Research map / Vision-language and multimodal / Multimodal LLMs

CreBench: Human-Aligned Creativity Evaluation from Idea to Process to Product

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CreBench, a 12-indicator creativity evaluation benchmark (idea, process, product) with the CreMIT instruction-tuning dataset, and fine-tunes LLaVA-1.5-7B into CreExpert, which outperforms GPT-4V and Gemini-Pro-Vision on human-aligned creativity scoring.

Problem

MLLMs struggle to assess human-defined, abstract and subjective creativity in open-ended visual design tasks, and no benchmark exists for multimodal creativity evaluation aligned with human judgments.

Approach

Decompose creativity into idea, process and product dimensions with 12 expert-rated indicators, convert expert feedback into multi-type GPT-4o instruction data, and instruction-tune an open MLLM (LLaVA-1.5-7B) into a creativity evaluation expert.

Key results

• CreExpert achieves overall score 65.50% vs 29.27% for GPT-4V and 27.78% for Gemini-Pro-Vision on CreBench (Pearson-based).
• CreExpert surpasses its LLaVA-1.5-7B baseline by roughly 45 points overall (65.50% vs 20.57%).
• Inter-rater reliability reported as average Fleiss' kappa 0.71 and ICC(2,1) 0.78 among expert annotators.

Surprising findings

• Product-level scores remain low even for CreExpert (about 40% average), with Manufacturability and Novelty gains varying across tasks and a slight drop on Reach.

Limitations

• Evaluation limited to four drawing-based tasks with secondary students, limiting generalizability.
• Only 50/50 split of a small 2.2K-instance dataset used for fine-tuning and evaluation.
• Instruction data generated by GPT-4o from expert feedback may inherit its biases.

Experimental setup

Datasets: CreMIT (CreBench training split)

Benchmarks: CreBench

Metrics: Pearson correlation coefficient

Baselines: GPT-4V Gemini-Pro-Vision mPLUG-Owl2 InstructBLIP Seed1.5-VL Qwen2.5-VL MiniGPT-v2 GLM MiniGPT-4 TinyGPT LLaVA-1.5-7B

Builds on: LLaVA-1.5 Vicuna CLIP LLaMA-Factory

Closest papers by meaning

• CREval: An Automated Interpretable Evaluation for Creative Image Manipulation under Complex Instructions (CVPR 2026 · similarity 0.87)
• Redefining <Creative> in Dictionary: Towards an Enhanced Semantic Understanding of Creative Generation (CVPR 2025 · similarity 0.86)
• IDEAFix: Evaluation Framework for Creative Defixation Prompting in LLMs (NeurIPS 2026 · similarity 0.86)
• AI as Humanity’s Salieri: Quantifying Linguistic Creativity of Language Models via Systematic Attribution of Machine Text against Web Text (ICLR 2025 · similarity 0.86)
• IDEA-Bench: How Far are Generative Models from Professional Designing? (CVPR 2025 · similarity 0.85)
• CLAWS:Creativity detection for LLM-generated solutions using Attention Window of Sections (NeurIPS 2025 · similarity 0.85)