Prior works have shown that fine-tuning on new knowledge can induce factual hallucinations in large language models (LLMs), leading to incorrect outputs when evaluated on previously known information. However, the specific manifestations of such hallucination and its underlying mechanisms remain insufficiently understood. Our work addresses this gap by designing a controlled dataset Biography-Reasoning, and conducting a fine-grained analysis across multiple knowledge types and two task types, including knowledge question answering (QA) and knowledge reasoning tasks. We find that hallucinations not only severely affect tasks involving newly introduced knowledge, but also propagate to other evaluation tasks. Moreover, when fine-tuning on a dataset in which a specific knowledge type consists entirely of new knowledge, LLMs exhibit elevated hallucination tendencies. This suggests that the degree of unfamiliarity within a particular knowledge type, rather than the overall proportion of new knowledge, is a stronger driver of hallucinations. Through interpretability analysis, we show that learning new knowledge weakens the model’s attention to key entities in the input question, leading to an over-reliance on surrounding context and a higher risk of hallucination. Conversely, reintroducing a small amount of known knowledge during the later stages of training restores attention to key entities and substantially mitigates hallucination behavior. Finally, we demonstrate that disrupted attention patterns can propagate across lexically similar contexts, facilitating the spread of hallucinations beyond the original task.

Large Language Models (LLMs) demonstrate strong semantic understanding ability and extensive knowledge, but struggle with Named Entity Recognition (NER) due to hallucination and high training costs. Meanwhile, supervised Small Language Models (SLMs) efficiently provide structured predictions but lack adaptability to unseen entities and complex contexts. In this study, we investigate how a relatively weaker LLM can effectively support a supervised model in NER tasks. We first improve the LLM using LoRA-based fine-tuning and similarity-based prompting, achieving performance comparable to a SLM baseline. To further improve results, we propose a fusion strategy that integrates both models: prioritising SLM’s predictions while using LLM guidance in low confidence cases. Our hybrid approach outperforms both baselines on three classic Chinese NER datasets.