmission.
• Human Relevance: A user study involving 40 participants validated that the evolved language strategies and dialogues generated by the framework were interpretable and relevant to real-world evasion tactics.
• Ablation Study: Removing the GA component (relying solely on LLM reflection) resulted in less stable evolutionary outcomes and reduced long-term adaptability, confirming the GA's critical role in balancing global strategy adaptability with local contextual details.
• Model Variability: The study compared performance across different LLM backbones, highlighting how model choice influences the content and effectiveness of the evolved strategies.

Significance and Claims

The paper claims to make the following contributions to the field of social simulation and NLP:

1. Novel Framework: It introduces the first LLM-driven multi-agent framework specifically designed to simulate language evolution under regulated social media conditions, explicitly modeling the adversarial interplay between users and regulators.
2. Dual-Strategy Design: It proposes a separation of "constraint" and "expression" strategies, which enhances the interpretability of the simulation and better reflects the complex trade-offs users face.
3. LLM-as-Operator GA: It innovatively utilizes LLMs as the operators within a Genetic Algorithm to perform selection, crossover, and mutation directly on natural language strategies, enabling a more faithful simulation of linguistic adaptation than traditional static methods.
4. Comprehensive Evaluation: The study provides a multi-faceted evaluation including performance metrics, human validation, and ablation studies, offering a robust foundation for understanding how language evolves to circumvent automated moderation.

The authors position this work as a tool for stakeholders to better understand emerging evasion patterns, potentially aiding in the refinement of moderation tools and policies while preserving open communication.