Most prior works focused on conflicts between an LLM’s internal parametric knowledge and externally provided context. In contrast, we investigate how LLMs handle conflicts that arise within contextual knowledge itself. We introduce a taxonomy of six types of contextual conflicts (factual, inferential, temporal, granularity, perspective, and ambiguity) and contribute a comprehensive dataset ContextConflict for this setting. The dataset contains 5,781 samples, covers both reasoning and summarization tasks, and includes both explicit contradictions and implicit conflicts that require multi-step reasoning. Experiments on nine LLMs show that current models still fall short in resolving contextual knowledge conflicts. We further provide mechanistic interpretability insights into how LLMs process such conflicts, revealing their latent awareness of conflicts and the representational geometry underlying conflict processing. In addition, our analysis uncovers a consistent model bias towards earlier evidence, and this positional preference serves as a key obstacle to effective conflict resolution. Motivated by these findings, we further propose a simple training-free, label-free steering method that steers activations to encourage a more comprehensive incorporation of evidences for better conflict resolution. On our dataset, the method consistently improves accuracy on reasoning tasks and generates higher-quality, more balanced summaries for summarization tasks.

This paper describes the machine translation system submitted to the SemEval-2025 Entity-Aware Machine Translation Task by the SheffieldGATE Team. We proposed a multi-agent entity-aware machine translation system that operates through three distinct reasoning stages: entity recognition, knowledge enhancement, and translation decision-making. The innovation in our approach lies in leveraging large language models to generate contextually relevant queries during the knowledge enhancement stage, extracting candidate entities and their translations from external knowledge bases. In the final translation decision-making stage, we employ fine-tuned large language models to denoise the retrieved knowledge, selecting the most relevant entity information to ensure accurate translation of the original text. Experimental results demonstrate our system’s effectiveness. In emEval-2025 Task 2, our system ranks first among all systems in Spanish entity translation metrics and third in Italian. For systems that do not use gold standard entity IDs during test set inference, ours achieves the highest overall scores across four language pairs: German, French, Italian, and Spanish.

This paper describes SheffieldGATE’s submission to WMT 2024 Chat Shared Translation Task. We participate in three language pairs: English-German, English-Dutch, and English-Portuguese (Brazil). In this work, we introduce a context-aware sliding window decoding method to track dependencies between chat messages. We fine-tune a large pre-trained language model based on the training data provided by the shared task Our experiments (i) compare the model performance between multilingual and bilingual fine-tuning and (ii) assess the impact of different window sizes. Our experimental results demonstrate that utilising contextual information yields superior performance in document-level translation compared to translating documents as isolated text segments, and that models fine-tuned with multilingual data perform better than those fine-tuned with bilingual data.