MCAD-EUC: Multi-context adaptive decoding with entropy-based uncertainty calibration for knowledge conflict mitigation

The knowledge sources of large language models (LLMs) encompass both parametric internal knowledge and external contextual information. However, conflicts between these two sources can significantly impair model performance. Existing methods typically assume a priori correctness of either the context or the parametric knowledge, lacking dynamic coordination mechanisms and being limited to single-context scenarios. To address this issue, this work proposes a lightweight and training-free decoding method,Multi-Context Adaptive Decoding (MCAD-EUC), which dynamically measures the effectiveness of both knowledge through Entropy based Uncertainty Calibration. It does not concern itself with whether the knowledge is false or true, the internal or the external, but balancing them according to their contributions to correctly answering the question. Particularly, MCAD-EUC is naturally multi-contextual. It can dynamically amplify the distribution of golden context while mitigating the influence of noisy context, thereby optimizing the final logits for predicting the next token during the decoding process. To comprehensively evaluate the model performance in multi-context scenarios, this work constructs MCQA, a multi-context question answering dataset that includes golden context, irrelevant context, and six categories of misleading context (crowd, logic, temporal, authority, emotional, numeric), simulating the diversity of noise in real-world settings. Extensive experiments on four LLMs and four MCQA datasets demonstrate that MCAD-EUC achieves an average accuracy improvement of 3.17 % over the best-performing baseline methods. Further sensitivity analysis confirms that the entropy-based adaptive weighting mechanism consistently outperforms all fixed-weight settings. Our dataset and code will be publicly available.