The Big Picture
A shared, verifiable fact graph plus a coordinator and many parallel worker agents enables AI systems to assemble long, detailed mathematical proofs with relatively little human guidance.
ON THIS PAGE
The Evidence
Combining a main coordinator, many parallel proof-search workers, and a single stateless verifier that admits only checked facts into a global fact graph lets the system accumulate long, reliable arguments. The fact graph (a directed graph of verified claims and their dependencies) prevents workers from trampling one another’s progress and turns many short model calls into a coherent long proof. In six research-level case studies the system produced manuscripts in days; human experts provided occasional high-level guidance and performed final checks. Emergence-Aware Monitoring Pattern
Not sure where to start?Get personalized recommendations
Data Highlights
1One nontrivial proof run used five parallel workers and recorded 63 verified facts and 239 failed paths while building the final argument.
2Danus produced full manuscripts for 6 research-level problems (algebraic geometry, singularity theory, combinatorics) in days, whereas a direct call to the stronger baseline model produced no meaningful result for these problems.
3On a benchmark problem, the predecessor single-worker system failed 3 times where the new multi-worker orchestration succeeded.
What This Means
Engineers building AI-driven research assistants and toolchains should care because the fact-graph pattern shows how to combine many short, fallible model calls into long, reliable reasoning. Technical leads and product owners evaluating agent orchestration will find a concrete approach for scaling complex, long-horizon tasks while keeping the shared state auditable for human reviewers. Blackboard Pattern
Key Figures
Fig 1: Figure 1 : The overall architecture of Danus.
Fig 2: Figure 3 : The submit–verify–repair cycle between a worker and the verifier.
Fig 3: Figure 4 : The three kinds of agent, each with its own skills and its own role-gated set of tools.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Proofs still require expert review and occasional human hints to unlock hardest steps; the system is a collaborator, not an autonomous mathematician. The design depends on heavy model use (including high-performance expert calls at low frequency), which affects cost and availability. Results come from six case studies—broader generalization and robustness across domains remain to be validated. Human-in-the-Loop Pattern
Methodology & More
Danus organizes mathematical reasoning around three roles: a main agent that plans and coordinates, many worker agents that propose local claims and supporting proofs in parallel, and a stateless verifier that is the sole authority admitting only fully checked claims into a single global fact graph. The fact graph stores verified facts and their logical dependencies as a directed acyclic graph; workers repeatedly submit candidate statements and proofs to the verifier, which either admits them as facts or rejects them (recorded as failed paths). The main agent periodically reads the graph, summarizes progress, reassigns workers, and can consult a high-end model as a low-frequency advisor. When the target theorem appears among the verified facts, the system runs a writing stage that converts the fact graph into a paper-style manuscript for human experts to finalize. Agent Registry Pattern Tool Use Pattern
In six research-level case studies across algebraic geometry, singularity theory, and combinatorics, the system produced long, detailed proofs built from dozens or hundreds of verified facts and finished manuscripts in days. Parallel workers broaden search and make long multi-step arguments tractable, while the fact graph prevents contradictory or irrelevant intermediate artifacts from confusing the process. The design shows the harness (how models are coordinated and memory is managed) can matter more than any single model call: direct calls to a strong model produced no meaningful results on these tasks, but the orchestrated system succeeded repeatedly. The main trade-offs are cost, the need for focused human guidance on the hardest steps, and the current limited evaluation scope; nevertheless, the fact-graph orchestration offers a practical blueprint for scaling agentic systems on long-horizon reasoning tasks. Tool Use Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Multiple authors with modest h-indexes (mostly <10) and no top-institution or top-venue signals; team work suggests more credibility than solo unknown authors but lacks strong institutional or citation evidence.