Thesis: An Orchestration Framework for LLM Hallucination Suppression Based on Multi-Agent Debate

The Thesis framework reduces hallucinations in large language models through a structured multi-agent debate mechanism, with the core idea of using reasoning diversity across different models for cross-validation; the framework adopts role division (Solver/Critic/Validator) and flexible debate depth design, and achieves scalability through a modular architecture, aiming to build a more reliable collaborative AI system.

Large language models have hallucination issues: they confidently generate incorrect information, fabricate facts, or misinterpret contextual details, and a single model lacks a self-verification mechanism, making this flaw particularly fatal in complex tasks.

Core Insight: The reasoning diversity formed by different models due to differences in training data and processing methods can be transformed into cross-verification capabilities; the architecture includes an input preprocessing layer (information extraction, task structuring), role division (Solver generates initial answers/Critic detects loopholes/Validator synthesizes results), and configurable debate depth (rounds/reasoning depth/model selection).

The backend uses Python FastAPI/Uvicorn to provide high-performance APIs; the model layer supports expansion based on the OpenAI API; the architecture pattern is an Orchestrator coordinating Roles to execute the Pipeline, ensuring system scalability.

Current areas to improve: fine-tuning dedicated models (context extraction/task decomposition), supporting local execution, intelligent routing (dynamic assignment of model roles), persistent memory (long context optimization), and introducing a fact-checking layer.

Thesis represents a paradigm shift: from pursuing a single strong model to building a reliable collaborative system, which aligns with human decision-making wisdom; it is suitable for high-reliability scenarios such as medical diagnosis and legal analysis; the vision is to build a trustworthy collaborative AI system and provide an engineering solution to the LLM credibility problem.