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

A-CAEF, or Adaptive Context-Aware Embedding Filtering, is a lightweight decoding-time method for potentially improving factual reliability in large language models. It uses context-aware semantic drift signals and entropy-aware adaptationto guide token selection without retraining the model, modifying model parameters, or relying on external retrieval. This record contains the first public preprint version of A-CAEF. The method adjusts token selection during inference by applying context-aware drift signals and adaptive filtering strength based on model uncertainty. The paper reports preliminary evaluation on TruthfulQA using a Qwen-based decoder model and compares A-CAEF against nucleus sampling. In this setting, A-CAEF improves the joint TruthInfo metric by 6.67 absolute points, corresponding to a 28.6% relative gain over the baseline, while maintaining similar ROUGE-L and BLEU scores. These results provide preliminary evidence that context-aware decoding constraints can improve factual-informativeness without substantially degrading fluency. Broader validation across additional models, datasets, and decoding baselines is left for future work. Keywords: A-CAEF, Adaptive Context-Aware Embedding Filtering, context-aware semantic drift, entropy-aware adaptation, entropy-aware decoding, decoding-time token selection, factuality-aware decoding, hallucination mitigation, large language models, TruthfulQA, logit adjustment.