acceptodds
Under review as a conference paper at ICLR 2027

Lean2Isabelle: Verifier-Grounded Cross-Assistant Proof Translation

Abstract

Formal proof artifacts are usually tied to the assistant in which they were written. We study theorem-level Lean-to-Isabelle migration as a verifier-grounded learning problem: given a Lean theorem and proof, construct an Isabelle/HOL theory that passes Isabelle/PISA checking and an audited semantic evaluation against the source theorem. Lean2Isabelle factorizes migration into target-statement alignment and statement-conditioned theory reconstruction, and uses Minimal Valid Progress (MVP), a bounded Isabelle proof-context signal, to supplement sparse pass/fail rewards. We construct 26,104 complete-theory Herald-ISA examples and 9,683 LeanWorkbook-ISA statement-level examples, and evaluate external Lean inputs in MiniF2F-DSP. On LeanWorkbook-ISA, the full-input cascade reaches 39.1% end-to-end SemPass@1, while a statement-only frozen structural translator followed by Isabelle automation reaches 24.3%. With a verified target statement, GRPO-MVP reaches 54.3% SemPass@1, improving over Theory-SFT by 4.5 percentage points. Removing tactics, shuffling proof steps, or retaining only the Lean statement lowers Reference-Statement SemPass@1 to 46.9%, 46.1%, and 43.2%, respectively. On MiniF2F-DSP, the system reaches 85.7% statement accuracy and 24.0% end-to-end SemPass@1 without a reference Isabelle target at inference. These results establish a measurable theorem-level interface for cross-assistant proof migration and identify target-statement alignment, target-context reconstruction, and source-proof conditioning as distinct empirical bottlenecks.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Formal theorem proving, cross-assistant proof translation, Lean, Isabelle/HOL, verifier-guided learning, reinforcement learning, semantic preservation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ic3eOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Lean2Isabelle, a framework for translating Lean theorems and proofs into Isabelle/HOL. It separates statement translation from proof reconstruction and uses supervised fine-tuning followed by verifier-guided RL with a Minimal Valid Progress (MVP) reward.

The authors construct two datasets, Herald-ISA (26K examples) and LeanWorkbook-ISA (9.7K examples), and evaluate on MiniF2F-derived problems as well. The system achieves 54.3% SemPass@1 with reference Isabelle statements and 39.1% end-to-end on LeanWorkbook-ISA. The main idea is to make cross-assistant proof translation verifiable while explicitly measuring semantic drift.

Critical strengths

I really like the motivation behind this work :) As more mathematics gets formalized in different proof assistants, finding ways to reuse proofs across systems seems increasingly important. I particularly appreciate that the authors distinguish between generating a valid Isabelle theorem and actually preserving the meaning of the original Lean theorem.

The dataset construction is also a valuable contribution. I liked the source-proof ablations, especially the observation that removing tactics or shuffling proof steps significantly reduces performance. This provides useful evidence that the model is learning something from the original Lean proof rather than simply proving the target Isabelle statement independently.

Critical weaknesses

1. Semantic preservation is still difficult to guarantee.

My biggest concern is how we actually know that the translated Isabelle theorem means the same thing as the Lean theorem. Isabelle checking alone cannot establish this, so the authors use an LLM-based semantic judge. I appreciate the manual audits, but Appendix B reports that 92.5% of sampled judge-rejected Reference-Statement candidates were actually considered valid by human reviewers. This makes me wonder how much the reported success rates depend on the judge's decisions.

2. The improvement from MVP seems relatively small.

I find MVP interesting, but closing proof contexts does not necessarily mean making mathematical progress :) In the equal-budget experiments, MVP improves SemPass@1 from 51.0% to 52.3% over verifier-only GRPO. I would have liked stronger evidence that this reward correlates with actual progress and cannot be exploited through unnecessary intermediate proof steps.

3. Stronger baselines and larger migration experiments would help.

The structural translation baseline is fairly limited. I would be interested in comparing against an LLM agent with Isabelle execution, library search, and iterative repair. The 10-theorem matrix pilot is encouraging, but it remains unclear how well the approach scales to larger libraries with shared definitions and dependencies.

Questions for authors
  1. Semantic equivalence: How sensitive are the SemPass results to errors in the LLM judge? Could some translations be checked through stronger, possibly machine-verified correspondence conditions?

  2. MVP reward: Does closing Isabelle proof contexts reliably predict eventual proof success? Could a model game this reward by proving easy but irrelevant intermediate lemmas?

  3. RL improvements: The equal-budget MVP gain is only 1.3 percentage points. Is this consistent across multiple training seeds, and is it statistically significant?

  4. Stronger baselines: How would the system compare against a general-purpose LLM agent with Isabelle/PISA access, lemma retrieval, and iterative feedback? I think this would be a particularly interesting comparison :)

  5. Library migration: Could the authors evaluate a larger dependency-connected collection of Lean theorems? I am curious how the system handles shared definitions, typeclasses, and reused lemmas.

  6. Source-proof information: Which aspects of the Lean proof contribute most to translation success: tactic names, intermediate statements, or dependencies on existing lemmas?

Reviewer_48eXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the translation of mathematical proofs from Lean to Isabelle/HOL. The paper leverages MVP (Minimal Valid Progress) as a signal to measure the success in translation to Isabelle. It goes on to develop its own data both in Lean and Isabelle/HOL and studies its MVP based measure. Experiments are on 243 + 262 + 217 problems from LeanWorkbook-ISA, Herald-ISA, and MiniF2F-DSP. The paper reports that its MVP measure can meaningfully measure the correct translation of proofs.

Critical strengths

The topic is interesting and can be potentially useful, especially for the members of community that are working with the Isabelle/HOL system and might need automated translation of proofs from Lean to their system. Hence, it has the potential to be a useful tool.

Critical weaknesses

I was expecting to see experiments on ITPEval [1] which is a benchmark specifically designed for translation between various formal languages including Lean and Isabelle/HOL. It would have also been informative to experiment on PutnamBench.

Literature review does not even mention ITPEval as a related work which was presented at the 3rd AI for Math Workshop at ICML 2026 and now accepted at NeurIPS 2026. The paper seems to be unaware of it not mentioning and also not using it in the experiments.

  • [1] Wu, J., George, R.J. and Anandkumar, A., 2026. ITPEval: Benchmarking Formal Translation Across Interactive Theorem Provers. arXiv preprint arXiv:2607.19407. 3rd AI for Math Workshop at ICML 2026 and now at NeurIPS 2026.

The contribution is unclear to me. I do not see much value in the MVP-based measure alone unless it is shown that it can help improve the correct complete translation of proofs as a reliable tool. The experiments on incomplete proofs may provide some insights, but, in my view, such insights are useful if they are leveraged to improve complete and accurate translations.

In my view, experiments are weak and perhaps not focused enough on improving the translation accuracy, rather focused more on the MVP-based measure.


The writing and presentation is not clear. I had to put plenty of effort to see through what the paper has actually done, its implementation, results, and what it claims to be its contribution. The abstract and much of the paper is more focused on the data while not clearly explaining the method and implementation structures. Much of the main body of the main body of the paper deals with generalities while some critical information are in the appendices or not mentioned.


Some references that authors might want to consider for their literature review.

  • Obua, S. and Skalberg, S., 2006, August. Importing hol into isabelle/hol. In International Joint Conference on Automated Reasoning (pp. 298-302). Berlin, Heidelberg: Springer Berlin Heidelberg.

  • Bergeron, G., Glondu, S. and Tourret, S., 2024. HOL Light to Isabelle/HOL Translation, Rebooted.

  • Vaishnav, R., 2025, November. Lean4Less: eliminating definitional equalities from lean via an extensional-to-intensional translation. In International Colloquium on Theoretical Aspects of Computing (pp. 202-219). Cham: Springer Nature Switzerland.

  • Blanqui, F. and Vaishnav, R., 2026. Encoding Lean's Type Theory in Dedukti. arXiv preprint arXiv:2609.24604.

  • Rothgang, C., Rabe, F. and Benzmüller, C., 2023, July. Theorem proving in dependently-typed higher-order logic. In International Conference on Automated Deduction (pp. 438-455). Cham: Springer Nature Switzerland.

Questions for authors

Please see weaknesses. I'd be happy to see any additional results or clarifications.

Reviewer_LuN6OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper trains a translation model from lean statements (and optionally proofs) to Isabelle theory. The translation process is decomposed into statement translation and proof generation. GPT generates (filtered) training samples for both stages. Then two models are responsible for two stages, respectively. They are both trained with SFT on GPT trajectories. Thereafter, the prover model is trained with reinforcement learning, with a partial reward model for intermediate progress. The resulting algorithm is empirically validated on LeanWorkbook, Herald, and MiniF2F.

Critical strengths

The paper shows a clear effect of an existing source proof for correct proof translation, even though the existence of source proof is optional for the task. This is a reasonable and interesting empirical result, which is firmly supported by sufficient ablation studies that incrementally remove information from the source proof, and observes degrading performance.

Critical weaknesses

The motivation is unclear to distill GPT's translation capability into another model. The distillation techniques are standard (teacher forcing + rl), though a partial credit reward is designed but the ablation study shows relatively marginal improvement. The performance of the distilled model is similar to GPT. So why not use GPT directly, considering the budget for collecting training samples from GPT, and the compute to train the smaller model. Admittedly, the distilled model will become more cost-efficient if we have infinite number of samples to translate. But I suppose the number of test samples should be very large when the combined cost of distillation and inference is smaller than directly calling GPT.

Questions for authors

Can we get a comparison of actual cost of calling GPT API directly to solve the evaluation suites, versus the cost of full training and deployment pipeline: collecting training samples from GPT, training the model, and inference.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.