Hugging Face researchers have introduced MBA-Bench, the first multimodal benchmark designed to train and evaluate AI agents for business ideation. The benchmark addresses a key limitation of existing systems, which rely solely on text despite the inherently visual nature of real-world business contexts.
MBA-Bench comprises 30,000 samples across six domains, each featuring distinct visual cues that text alone cannot fully convey. To build the benchmark, the team automatically captioned images and used GPT-4o to generate five reference ideas per business question, employing retrieval query generation, market evidence retrieval, and evidence-augmented synthesis.
Alongside the benchmark, the researchers propose two agents: MBA-b (blind) and MBA-k (known). Both are trained using LoRA-based supervised fine-tuning followed by group relative policy optimization, with novel reward objectives for creativity and feasibility. MBA-k additionally optimizes six disclosed business-oriented criteria, bringing the total to eight.
In extensive experiments, MBA-b and MBA-k outperformed caption-only baselines by 63.9% and 77.1%, respectively, and surpassed open-source multimodal baselines by 25.6% and 35.8%. The results demonstrate the value of incorporating visual information into business ideation, moving beyond text-centric approaches.
The project is open-sourced, with code, models, dataset, and an interactive demo available on Hugging Face and GitHub.