Middleware that catches when LLM conversations slowly drift off-topic (github.com)

A new recursive middleware called HALLUCINATIOFF has been introduced to enhance the regulation of outputs from large language models (LLMs) by detecting subtle issues such as semantic drift and premise flaws. Positioned between the LLM and the user, HALLUCINATIOFF operates without altering the model’s weights; instead, it observes responses before they reach the user and can decide on optimal actions, ranging from outright emission to cautious issuance or even blocking responses. This middleware employs a unique second-order observation technique that assesses not just whether responses are coherent, but also monitors the stability of the coherence criterion itself. The significance of HALLUCINATIOFF lies in its advanced psychological framework rooted in clinical practice, providing a novel approach to safeguarding LLM interactions. With an impressive 95% accuracy on synthetic test cases, it showcases potential for real-world applications, despite current limitations in fully assessing performance across diverse conversation scenarios. This innovative tool may ultimately improve user trust and safety in AI conversations by dynamically identifying and responding to sources of uncertainty and ambiguity.