People often make decisions by holistically reasoning over heterogeneous personal information. This paper investigates the relatively under‑explored capability of multimodal large language models (MLLMs) to perform personalized reasoning. To this end we introduce MPRBench, a novel benchmark specifically designed for evaluating personalized reasoning across a broad spectrum of tasks and real‑world challenges. MPRBench comprises 12 sub‑tasks and supports interpretable analysis of representative error cases.\n\nThe benchmark covers scenarios such as:\n- Predicting user interests from photos and textual descriptions\n- Sentiment inference based on historical chat logs\n- Context‑aware QA using personal location data\n\nExtensive experiments reveal two major shortcomings of current personalized MLLMs:\n1. Inaccurate recognition of personalized concepts, leading to reasoning that deviates from true user preferences;\n2. Over‑sensitivity to irrelevant personal information, causing noise‑induced errors.\n\nWe hope MPRBench will stimulate further research on personalized reasoning in MLLMs and serve as a systematic, interpretable evaluation platform. The code and data are released at https://github.com/kkdh15/MPRBench.\n\nReview: MPRBench offers a comprehensive and interpretable framework for assessing multimodal models' ability to leverage personal information, highlighting critical gaps in concept recognition and noise robustness that future work must address.