<p>Accurate and reliable bearing fault diagnosis is critical for ensuring the safe operation of mechanical equipment. Previous data-driven methods encounter challenges in training advanced deep learning models, primarily due to the scarcity of fault data and inconsistencies in data distributions. Additionally, these methods often suffer from limited interpretability and reliability, as they lack constraint-guided learning based on the physical mechanisms underlying bearing failures, which hampers their utility in machine condition monitoring. Recent advancements in large language models (LLMs) demonstrate their potential to address these challenges. To this end, we aim to enhance the generalization and interpretability of bearing fault diagnosis by leveraging the capabilities of multimodal LLMs. Specifically, a novel framework called DiagLLM is designed to achieve this goal. DiagLLM leverages the powerful reasoning capabilities of large language models and incorporates contextual information from both envelope spectrum images and expert knowledge to accurately diagnose bearing faults. To effectively tune DiagLLM, diagnostic visual instruction-following data are constructed to link fault feature descriptions with signal characteristics, and the entire model is fine-tuned using a parameter-efficient training pipeline. Extensive experiments are conducted on two publicly available bearing fault diagnosis datasets, and the results show that DiagLLM outperforms leading baselines, particularly in scenarios with limited data and cross-data generalization.</p>