acceptodds
Under review as a conference paper at ICLR 2027

Lean Pool: An AI-Maintained Archive of Formalized Mathematics

Abstract

As AI systems increasingly contribute to mathematical research, formal proofs make it possible to check arguments before they have been fully examined by mathematicians. Reusing these proofs requires more than preserving their source: formalizations must remain compatible with evolving libraries, and their results must be easy to find and understand. We present Lean Pool, a living archive of mathematical formalizations maintained together as Lean and Mathlib evolve. It combines AI agents, automated checks, and human oversight to maintain independently developed projects while preserving their attribution. We analyze the archive’s contribution and maintenance history, accepted optimizations, mathematical reviews, and evidence of reuse. The operational record shows that agent-assisted maintenance can restore compatibility across dependency upgrades and support library-wide proof shortening and compilation improvements. It also records follow-up repairs after initial automation, the resource demands of large reviews, and tradeoffs between reusable interfaces and compilation cost. Archived mathematics is reused in subsequent research-level formalization. We release the archive, its maintenance workflows, exposition site, and supporting analysis artifacts, aiming to provide a maintained formal counterpart to arXiv.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
formalized mathematics, Lean, AI agents, proof maintenance, mathematical libraries, automated theorem proving

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_L4j4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Lean Pool, a living archive that gives completed formalizations a persistent, searchable home. The paper details the intended contributors, sources, and functionality of the system, and analyzes its evolution and capability to maintain compatibility.

Critical strengths

The system itself is practically useful and serves as a good resource for formal mathematics.

Critical weaknesses

This paper seems to be more of a system demonstration or system card than a research paper. Although ICLR accepts papers on infrastructure and software systems, I find the research contributions of this paper insufficient for an ICLR submission.

Questions for authors

As the paper does not present a sufficiently clear scientific contribution beyond the development and demonstration of the system, I have no further questions.

Reviewer_QnTHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Lean Pool, a living archive of mathematical formalizations maintained as Lean and Mathlib evolve.

Its main contributions are as follows:

  1. Lean Pool brings together formalization projects developed independently by different authors in a common environment, while preserving authorship and source information.
  2. It proposes using Lean Pool to connect citations of mathematical papers with imports of their corresponding formalizations.
  3. Through continued maintenance, Lean Pool provides a lasting, searchable home for completed formalizations and makes them available for reuse.

The paper describes the archive's growth and day-to-day maintenance, including six dependency upgrades and several proof and compilation optimizations. It also presents a mathematical review service, search and exposition tools, and evidence of code reuse from LeanEval and later projects.

Critical strengths
  1. Lean Pool tackles a practical problem: a completed formalization can become difficult to reuse as Lean and Mathlib change. The archive already brings together many independent projects, and the six recorded upgrades show that the authors have kept these projects working over time.
  2. I appreciate the concrete build measurements and examples of later projects using archived results. The paper also reports the extra repair work and explains the limits of its measurements. These details give a useful picture of what maintaining this archive involves.
Critical weaknesses
  1. My concern is how changes to the scope of archived results are handled during maintenance. Appendix D records stronger auxiliary assumptions, including measurability and sigma-finiteness, and explicitly notes that compatibility does not establish equivalence of declaration types. Appendix B also describes statement-change summaries. However, the discussion does not establish whether the added conditions already followed from the main theorems' hypotheses or made any registered main result more restrictive. It also gives little detail about how such changes are reviewed and reflected in project cards. This matters because users rely on those cards to identify results they can build on. The maintenance evidence therefore leaves an important question about the continued availability of the advertised mathematics unanswered.
  2. I appreciate the value of a maintained Lean archive. I would have liked a clearer sense of what this work adds to existing approaches to proof-library maintenance. Several parts of the approach have precedents in the literature. AFP (https://isa-afp.org/submission/) provides a model for maintaining contributed formalizations, while Baanen et al.'s Growing Mathlib (https://arxiv.org/abs/2508.21593) discusses compatibility management, linting, and compilation optimization. Compiler-guided proof repair has also been explored in APOLLO (https://arxiv.org/abs/2505.05758) and APRIL (https://arxiv.org/abs/2602.02990). These works address different settings, and bringing independently developed projects into a maintained common environment is a worthwhile undertaking. Sections 2 and 4 and Appendix C explain how this is organized, and the upgrade records and optimization measurements provide useful evidence of its operation. What I found less developed is whether the challenges of maintaining these projects together called for different technical solutions, or led to insights beyond those already discussed in earlier work. On balance, I see the paper's strongest contribution in the archive and its deployment. The case for a further contribution to maintenance methods or our understanding of the maintenance problem is less convincing to me.
  3. The reuse cases support the claim that archived mathematics enters later work. I would, however, give limited weight to Section 3's 'most reused external repository' ranking when assessing Pool's contribution. Figure 6 counts declaration matches, including copied code, excludes ordinary package imports, and allows the same solution to match Pool and an upstream source. These choices are disclosed. The concern is that a collection of projects can rank above individual repositories because it covers more material, without showing what its shared maintenance adds. Table 24 and the cited agent traces provide evidence beyond these counts, but the paper gives only a brief account of which parts of Pool's maintenance or discovery workflow helped in those cases. This leaves the practical benefits of the maintained common environment less well supported than the fact of code reuse itself.
Questions for authors
  1. For the stronger auxiliary assumptions in Appendix D, did the additional conditions already follow from the hypotheses of the relevant main results, or did any main result become more restrictive? How was the outcome checked and, where necessary, reflected in its project card?
  2. Compared with earlier work on archive maintenance, proof repair, and optimization, what would you highlight as Lean Pool's main research contribution? In particular, did maintaining independent projects together call for different technical solutions or lead to new insights, and which parts of the reported experience best illustrate them?
  3. For the reuse cases in Table 24 and the LeanEval traces cited in Figure 6, what do the records show about how results were found and used through Pool? Do any cases document a specific benefit from Pool's maintenance, shared environment, or discovery tools, beyond the presence of matching upstream code?
Reviewer_iN5YOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors present and analyze a history of LeanPool, an AI-maintained repository of Lean formalizations of research mathematics. The authors intend it to be a formal counterpart to e.g., arXiv, to help share and re-use mathematics.

Critical strengths
  1. The proposed project could be a potentially useful archive of formal mathematics. I can imagine how it's useful for AI autoformalization or theorem proving agents to search through prior results. Staying up-to-date automatically helps avoid version mismatches for dependencies.
Critical weaknesses
  1. Unsubstantiated claims. The core "claim" in this paper is that LeanPool is an archive of formal mathematics which can stay up-to-date by emplying AI agents to update projects to new Lean versions. The authors provide no evidence that this is a current/future problem. Additionally, if AI agents can update projects to a new Lean version, why not do so only when needed? To be clear, there may be a valid reason for maintaining such an archive (e.g., potentially it's more efficient to keep things up-to-date), but the authors provide no such reason or support. In particular, why do previous archives not achieve the goals of LeanPool?

    Other claims are also unsupported. Take section 5 for example. The authors claim that a shared archive permits improvements across independent projects, but this is not supported in any way - there is no comparison to improvement attempts across non-shared archives, etc. Additionally, section 3 seeks to claim that this archive encourages reuse of existing code, but makes only one citation without any relevant statistics or analysis. Again, was this not possible with separately maintained archives (i.e., using websearch to find relevant libraries)?

  2. At a higher level, I am struggling to understand the purpose of this paper/project in context of the existing field. To some extent, I understand the intended purpose of LeanPool, but the paper does not clearly define the gap(s) this system fills in context of prior work and existing methods. Table 9 makes some attempt at placing this work, but stops short of a proper comparison. While it compares the contribution and maintenance policies of each archive, it does not detail the ramifications of those choices, and why LeanPool's particular choices fill a gap that others leave open. I believe more clear comparisons and, where needed, experimental evidence is necessary.

In general, I believe this paper is very disorganized, and does not clearly argue why the proposed system is useful compared to prior systems, or necessary in the face of current problems. There are a number of sections describing LeanPool (e.g., 3, 5, 8), but their connection to the main claims of the paper is unclear. I believe this can in part be solved with a careful rewrite, though will likely need additional experimental evidence/statistical data to support the claims.

Questions for authors

Have you learned anything surprising or new from your time maintaining/working with LeanPool? If so, I think incorporating this into the paper would be interesting for readers.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.