Adaptive Context-Aware Embedding Filtering: Balancing Factual Accuracy and Creative Freedom in Language Model Decoding

This record contains the first public preprint version of “Adaptive Context-Aware Embedding Filtering (A-CAEF),” a decoding-time approach for improving factual reliability in large language models. A-CAEF adjusts token selection using context-aware drift signals without retraining the model, modifying model parameters, or relying on external retrieval. The preprint reports preliminary evaluation on factuality-oriented generation and discusses the method’s potential use as a lightweight inference-time strategy for reducing factual drift.