Research map / Vision-language and multimodal / Multimodal LLMs

AnyDoc: Enhancing Document Generation via Large-Scale HTML/CSS Data Synthesis and Height-Aware Reinforcement Optimization

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds DocHTML, a 265K-sample HTML/CSS document dataset synthesized by MLLM/code/image models, fine-tunes Qwen2.5-VL-7B for three document generation tasks, and adds height-aware GRPO (HARL) to reduce content overflow.

Problem

Document generation is limited by narrow category coverage, flat coordinate or raster representations that hurt editability and structure, and scarce large-scale data; generated documents also often overflow their specified heights.

Approach

Represent documents as hierarchical HTML/CSS, synthesize a large multi-category HTML/CSS dataset with an automated pipeline, and post-train with GRPO using a reward based on the ratio of rendered height to target height so overflow (and underflow) is penalized.

Key results

• On intention-to-document, AnyDoc scores 8.64 layout and 8.92 image vs GPT-4o 8.59 and 8.75, with Height overflow 0.005 vs GPT-4o 0.047.
• On element-to-document, AnyDoc reaches 7.36 innovation and 0.085 height versus GPT-4o's 6.87 and 0.212.
• Ablation on E2D shows HARL reduces Height from 0.392 to 0.085 and flat coordinate representation yields Overlap 0.2627 vs 0.0494 for HTML/CSS.

Surprising findings

• Flat coordinate representation gets Height 0 but much worse overlap (0.2627 vs 0.0494), showing a trade-off between overflow and layout quality.
• Scaling the synthetic data from 10K to full set did not noticeably change metrics, suggesting robustness but limited gains from more data.

Limitations

• Height metric on derendering remains non-trivial (0.309 overflow), attributed to small model size.
• Evaluation relies on MLLM judges (InternVL3-78B) and a 1,000-sample test subset.
• Baselines like GPT-4o and InternVL3 are zero-shot and prompt-based, so comparison is not fully controlled.

Experimental setup

Datasets: DocHTML

Benchmarks: DocHTML test split

Metrics: Layout Image Typography Content Innovation Intention Overlap Alignment Height Block

Baselines: GPT-4o InternVL3-78B OpenCOLE FLUX.1-dev WebSight Design2Code LaDeCo

Builds on: GRPO Qwen2.5-VL LLaMA-Factory LoRA VeRL

Closest papers by meaning

• BigDocs: An Open Dataset for Training Multimodal Models on Document and Code Tasks (ICLR 2025 · similarity 0.84)
• Adaptive Markup Language Generation for Contextually-Grounded Visual Document Understanding (CVPR 2025 · similarity 0.84)
• AesthetiQ: Enhancing Graphic Layout Design via Aesthetic-Aware Preference Alignment of Multi-modal Large Language Models (CVPR 2025 · similarity 0.83)
• Seeing is Improving: Visual Feedback for Iterative Text Layout Refinement (CVPR 2026 · similarity 0.83)
• Table2LaTeX-RL: High-Fidelity LaTeX Code Generation from Table Images via Reinforced Multimodal Language Models (NeurIPS 2025 · similarity 0.83)
• AnyText: Multilingual Visual Text Generation and Editing (ICLR 2024 · similarity 0.83)