C4 Evaluates MLLM Creative Understanding of Cross-Concepts.

Evaluating the creative capabilities of Multimodal Large Language Models (MLLMs) is challenging due to the lack of clear targets and reward signals. This research proposes C4, a new framework designed to assess "receptive creativity," specifically focusing on cross-concept understanding – the ability to grasp intended meaning from non-obvious conceptual relationships. C4 operationalizes this by encoding cross-concept relations, often inspired by Chinese idioms (Chengyu), into items and then evaluating the MLLM's ability to decode them. The C4 Evaluation Set (C4-Eval) includes both synthetic and human-created items, each with detailed cross-concept relations and reasoning paths. Tested across ten MLLMs in various task settings, the strongest closed models achieved only around 50% accuracy, while open-source models performed significantly worse. The study found that while candidate constraints improved accuracy, hints and explanation requests offered only modest gains. These results highlight a considerable deficiency in how current MLLMs process and understand creatively encoded meanings through complex conceptual connections.