Researchers from Tianjin University of Technology developed the Memory Decision Layer (MDL) to improve the reliability of LLM agents using RAG. The system evaluates documents based on three independent metrics: relevance, reliability, and task risk, using QR decomposition to prevent signal interference.
The MDL controller assigns a risk coefficient of 0.70 or higher to sensitive topics like medicine, law, and finance. If confidence falls below a set threshold, the system refuses to generate an answer. The process takes approximately 0.14 milliseconds, which is significantly faster than self-verification methods.
Testing showed that while standard RAG systems produce 53% hallucinations in conflict-heavy scenarios, the MDL controller reduces this to 23.3%. However, the system still produces hallucinations in 8% of approved cases. Performance varies by base model, with minimal gains observed on more advanced models like deepseek-v4-flash.
Comparative analysis indicates that simpler machine learning approaches, such as logistic regression and gradient boosting, achieved higher accuracy rates of 63.3% and 64.6% respectively, compared to the 61.2% accuracy of the geometric MDL controller. The system also struggles with high-risk, high-relevance content.
