Reproduce “LEDGER: Scaling Agentic Document Editing with Dependency-aware Graph Retrieval” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mike Hang Wang, Utkarsh Garg, Reza Davari, Huitian Jiao, Hao Cheng, Baolin Peng, Tao Ge, Si-Qing Chen
Microsoft researchers developed LEDGER, a framework for scaling agentic document editing by constructing a dependency-aware graph of document units. This approach significantly improved editing consistency by 20 percentage points to an average of 76% across models, while achieving an 88.5% token reduction and maintaining constant token usage for documents of varying lengths.
Key insights
- Explicit external structural representations, such as dependency graphs, can effectively substitute for expensive internal reasoning within large language models, leading to more efficient and reliable agentic systems.
- A hierarchical prioritization of dependency types (explicit references, implicit dependencies, hierarchical, semantic) is crucial for efficiently selecting context that maintains document consistency.
- Agentic document editing can achieve constant (O(1)) context size with respect to document length by leveraging structured dependency graphs, thereby enabling scalability for very long documents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.