跳转至

MBA:面向现实世界商业构思的多模态基准与智能体

文章背景与核心概要

现实世界的商业决策与构思往往高度依赖于多模态信息(如市场图表、产品外观及视觉环境),然而现有的基于大语言模型(LLM)的智能体框架大都局限于纯文本范式,难以应对视觉线索至关重要的实际商业场景。为此,本文推出了 MBA-Bench——首个专为训练和评估现实世界商业构思智能体而设计的综合多模态基准。该基准涵盖来自六大商业领域的 30,000 个样本,每个样本均融合了关键的视觉线索。

为了突破现有局限,研究团队开发了两种智能体变体:针对评估标准隐藏场景的 MBA-b 以及针对评估标准公开场景的 MBA-k。在训练方法上,两款智能体均通过基于 LoRA 的监督微调(SFT)以及引入全新“创造力”与“可行性”奖励的群组相对策略优化(GRPO)进行训练。实验结果表明,在 MBA-Bench 的各项严格测试中,MBA-b 和 MBA-k 的表现均显著超越了传统的纯文本基线以及现有的多模态基线,展现出强大的现实世界商业构思能力。


摘要 (Abstract)

由大语言模型(LLMs)驱动的智能体系统为商业构思开辟了新的机遇。然而,尽管现实世界的上下文本质上具有多模态特征,现有方法仍局限于纯文本范式。

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.

因此,我们推出了 MBA-Bench,这是首个用于训练和评估商业构思智能体的多模态基准,包含跨越六个领域的 30K 个样本,每个领域都具有无法仅通过文本完全传达的独特视觉线索。具体而言,我们自动为图像添加标题(caption),并利用 GPT-4o 通过检索查询生成、市场证据检索和证据增强合成,为三个商业问题中的每一个生成五个参考构思。

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.

遵循先前的工作,我们使用 MLLM-as-a-Judge(多模态大模型作为裁判)在六个面向商业的标准上评估智能体。为了考虑标准隐藏或公开的场景,我们分别提出了用于隐藏场景的 MBA-b 和用于公开场景的 MBA-k。我们使用两个新颖的奖励目标——创造力和可行性——来训练这两者,同时 MBA-k 进一步优化了六个公开标准,总共八个标准。两者的训练均通过基于 LoRA 的监督微调进行,随后利用这些针对特定场景的奖励进行群组相对策略优化(group relative policy optimization)。

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.

为了在 MBA-Bench 上进行广泛的实验,我们设置了两个仅容纳图像标题或多模态输入的基线,后者在多项指标上接近闭源模型的性能。在基准测试中,MBA-b 和 MBA-k 分别比标题基线高出 63.9% 和 77.1%,比多模态基线高出 25.6% 和 35.8%。

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.


元数据与参考信息 (Metadata & Reference Information)

  • arXiv ID: arXiv:2608.11616 [cs.AI]
  • 学科分类 (Subjects): 人工智能 (cs.AI);计算机视觉与模式识别 (cs.CV);机器学习 (cs.LG)
    • Subjects: Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
  • 作者 (Authors): Hojun Choi, Jaeyo Shin, Suin Lee, Hyunjung Shim
    • Authors: Hojun Choi, Jaeyo Shin, Suin Lee, Hyunjung Shim
  • 提交时间 (Submitted): 2026年8月12日 (修订版: 2026年8月13日)
    • Submitted: August 12, 2026 (Revised: August 13, 2026)
  • 许可与权利 (License & Rights): 知识共享署名 4.0 license icon