AutoGraphForge tests a full pipeline for AI-assisted graph theory discovery
A September 3, 2026 arXiv study links conjecture generation, counterexample search, novelty filtering, Lean formalization, and neural theorem proving.
Graph theory becomes a test bed for an end-to-end discovery workflow
The new arXiv paper AutoGraphForge: Towards Automated Graph Theory Discovery describes an unusually broad mathematical pipeline. Rather than asking one language model to invent and prove a theorem in a single pass, the system separates the work into conjecturing, novelty checking, counterexample search, formalization, and kernel-verified proof attempts.
How the pipeline narrows the search
AutoGraphForge starts with a few hundred graphs and their computed invariants. A Graffiti3-based generator proposes inequalities, while counterexamples are added back into the table for later rounds. Candidates also face a novelty filter built from 559 classical and folklore relations, followed by tests against about 348,000 graphs and several active counterexample-search methods. The authors report 6,522 conjectures that survived those filters and searches, including nontrivial relations that they then proved by hand.
Where neural theorem provers enter
Surviving statements can be translated deterministically into Lean 4 skeletons and passed to DeepSeek-Prover-V2 and OProver-32B. Lean's kernel independently checks any proposed proof against a pinned version of mathlib. That division is important: the neural systems suggest proof code, but they do not get to certify their own answers.
The paper's important limitation
This is a progress report, not a claim that automated graph-theory discovery is solved. The paper says the formalization and proving layer has passed initial sanity checks, but has not yet undergone a controlled evaluation across the surviving conjectures. The full loop is therefore not closed yet. That candour makes the work more useful: it shows exactly which stages have evidence behind them and which remain engineering and research goals.
Why MathsAI readers should care
AutoGraphForge illustrates a practical direction for mathematical AI: use different tools for creativity, refutation, formal expression, and verification. Its most valuable idea may be the architecture rather than any single conjecture. If the proving stage scales, researchers could spend less time sorting obvious, previously known, or easily refuted statements and more time judging the genuinely interesting survivors.
Read and explore
This article is an original MathsAI summary based on the linked paper.