Hojun Choi, Jaeyo Shin, Suin Lee, Hyunjung Shim
4 min
Abstract
Agentic systems powered by large language models (LLMs) have opened new opportunities for business ideation. Yet existing approaches remain confined to a text-only paradigm, despite the inherently multimodal nature of real-world contexts. We thus introduce MBA-Bench, the first multimodal benchmark for training and evaluating business ideation agents, comprising 30K samples across six domains, each domain characterized by distinct visual cues not fully conveyed by text alone. Concretely, we automatically caption images and employ GPT-4o to generate five reference ideas for each of three business questions through retrieval query generation, market evidence retrieval, and evidence-augmented synthesis. Following prior work, we evaluate agents across six business-oriented criteria using MLLM-as-a-Judge. To consider settings where criteria are hidden or disclosed, we present MBA-b and MBA-k for blind and known, respectively. We train both with two novel reward objectives---creativity and feasibility---while MBA-k further optimizes the six disclosed criteria for eight in total. Both are trained via LoRA-based supervised fine-tuning followed by group relative policy optimization with these setting-specific rewards. For extensive experiments on MBA-Bench, we set up two baselines accommodating either captions only or multimodal inputs, with the latter nearing closed-source performance on several metrics. MBA-b and MBA-k outperform caption baselines by 63.9% and 77.1%, and multimodal baselines by 25.6% and 35.8%, respectively.
Sam: What are the limits, though? I'm guessing a single photograph only tells you so much.
Alex: That's the key limitation the paper acknowledges. The model analyzes a single snapshot. It can't yet process video or track how an environment changes over time — which matters a great deal in places like a busy retail store or a factory floor where things are constantly shifting.
Sam: It's a bit like trying to plan a traffic route by looking at one still photo of an intersection, rather than watching the cars actually move.
Alex: That's an accurate comparison. There's also another gap: the system doesn't yet account for an individual entrepreneur's specific constraints — their budget, their expertise, the scale of their operation. A solid idea for a large corporation might be completely out of reach for a small business owner.
Sam: So where does the research go from here?
Alex: The paper points toward what it calls temporal dynamics — which is essentially the ability to watch and interpret live video rather than static images. The researchers suggest that future versions could, for instance, observe how customers move through a store in real time and suggest layout changes on the fly.
Sam: So the shift is from "what does this image mean" to "what is actually happening here, right now."
Alex: That's the direction. The current model is a meaningful step toward AI that can genuinely observe the world rather than just read about it. Whether it can eventually keep pace with the complexity of real environments in real time — that's the open question the field is now working toward. Thanks for listening to ResearchPod.