MBA Benchmark and Agents Boost Multimodal Business Ideation

Agentic systems powered by large language models (LLMs) have opened new avenues for business ideation, but current approaches are largely confined to text-only inputs. This limitation overlooks the inherently multimodal nature of real-world business contexts, where visual cues often provide critical information. To address this, researchers introduce MBA-Bench, the first multimodal benchmark specifically designed for training and evaluating business ideation agents. MBA-Bench consists of 30,000 samples spanning six diverse domains, each featuring unique visual information not fully conveyed by text alone. The benchmark uses automatically captioned images and GPT-4o to generate reference ideas based on retrieval, market evidence, and evidence-augmented synthesis. Agents are evaluated across six business-oriented criteria using an MLLM-as-a-Judge approach. The study presents two novel agents, MBA-b (blind criteria) and MBA-k (known criteria), trained with new reward objectives focusing on creativity and feasibility. MBA-k further optimizes for the six disclosed criteria. Both agents, trained via LoRA-based supervised fine-tuning and group relative policy optimization, significantly outperform both caption-only and multimodal baselines, demonstrating substantial improvements in business ideation performance.