Conditional Mechanistic Weight Edits: A Potential Framework for Targeted Real-Time Hallucination Reduction in LLMs

ABSTRACT The explosive growth and power of large language models (LLMs) have captured global attention. In just a few years, tools like ChatGPT, Gemini, and Claude have become integrated into everyday life. Despite their strengths, however, LLMs sometimes generate misleading or false information, known as “hallucinations.” To address this, we investigated Conditional Mechanistic Weight Edits (CMWE), a novel technique designed to combat two specific types of hallucinations—providing incorrect answers to undefined math problems and generating fake citations—while preserving the model's general capabilities. CMWE, in theory, combines a hallucination detector and custom Low-Rank Adaptations (LoRAs) [1] to adjust the model's internal parameters in real-time, thus reducing the likelihood of a hallucination. We evaluated CMWE on Mistral 7B-Instruct (v0.3), a 7-billion-parameter LLM from Mistral AI. While our results failed to show a meaningful reduction in hallucinations, further investigations are needed to prove or disprove the true merit of CMWE.