Research map / Multi-agent systems and games / LLM agents
FormAct: Agentic Source Editing for Rich-Format Document Generation
Eugene J. Yu, Xingxing Zhang, Yuan Xia, Tao Ge, Xun Wang, FNU Kartik, Vishwas Suryanarayanan, Cheng Yang, Amanda Jiang, Jiayu Ding, Xiangyu Wong, Tengchao Lv and 4 more
Read the paperCheck your idea against 137,217 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.
Presents FormAct, an agentic system that generates professional rich-format documents from scratch by iteratively refining an HTML source with an editing agent that calls a suite of tools.
Problem
Rich-format documents are costly to author manually, motivating automated generation of professional documents.
Approach
Operates on an HTML source representation and performs iterative source refinement using an editing agent that invokes a suite of tools, rather than generating documents in one shot.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add review-guided refinement loop (ablation) | FormAct · RichDocBench | rendering correctness: 4.59 -> 4.81 | Improved |
| apply refinement loop to Direct Generation drafts (ablation) | DG · RichDocBench | rendering correctness: 4.34 -> 4.82 (+11.1%) | Improved |
| apply refinement loop to Codex drafts (ablation) | Codex · RichDocBench | rendering correctness: 4.40 -> 4.92 (+11.8%) | Improved |
| apply refinement loop to TAG drafts (ablation) | TAG · RichDocBench | layout appropriateness: 3.78 -> 3.76 (-0.5%) | No clear effect |
| apply refinement loop to Codex drafts (ablation) | Codex · RichDocBench | layout appropriateness: 3.84 -> 3.98 (+3.6%) | Improved |
| add retrieved templates as in-context examples (TAG vs DG) | Direct Generation · RichDocBench | layout appropriateness: 3.76 -> 3.70 | Hurt |
| add retrieved templates as in-context examples (TAG vs DG) | Direct Generation · RichDocBench | professionalism: 3.75 -> 3.80 | Improved |
| use general-purpose coding agent (Codex) instead of direct generation | Direct Generation · RichDocBench | rendering correctness: 4.36 -> 4.39 | No clear effect |
| construct documents via python-docx library instead of editing HTML source | FormAct · RichDocBench | layout appropriateness: 3.92 -> 3.39 | Hurt |
| FormAct source editing with template retrieval vs DG | Direct Generation · RichDocBench | layout appropriateness: 3.76 -> 3.92 | Improved |
| FormAct compared with Codex in human ranking | Codex · RichDocBench human preference (50 examples) | rank-1 fraction: 0.140 -> 0.760 | Improved |
Closest papers by meaning
- WebGen-Agent: Enhancing Interactive Website Generation with Multi-Level Feedback and Step-Level Reinforcement Learning (ICLR 2026 · similarity 0.84)
- PPT-Eval: A Benchmark for Computer-Use Agents on PowerPoint Tasks (ICML 2026 · similarity 0.83)
- FullStack-Agent: Enhancing Agentic Full-Stack Web Coding via Development-Oriented Testing and Repository Back-Translation (ICML 2026 · similarity 0.83)
- AnyDoc: Enhancing Document Generation via Large-Scale HTML/CSS Data Synthesis and Height-Aware Reinforcement Optimization (CVPR 2026 · similarity 0.83)
- NL2Repo-Bench: Towards Long-Horizon Repository Generation Evaluation of Coding Agents (ICML 2026 · similarity 0.82)
- WebGen-Bench: Evaluating LLMs on Generating Interactive and Functional Websites from Scratch (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
