LARGE LANGUAGE MODEL HALLUCINATION ATTENUATION BASED ON MULTI-AGENT DEBATE

Smart overview of the Invention

The patent application introduces a novel approach to mitigate hallucinations in large language models (LLMs) by employing a multi-agent debate framework. This method involves selecting agents and critics based on a user's query to form a personality council. Within this council, agents with specific personalities engage in debates with critics, who provide feedback on the agents' responses. The process continues until a satisfactory consensus is reached, ensuring more reliable outputs from the LLMs.

A significant challenge addressed by this approach is the occurrence of hallucinations in LLMs, especially in business-critical applications where reliability is paramount. Hallucinations can lead to incorrect outputs, which may undermine user trust and result in operational disruptions. Even with retrieval-augmented generation techniques, hallucinations remain a concern, highlighting the need for improved methodologies to ensure accurate and dependable information generation.

The proposed system enhances the accuracy of LLM outputs by fostering rigorous critical evaluation through debates among agents and critics. This process not only reduces the frequency of hallucinations but also improves user experiences by delivering trustworthy and relevant responses. Consequently, businesses can rely on LLMs for informed decision-making, while maintaining compliance with regulatory standards and avoiding potential legal issues from erroneous outputs.

The multi-stage process begins with receiving a user's query, followed by selecting personalities for the agents involved. The query is then processed through a distributed workflow managed by the selected personalities. The interaction and debate among these personalities are overseen to ensure a comprehensive evaluation of responses. Finally, the LLM responds to the query, with the entire process designed to function flexibly across various stages.

An example of the implementation involves constructing a personality council and conducting debates among agents and critics until a consensus is reached. This method facilitates the creation of an action-flow comprising proposed actions based on critiqued responses, ultimately enhancing the reliability and accuracy of the LLM's outputs. The approach provides a technological advancement over existing systems, promoting operational integrity and user satisfaction.