MBA: Multimodal Benchmark and Agents for Real-World Business Ideation
Hojun Choi, Jaeyo Shin, Suin Lee, Hyunjung Shim
cs.AI, cs.CV, cs.LG
2026-08-12
The first multimodal business-ideation benchmark MBA-Bench (30K samples, six domains), with a LoRA+GRPO 7B agent that beats GPT-5-mini on most dimensions and exceeds text and multimodal baselines by 64-77%.
Using an LLM agent to help generate business ideas is a new direction, but existing work is stuck in text only: feed the model a scene description and ask for ideas. The real world is multimodal, though. A crowded street scene, a mobile-interface layout, or a circuit board holds visual cues that text cannot convey, and those cues may point straight at an unmet need.
This paper pushes business ideation from text into multimodal input, and for that it builds the first multimodal benchmark, MBA-Bench, plus an agent trained on it.
MBA-Bench has 30K samples across six domains, each chosen for a trait that text alone cannot capture: everyday scenes (ADE20K), dense mobile-interface layouts (RICO), crowded images (COCO, at least 10 people), surface defects (VisA anomaly images), material textures (DTD), and circuit-board structure (DeepPCB).
Each image is captioned, and around three business questions (cost efficiency, technology, user experience) five reference ideas are generated with GPT-4o via retrieval-query generation, market-evidence retrieval, and evidence-augmented synthesis. Evaluation uses MLLM-as-a-Judge on six criteria: specificity, technical validity, innovativeness, competitive advantage, need validity, and market size.
The agent is built on Qwen2.5-VL-7B-Instruct, first LoRA-supervised fine-tuned then improved with GRPO. Two variants: MBA-b (blind, with only creativity and feasibility rewards) and MBA-k (known, which also folds the six disclosed criteria into the reward for eight objectives total).
Both trained small models clearly beat the baselines.
| Model | vs caption baseline | vs multimodal baseline |
| MBA-b-7B | +63.9% | +25.6% |
| MBA-k-7B | +77.1% | +35.8% |
MBA-k (given the scoring criteria) beats MBA-b (blind), showing that disclosing evaluation dimensions to training helps. Across the six criteria, the 7B MBA-k surpasses GPT-5-mini on innovativeness (4.00), competitive advantage (3.32), need validity (2.94), and market size (2.75), essentially ties on specificity (3.99), and is only slightly behind on technical validity (3.00 vs 3.07). A fine-tuned 7B model matching or beating a frontier closed model on most dimensions wins mainly on divergent metrics like innovativeness and market judgment.
For teams building vertical agents, this is a case study in how a small model with task-aligned training and clear rewards can match a large model on a specific task. It also demonstrates a benchmark-design principle: set visual information as a hard gate that text cannot convey, forcing the agent to actually look at the image.
Turning an inherently subjective, divergent task like ideation into six scorable dimensions plus a reinforcement-learning reward is a reusable engineering pattern.
The authors concede three points. Only image and text are covered, missing audio, tactile, and olfactory signals. There is no temporal reasoning, even though video could supply behavioral and causal context. And ideas are not personalized: they ignore the entrepreneur's capital, expertise, location, network, and risk appetite, so an idea right for one person may be unrealistic for another.
One more concern: the six criteria are scored by an MLLM judge, and MBA-k uses those same criteria as reward, so judge and reward share a source. This invites a self-evaluation inflation risk that the paper does not cross-check with separate human scoring.