Reproduce “Danus: Orchestrating Mathematical Reasoning Agents with Fact-Graph Memory” 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
Jihao Liu, Guoxiong Gao, Zeming Sun, Bin Wu, Shurui Liu, Jiedong Jiang, Haocheng Ju, Leheng Chen and 3 more
Danus presents an orchestration system for large language model-based mathematical reasoning agents, employing a shared fact-graph memory to manage and coordinate parallel proof generation. The system successfully resolved six research-level mathematical problems, autonomously decomposing tasks and building proofs with thousands of verified facts and dependency chains up to 54 steps deep.
Key insights
- Orchestration via a shared fact graph provides a robust mechanism for coordinating numerous mathematical reasoning agents and managing complex logical dependencies.
- A stateless verifier, acting as the sole authority on mathematical correctness, is crucial for building reliable, long-form proofs and preventing error propagation.
- The system's architecture allows for autonomous problem decomposition and synthesis of diverse mathematical techniques, often leading to unanticipated solution paths.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.