acceptodds
Under review as a conference paper at ICLR 2027

Towards Lean-grounded Mathematical Reasoning

15 views

Abstract

Formal verification and informal mathematical reasoning in LLMs have largely evolved independently. Informal outcome-level supervision provides limited guidance on how rigorous proofs should be structured, while invoking proof assistants such as Lean at inference time incurs recurring formalization costs. To address this, we propose Lean-grounded Learning, which uses Lean as privileged training-time supervision to improve informal reasoning without requiring Lean at inference. A teacher first generates an informal proof, then revises it using a verified Lean proof. A student is subsequently trained on these Lean-grounded revisions via SFT. On Qwen3-30B-A3B, Lean grounding improves AIME24 accuracy from 45.0 to 55.0 and ProofNet from 72.1 to 80.9 over the baseline with natural-language privileged information. With only 30K Lean-grounded examples, Qwen2.5-Math-7B-Base surpasses Qwen2.5-Math-7B-Instruct on MATH, AMC, and Olympiad Math, despite the latter being trained on roughly 2.5M examples followed by RL. Experiments show that Lean-grounded models more often derive claims from established facts, whereas Qwen3-30B-A3B more often relies on candidate testing and extrapolation from concrete cases.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Formal Reasoning, LLM

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ncn2OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates a long-standing question in the reasoning community: Can formal reasoning capabilities benefit informal reasoning? Although the paper does not employ complex training methods, it provides a convincing answer to this question, supported by extensive experimental designs and insightful analyses of related research questions.

First, the paper develops a method for leveraging Lean-verified formal reasoning trajectories for training, demonstrating that even simple SFT can improve a model's reasoning capabilities. Furthermore, by comparing Lean reasoning trajectories with natural-language solution trajectories, the authors confirm that the improvement comes from the logical structure of formal reasoning, rather than merely from distillation using stronger models or privileged information.

Second, the stability study demonstrates that these findings hold across problems of varying difficulty levels and that the improvements can scale with larger pass@k sampling budgets.

Finally, the paper experimentally validates several interesting findings. For example, the proposed method not only improves model capabilities but can also provide a better initialization for subsequent RL training. The paper also compares the efficiency of using Lean during training versus inference, and the proposed approach, to some extent, addresses the long-standing challenges of high inference costs and low efficiency associated with formal reasoning.

Critical strengths
  1. The paper is very clearly written and well presented.
  2. The proposed method is simple yet effective.
  3. The experimental design and analysis are very well conducted, providing many interesting and insightful findings.
Critical weaknesses
  1. The base models used for SFT in this paper are primarily from the Qwen3 series, rather than stronger models such as Qwen3.5 or Qwen3.8. We would be very interested to see whether the proposed method can still effectively improve the reasoning capabilities of state-of-the-art models. If so, this would make the experimental findings even more exciting.

  2. The paper only explores using Lean reasoning trajectories for SFT, without considering other training methods. However, I do not believe this limitation undermines the soundness of the experimental design or the completeness of the experimental analysis.

Questions for authors
  1. Have the authors tried applying the proposed method to SFT on stronger base models, such as Qwen3.5 or Qwen3.8, which have already undergone substantial post-training? We are curious whether the method can still provide meaningful improvements on these models.

  2. Have the authors considered training with OPSD, using Lean proofs as privileged information for the teacher model? How would OPSD compare with the SFT approach proposed in this paper? Which method would perform better?

  3. The training data primarily comes from Nemotron-Math-Proofs-v1, where Lean proofs are generated by Goedel-Prover. We wonder whether this dataset exhibits any domain or difficulty biases. As we know, formal provers with Lean verifiers tend to perform well in certain mathematical domains (e.g., linear algebra), while being relatively weaker in others (e.g., geometry and combinatorics). Does the training dataset exhibit such domain biases? Regarding difficulty, does the difficulty distribution of the training problems selected through this process differ from the distribution of problems that the student model itself is capable of solving? If these two types of bias exist, how do they affect model training and downstream reasoning performance?

4 comments

Thanks for your insightful suggestions. We have run our pipeline with a stronger student, added an OPSD variant in which the Lean proof serves as privileged information, and analyzed the domain and difficulty bias of the training data.


W1 & Q1: Results on stronger students

We applied Teacher Lean to Qwen3.5-9B, using the identical 10K Lean-grounded dataset generated by Qwen3.6 and SFT reciepe in our paper. We report Pass@16 for AIME24—26, Pass@8 for IMOAnswerBench on Qwen3.5-9B based models.

AIME24AIME25AIME26IMOAnswer
Qwen3.5-9B90.073.380.050.7
Qwen3.5-9B SFTed via Lean Teacher90.080.090.053.3

Lean grounding still improves the stronger student: +6.7 on AIME25, +10.0 on AIME26, and +2.6 on IMOAnswerBench.

W2&Q2: OPSD with Lean proofs as privileged information

We discussed OPD in Appendix A.4 as future work, and we have now run it. The Qwen3-8B student samples proof on-policy from the natural-language problem, and a frozen copy of the same model, conditioned on the verified Lean proof ρ\rho as privileged information, scores the student’s own completion using soft labels (exactly OPSD), the objective is the exact full-vocabulary reverse KL between the two distributions.

AIME24AIME25AIME26MATH500IMOAnswer
Qwen3-8B26.519.216.084.816.4
OPSD (Lean)30.424.617.584.416.8
Teacher Lean (SFT)40.831.732.987.116.9

On Qwen3-8B, we observe that SFT outperforms OPSD with Lean proof reference, while OPSD outperforms base models on AIME24, 25, 26, IMO-AnswerBench. A likely reason lies in the teacher: in OPSD, the teacher is a frozen copy of Qwen3-8B itself, whose ability to understand Lean 4 code is limited, whereas the SFT targets are produced by the stronger model (Qwen3.6) in understanding Lean.

Q3. Domain and Difficulties Bias and Resulting Downstream Behavior

Domain bias and Lean-grounding’s generalization to sparse domains. We drew a random sample of 250 problems uniformly from the 10K training set and labeled the mathematical domains of each sample using Claude Opus 5.

Table 1: Domain distribution of our sampled training data.

DomainPercentage
Algebra64.4%
Number theory13.6%
Geometry9.2%
Analysis9.2%
Combinatorics2.0%
Linear algebra1.6%

We observe that algebra accounts for the largest portion (64.4%), while combinatorics (2.0%) and linear algebra (1.6%) are the smallest. Table 2 splits AIME24--26 performance of the Qwen3-30B-A3B variants by sub-domain. On the domains with a large portion, Lean grounding improves over the base model by +17.1 points on algebra and +37.5 points on number theory, on the sparse domains, by +21.5 points on geometry and +18.8 points on combinatorics, showing that our method also generalizes to domains with few training examples.

Table 2: Pass@1 accuracy (%) by subdomain on AIME24—26 (90 problems) for Qwen3-30B-A3B. Our method generalizes to domains with sparse training data.

Domain#problemsQwen3-30B-A3BTeacher CoT (Student: Qwen3-30B-A3B)Teacher Lean (Student: Qwen3-30B-A3B)
Algebra1541.251.758.3
Combinatorics3020.232.539.0
Geometry3017.938.339.4
Number theory1524.646.262.1
All9023.739.946.2

Difficulty bias and consistent gains at every difficulty level.

In our work, we constructed the training set by sampling uniformly from Nemotron-Math-Proofs-v1 after decontamination, with no preference for difficulty. To measure the difficulty distribution, we use Claude Opus 5 to assign a difficulty score from 1 to 10 for 100 problems sampled uniformly from the 10K training set, for all 90 problems of AIME24--26, and for 50 problems sampled uniformly from IMO-AnswerBench. The judge sees only the problem statement.

Table 3: Difficulty distribution of the training data and of the benchmark problems, with Pass@1 accuracy (%) at each level. We also show the comparison between our Lean-grounded method with Teacher CoT and the Base model.

DifficultyTraining data (%)Benchmark (%)Qwen3-30B-A3B (Base)Qwen3-30B-A3B (Teacher CoT)Qwen3-30B-A3B (Teacher Lean)
≤315%1%100.0100.0100.0
415%13%67.783.796.5
529%26%24.337.247.1
634%30%6.218.820.8
≥77%30%17.919.621.7
All100%100%23.132.838.3

We find that the difficulty distribution of sampled problems sits at the boundary of what the student can already solve, and the gains do not come from training on problems harder than the benchmark contains. Teacher Lean has the highest accuracy at every level, improving over the base model by +28.8 points at level 4, +22.8 at level 5, +14.6 at level 6, and +3.8 at level 7 and above. In our paper (Figure 3(b)), we have already found similar pattern of Lean-grounded method improves against all base solve rate on AIME24—26.

I thank the authors for the additional experiments, which provide stronger evidence of the method's effectiveness. Although OPSD does not outperform SFT, it still brings meaningful improvements.

The cross-domain and difficulty-level analyses are particularly insightful. They suggest that learning from Lean solutions benefits reasoning beyond algebra or simpler problems, generalizing across different domains and difficulty levels. I interpret this as evidence that LLMs can acquire a more rigorous, Lean-style reasoning pattern.

Overall, I believe this paper offers valuable insights, and I would like to raise my rating from 3 to 4.

AuthorsOpenReview ↗

Thanks for recognizing our analyses insightful

Thank you for finding our additional experiments insightful. This is really encouraging for us.

Reviewer_pxMHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Lean-grounded learning, which uses verified Lean 4 proofs as privileged information during teacher-generated training-data construction. The teacher first independently drafts an informal proof and then revises it by consulting the formal derivation; the student is trained only on the revised outputs via SFT and has no access to Lean at inference time. Controlled experiments show that, compared with privileged information such as natural-language trajectories, Lean grounding yields consistent improvements on both answer-seeking and proof-generation tasks—for example, improving Qwen3-30B-A3B from 45.0 to 55.0 on AIME24 and from 72.1 to 80.9 on ProofNet. With only about 30K samples, it also enables Qwen2.5-Math-7B-Base to surpass its Instruct counterpart trained with roughly 2.5M samples plus RL.

Critical strengths
  1. The main controlled comparisons are valuable: all conditions use the same 10K problems, teacher, and SFT setup; conditions requiring revision also share the same initial draft and are revised only once. The results show that the gains do not merely come from giving the teacher an additional revision pass.

  2. The method leverages formal proofs during training while retaining ordinary natural-language inference at deployment, without requiring access to a verifier. Experiments cover multiple student models and tasks, and case studies further show that revision can preserve useful reasoning from the initial draft rather than simply translating Lean.

Critical weaknesses

The most important missing control is to have the same teacher directly generate a natural-language proof from the problem x and the verified Lean proof ρ , use it to train the student, and skip the initial draft. The submitted manuscript does not report this baseline. Teacher CoT does not have access to ρ , while Traj-Guided uses the natural-language trajectory accompanying the prover to revise the initial draft. The current results support the claim that Lean references can improve supervised data, but they do not yet establish the additional value of the draft-then-revise procedure. If a fair comparison shows that the simpler direct-generation baseline performs comparably, then the advantages of these additional stages have not been established, which would substantially weaken the contribution of the proposed pipeline.

Questions for authors

See weakness, maybe you can add an experiment about it?

I may consider raising my score if the experiment results supports the effectiveness of your pipeline.

6 comments
AuthorsOpenReview ↗

The Lean-grounded pipeline outperforms method that generate NL proof directly from Lean

Thanks for your insightful suggestions. We add the direct-generation baseline suggested by the reviewer: the same teacher (Qwen3.6) receives the problem xx and the verified Lean proof ρ\rho and produces a natural-language proof in a single pass. We then SFT Qwen3-8B on these proofs under the same student training settings as in the main experiments, and denote this condition Lean Direct in the following experiments.

Numerical results. The table below compares Lean-Direct with our ``draft-then-revise'' method. Each entry reports Pass@1 on Qwen3-8B.

AIME24AIME25AIME26MATH500IMOAnswer
Qwen3-8B26.519.216.084.816.4
Lean Direct19.417.914.877.68.2
Ours40.831.732.987.116.9

We observe that Lean Direct method underperforms our draft-then-revise pipeline. As the reviewer notes, Teacher Self-Refine (Traj-Guided) already shows that the gains do not come from the additional revision pass alone. The present result adds the complementary control: access to ρ\rho without a draft does not help either, and in fact falls below the untrained student. The improvement therefore comes from revising a draft against ρ\rho, not from either component on its own.

One follow-up on evaluation. In A.1.1 you cited Mahdavi et al. (2025) and Petrov et al. (2026), but those agreement studies are not for yours right? Did you yourselves perform any human validation of this judge on your own data (both GPT-OSS-120B and Opus 5) ?

We cite Mahdavi et al. (2025) and Petrov et al. (2026) to show that LLM judges are reasonably consistent and reliable for proof evaluation. In the main paper, we use the LLM judge only for the proof-generation tasks in Section 3.3. These tasks follow a setting aligned with Mahdavi et al. (2025) and Petrov et al. (2026), so we adopt GPT-OSS-120B as the judge directly and grade all methods with the same judge for a controlled and fair comparison. We have not performed human validation on our own outputs. Would the reviewer suggest adding a human validation experiment? If so, we will do our best to add one before the end of the discussion period.

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

The paper proposes Lean-grounded learning: verified Lean 4 proofs are used as privileged information for a teacher model, and a student model is then fine-tuned on the revised informal proofs. At inference time, the student model has no access to Lean. The experimental design is a controlled comparison among four supervision conditions built from the same 10K problems. The results show the effectiveness of the proposed solution on a variety of tasks and cases.

Critical strengths
  • The paper is easy to follow. The motivation is clear and well positioned.

  • Extensive experiments are conducted. The paper covers answer-seeking and proof generation tasks, multiple student models and teacher models, and several variants considering RL initialization, thinking mode, etc.

  • The controlled supervision comparison clearly supports the claimed contribution.

Critical weaknesses

RQ1 is not really an open question. RQ1 asks "can the deduction structure of Lean proofs transfer to LLMs’ informal proof generation without access to Lean at inference time?" The answer to this question is quite obvious, has been widely studied in the more general domain of math proving/solving. A possible meaningful question can be quantitative. For example, how much improvements can be achieved; which sub-domain is especially beneficial?

No quantitative metric for structure transfer. Related to the previous point, the paper's central scientific claim is that Lean grounding transfers structure, but the evidence for this is almost entirely qualitative. Without such metrics, the paper cannot show how much structure transfers, whether the transfer persists at scale, or whether it correlates with the benchmark gains.

Teacher and student are from the same model family. Main experiments use Qwen3-8B and Qwen3-30B-A3B as students, and Qwen3.6-35B-A3B as the teacher. Including a cross-family teacher in the main experiments would be helpful to evaluate the generalizability of the proposed method.

Questions for authors
  1. Structure-transfer metrics. Can you operationalize Definition 1 into automatic metrics? This would convert the paper's central claim from an illustration into a measurement.
  2. How much of the gain survives with a cross-family teacher for the main students (not only the weaker ones in Section 3.4)?
3 comments

W1, W2 & Q1: Lean grounding improves structure score

We thank the reviewer for insightful suggestions. We turn Definition 1 into an automatic metric.

Metric. We use Claude Opus 5 as a judge. Given a problem with its proof yy, it (1) splits the proof into claims c1,…,cnc_1,\dots,c_n, (2) for each claim cic_i, lists all previous claims that it relies on, (3) labels each cic_i as sound (it satisfy derivation and dependency discipline in Def 1) or not sound. Let A(y)A(y) be the set of all claims. The structure score is defined as: S(y)=1∣A(y)∣∑i∈A(y)1[ci  and all claims it depends on are sound]S(y)=\frac{1}{|A(y)|}\sum_{i\in A(y)}\mathbb{1}\big[c_i\; \text{and all claims it depends on are sound}\big] To ensure fairness, the judge is blind to which model produced the proof, and work in independent sandbox.

Structure transfer. Table 1 reports the structure score by subdomain, pooled over AIME24, AIME25, AIME26 and IMO-ProofBench. The model fine-tuned on Teacher Lean data has the highest average structure score. It outperforms the other baselines in algebra, combinatorics, and number theory, showing its advantage of structure transfer in these domains. Same as our paper, we adopt Qwen3.6 as teacher.

Table 1. Structure score (%) by subdomain on problems mixed with AIME24—26 and IMOProofBench on Qwen3-30B-A3B variants.

DomainQwen3-30B-A3BTeacher CoT (Student: Qwen3-30B-A3B)Teacher Lean (Student: Qwen3-30B-A3B)
Algebra46.659.064.1
Combinatorics29.944.846.8
Geometry37.448.346.6
Number theory45.456.064.1
All39.051.254.2

Which subdomains benefit most. Table 2 reports Pass@1 accuracy by subdomain on AIME24, AIME25 and AIME26. Lean-grounded supervision gives the highest accuracy in every subdomain, and the average ranking agrees with structure score in Table 1 (Base < Teacher CoT < Teacher Lean). Its largest gain over Teacher CoT is in number theory (+15.9), followed by algebra (+6.6) and combinatorics (+6.5). In geometry the gain is small (+1.1).

Table 2. Pass@1 accuracy (%) by subdomain on AIME24—26 (90 problems) for Qwen3-30B-A3B.

Domain#problemsQwen3-30B-A3BTeacher CoT (Student: Qwen3-30B-A3B)Teacher Lean (Student: Qwen3-30B-A3B)
Algebra1541.251.758.3
Combinatorics3020.232.539.0
Geometry3017.938.339.4
Number theory1524.646.262.1
All9023.739.946.2

W2 & Q2: Lean grounding also helps with a cross-family teacher

We repeat the controlled comparison with Muse-Glimmer-30B-GGUF as cross-family teacher, using the identical generation pipeline, the same 10K problems and the same student recipe. The student is Qwen3-8B. Each entry reports Pass@1 accuracy.

Table 3. Lean-grounded supervision with a cross-family teacher (Muse) and the same-family teacher (Qwen3.6).

AIME24AIME25AIME26MATH500IMOAnswer
Qwen3-8B26.519.216.084.816.4
Teacher CoT (Student Qwen3-8B, Teacher: Muse)32.519.226.784.018.2
Teacher Lean (Student Qwen3-8B, Teacher: Muse)35.425.428.388.519.7
Teacher CoT (Student Qwen3-8B, Teacher: Qwen3.6)32.525.823.382.616.7
Teacher Lean (Student Qwen3-8B, Teacher: Qwen3.6)40.831.732.987.116.9

We show that the benefits of Lean grounding persist when using a cross-family teacher, Muse.

AuthorsOpenReview ↗

Thanks for your review

Dear Reviewer ZFTp,

Thank you again for your insightful and constructive feedback. Since the discussion period is ending, we would be grateful if you could take a look at our extra experiments and analysis when convenient.

Best,

The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their time and careful review. Reviewers found our work simple yet effective, the experimental design and analysis well conducted, and the findings interesting and insightful (ZFTp, pxMH, ncn2). Following insightful suggestions from the reviewers, we summarize the points and answer each reviewer separately.

  1. Turning Def.1 into Automatic Metrics (ZFTp). We transform Def.1 into automatic metrics for judging structure score with Opus 5, and show our method outperforms baselines under structure score.
  2. Lean-grounded learning is effective with cross-model teachers (ZFTp). We show that the gains for Lean-grounded learning survive for cross-model teachers to train the Qwen3-8B student with the Muse teacher.
  3. Add a baseline that the teacher generates NL proof directly from Lean (pxMH). We show the direct Lean translation baseline underperforms Lean-grounded learning, the improvement therefore comes from revising a draft against Lean proof, not from either component on its own.
  4. Add stronger students (ncn2). We train Qwen3.5-9B with the same pipeline and show Lean grounding improves accuracy on AIME24--26 (Pass@16) and IMO-AnswerBench (Pass@8).
  5. Add OPSD training results on Qwen3-8B with Lean privileged information (ncn2).
  6. Discuss the emergent behavior with the domain and difficulty bias of the training data (ncn2). We show that the gains hold in the sparse domains and at every difficulty level for the Lean-grounded learning method.

Reviewers have increased their scores accordingly: pxMH (2 -> 3), ncn2 (3->4), and we address ZFTp’s concerns with new experiments in points 1 and 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.