Prove2Me
An open collaborative platform for scaling mathematical formalization, designed by tianyi-peng and collaborators at Columbia (Chen, Marwaha, Lu, Yuen & Peng, arXiv 2608.28433, 2026). It is the scaffold that turned Anthropic's FLT attempt from a failure into a proof flt-formalization-anthropic.
What it does
- Keeps a directed acyclic graph of theorem statements that agents read to decide what to attempt next. Anthropic credits this specifically with "mitigating memory degradation and allowing multiple agents to work in parallel."
- Separates statements from proofs into different files, maintaining the links independently, which speeds Lean compilation and cuts resource use.
- Attaches a natural-language description to every statement, so an agent can search for what exists and reuse it instead of re-proving it.
The result it is evidence for
Before Prove2Me, Anthropic's agents "quickly lost track of the project's state and stopped collaborating effectively"; their work survives as roughly 7% of the non-boilerplate lines. After it, dozens of agents completed FLT in eleven days. The same platform let three personal Claude Max subscriptions formalize Vinogradov's three-primes theorem in three days.
So the load-bearing component is a shared external state store with a task graph, not agent
autonomy. That is the same shape this hub keeps finding elsewhere: the document, the graph, the
explicit plan file doing work the model's context cannot. It rhymes with gbrain and with the
harness arguments in ../agentic-tooling-wiki — an agent system's competence is largely a property
of what holds its state.
Not yet read
The arXiv paper itself. This page is built from Anthropic's description of using the platform, which is a user's account and not the authors'. The paper would settle what Prove2Me does for humans rather than for agents, and whether the DAG is a general mechanism or an FLT-shaped one.
Related
flt-formalization-anthropic · autoformalization · tianyi-peng · lean-theorem-prover · gbrain