Can MLLMs Decode the Creative Leap? Introducing C4 for Cross-Concept Understanding

The Gap in Creative AI Understanding

Multimodal language models (MLLMs)—systems that process both images and text—excel at factual tasks but struggle with creativity. This paper reveals that current MLLMs fail to decode creatively encoded meaning, where concepts connect through non-obvious but meaningful bridges. The authors find that even the strongest closed-source models achieve only 50.7% accuracy on their new evaluation benchmark, while open-source models perform substantially worse.

The core insight: creativity isn't magic—it's cross-concept understanding. When you interpret a metaphor or grasp a pun, you're recovering intended meaning by recognizing conceptual relations that aren't explicitly stated. This capacity is testable, measurable, and currently weak in MLLMs.

How C4 Works

The authors designed C4 (Cross-Concept Creativity for Chengyu) around Chinese idioms, which are inherently creative—they compress stories into four-character phrases that demand bridging disparate concepts to decode their meaning.

The evaluation framework operates in two steps:

Encoding: A human annotator maps a target concept (the idiom's intended meaning) to substitute concepts along bridge paths—chains of conceptual links. Think of it like encoding a secret message where each link is verifiable and logically defensible. For example, to convey "perseverance," one might bridge through "flowing water" → "erosion" → "persistence." The framework controls difficulty by varying the number and depth of bridges, creating structured, scalable test cases.

Decoding: MLLMs must recover the target concept from the substitute concept(s) by traversing the bridge path in reverse. Instead of asking "What does this idiom mean?" (vague, subjective), the task becomes: "Given these visual/linguistic proxies and bridge relations, identify the intended concept" (concrete, verifiable).

The C4-Eval dataset contains 184 synthetic items and 37 human-created cross-concept chengyu figures sourced from online communities. Each item is instantiated across five task settings, totaling 884 test cases. Bridge paths and reasoning chains were manually constructed and third-party reviewed to ensure quality.

Why This Matters

Current creativity metrics are soft—accuracy on creative tasks is subjective, rewards are sparse. C4 inverts this: it grounds creativity in explicit, verifiable, difficulty-indexed structure. This enables rigorous evaluation without sacrificing ecological validity (the idioms are real cultural artifacts).

The results expose concrete failure modes. While providing candidate constraints boosts accuracy sharply, merely asking MLLMs for explanations or hinting at bridges helps only modestly. This suggests the models struggle fundamentally with maintaining coherence across conceptual leaps, not just accessing bridge knowledge.

The 50.7% ceiling on closed models is striking—well above chance but far from human fluency. Open-source models lag further, highlighting that creative reasoning isn't simply a scaling phenomenon but a distinct capability gap. For applications in design, advertising, education, and human-AI collaboration, this gap is consequential.

By operationalizing creativity as a learnable, measurable task with controlled difficulty, C4 provides the first standardized benchmark for tracking progress on a capability that matters beyond accuracy: understanding what humans mean, not just what they say.