The rapid advancement of Large Multimodal Models (LMMs) has ushered in an era of AI systems capable of processing both text and images, demonstrating impressive abilities to answer complex questions. But are these models truly understanding the visual world, or are they simply leveraging clever shortcuts and linguistic tricks to arrive at the right answers? New research suggests the latter may be more common than we think, potentially undermining the trustworthiness of current evaluation methods. Researchers have introduced MMEvalPro, a novel benchmark designed to expose the limitations of multimodal AI. The problem lies in how LMMs are currently tested, often with multiple-choice questions (MCQs) that rely on datasets available online. Large Language Models (LLMs), which only process text, have been shown to perform surprisingly well on these tests. How? They exploit biases in the datasets and often correctly guess answers without "seeing" the image at all. This means the evaluations don't accurately reflect whether an LMM uses its visual processing capabilities. MMEvalPro introduces a three-part question system for each MCQ. In addition to the original question, the benchmark tests "perception" (understanding details within the image) and "knowledge" (applying reasoning based on the image and question). To truly "pass," an LMM must answer all three related questions correctly. The results are revealing. Even leading LMMs struggled to maintain consistency across the questions. While they might answer the original MCQ correctly, their performance on the perception and knowledge questions exposed gaps in their understanding. This suggests that impressive performance on standard MCQ benchmarks can be misleading, potentially inflating the perceived capabilities of LMMs. This discrepancy highlights a crucial challenge in evaluating LMMs: distinguishing genuine multimodal understanding from linguistic cleverness. MMEvalPro addresses this challenge by directly testing the prerequisite steps involved in visual reasoning. This new benchmark opens doors to develop more robust and trustworthy LMMs that move beyond superficial pattern matching to true visual reasoning, getting us closer to genuinely intelligent AI systems. The next generation of LMMs needs to move beyond simply identifying objects in images and start understanding the relationships and logic within visual contexts. This shift will require innovative training methods and evaluation benchmarks that focus on genuine understanding rather than test-taking skills. MMEvalPro represents an important step in that direction, paving the way for a future where AI truly "sees" and understands the world around it.