SO-Bench: A Structural Output Evaluation of Multimodal LLMs

2025-11-23

AI Summary

Existing benchmarks lack systematic evaluation of multimodal large language models’ (MLLMs) ability to perform patterned information extraction and structured reasoning from visual inputs. Method: We introduce SO-Bench—the first multimodal benchmark explicitly designed to assess structured output capabilities—spanning four visual domains: UI interfaces, natural images, documents, and charts. It comprises over 6.5K JSON schemas and 1.8K human-verified image–schema pairs, integrating diverse visual data, rigorous human-annotated alignment, and quantitative compliance metrics. Contribution/Results: Extensive evaluation across open- and closed-source MLLMs reveals substantial performance gaps in structured generation. Supervised fine-tuning on SO-Bench improves schema adherence by up to 32.7%. SO-Bench establishes a reproducible, scalable evaluation paradigm for structured understanding and generation in MLLMs.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Multi-modal Vision
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

Multimodal large language models (MLLMs) are increasingly deployed in real-world, agentic settings where outputs must not only be correct, but also conform to predefined data schemas. Despite recent progress in structured generation in textual domain, there is still no benchmark that systematically evaluates schema-grounded information extraction and reasoning over visual inputs. In this work, we conduct a comprehensive study of visual structural output capabilities for MLLMs with our carefully designed SO-Bench benchmark. Covering four visual domains, including UI screens, natural images, documents, and charts, SO-Bench is built from over 6.5K diverse JSON schemas and 1.8K curated image-schema pairs with human-verified quality. Benchmarking experiments on open-sourced and frontier proprietary models reveal persistent gaps in predicting accurate, schema compliant outputs, highlighting the need for better multimodal structured reasoning. Beyond benchmarking, we further conduct training experiments to largely improve the model's structured output capability. We plan to make the benchmark available to the community.

Problem

• Evaluates multimodal LLMs' structured output from visual inputs
• Assesses schema compliance and accuracy in visual information extraction
• Identifies gaps in multimodal structured reasoning across diverse domains

Innovation

• Benchmark for multimodal structured output evaluation
• Training experiments to improve structured output capability
• Diverse visual domains with curated image-schema pairs