acceptodds
Under review as a conference paper at ICLR 2027

daVinci-Prover: Agent-Native Mid-Training for Formal Theorem Proving

22 views

Abstract

Advances in large language models (LLMs) are rapidly transforming formal theorem proving. Yet further progress remains constrained by the lack of a transparent and effective data recipe for systematically improving theorem-proving models. In this work, we argue that effective agentic provers require strong mathematical skill priors to ground theorem proving and agentic trajectories to enable long-horizon, interactive proof construction. To this end, we synthesize an agent-native 33.7B-token mid-training corpus covering decomposition, search, self-verification, and revision, together with verifier-grounded agentic trajectories. We further introduce the daVinci-Prover harness, which enables models to construct proofs and interact with a Lean verifier in an isolated workspace, and use it to collect approximately 36K high-quality agent trajectories for supervised fine-tuning. Our mid-training recipe yields an average improvement of 13.7 points over the matched Qwen3-8B SFT-only baseline across MiniF2F, FATE-M, and ProofNet. When applied to the Qwen3.5 series, our approach also yields strong performance across theorem-proving benchmarks and achieves near-saturating MiniF2F performance with memory compaction. The results highlight agent-native mid-training as an effective foundation for developing stronger interactive theorem-proving agents.

Primary area:
datasets and benchmarks
Keywords:
Mid-training, Formal Prover, Agentic Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_BHWcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces daVinci-Prover, a framework for interactive theorem proving that combines a 33.7B-token mid-training corpus with approximately 36K agent trajectories for supervised fine-tuning. The corpus covers proof decomposition, search, revision, self-verification, and tool interactions. It also introduces a Lean-based agent harness and reports consistent improvements over SFT-only baselines across multiple models and theorem-proving benchmarks.

Critical strengths
  1. A well-motivated training recipe for interactive theorem proving. The paper combines mathematical skill supervision—decomposition, search, revision, and self-verification—with environment-native agent trajectories. This design directly addresses the gap between learning from static proofs and interacting with a verifier during proof construction.

  2. Broad empirical evaluation with consistent downstream gains. The approach is evaluated across multiple model backbones, model sizes, and theorem-proving benchmarks. Improvements over the corresponding SFT-only baselines, together with component ablations and data-scaling experiments, provide evidence for the practical utility of the overall training recipe.

Critical weaknesses
  1. The contribution of the synthesis strategy is confounded with teacher-model capability. The proposed pipeline relies on strong teacher LLMs to generate training data, but the experiments do not adequately distinguish the benefit of the synthesis strategy from that of distilling the teacher’s capabilities. The observed gains could therefore largely reflect the quality of teacher-generated supervision rather than the specific data construction method. To isolate the methodological contribution, the authors should compare alternative synthesis strategies using the same teacher, source problems, and comparable generation and training budgets. Evaluating the proposed pipeline with teachers of different capability levels would further clarify its robustness and dependence on teacher quality.

  2. Limited comparative evaluation of the agent harness. The paper presents the daVinci-Prover harness as a contribution, but does not directly compare it with alternative harnesses. Table 3 primarily evaluates the effect of adding retrieval within the proposed harness, which does not establish the effectiveness of the harness design relative to existing alternatives. A comparison with another Lean-oriented harness, or general-purpose coding-agent harnesses such as Codex or Claude Code equipped with Lean verification, would strengthen the evaluation. These comparisons should use the same underlying model where feasible, comparable tool access, and matched inference budgets to isolate the contribution of the harness itself.

  3. Insufficient reporting of data synthesis costs. Although data construction is a central contribution, the paper does not report the wall-clock time or computational and monetary costs required to construct the corpus. The teacher models and final corpus size are specified, but the total generation budget, including discarded samples, failed attempts, and Lean verification overhead, remains unclear. Moreover, the reported training hardware does not characterize the cost of data synthesis. A breakdown of these costs by data component would help assess the practical reproducibility and cost-effectiveness of the proposed pipeline, particularly when comparing it with alternative synthesis methods.

Questions for authors
  1. Directly compare with a data synthesis baseline.
  2. Change a weak teacher model to show the effectiveness of the proposed method.
3 comments
AuthorsOpenReview ↗

Response to Reviewer BHWc

Thank you for acknowledging our well-motivated data recipe and broad empirical evaluation.

【W1 & Q1】 The contribution of the synthesis strategy is confounded with teacher-model capability. Directly compare with a data synthesis baseline.

Our central empirical finding is that incorporating agentic trajectories during mid-training improves subsequent adaptation to interactive theorem proving. We compare skill-only, agent-only, and mixed-data configurations for Qwen3.5-9B under the same downstream SFT, with skill-only training serving as a baseline for assessing the contribution of agentic trajectories. Our skill and agent data are generated using the same teacher.

Mid-trainingMT tokensMiniF2FFATE-MProofNet
Base—87.488.442.9
Skill-only25.3B87.889.843.6
Expanded skill-only33.7B88.989.243.4
Agent-only8.4B88.892.243.5
Skill + agent33.7B90.193.344.5

Conclusion:

  1. Skill-only mid-training improves performance over the no-mid-training baseline. This supports the usefulness of the resulting skill corpus.
  2. At the same 33.7B-token budget, the skill–agent mixture outperforms expanded skill-only training. This supports the value of incorporating agentic trajectories beyond simply increasing training volume. Notably, agent-only training achieves performance comparable to the skill-only baseline with just 8.4B tokens.

In addition, Section 6.5 compares our mid-trained Qwen3-8B with OProver-8B-Base (mid-trained on 65B tokens) under identical downstream SFT settings, further supporting the effectiveness of our mid-training recipe. Another relevant work is Leanstral, which reports 6.5B tokens of Lean data but provides limited details on data collection.

【W2】 Limited comparative evaluation of the agent harness.

We compare four harnesses using the same DeepSeek-V4.1-1012 checkpoint and sampling parameters, with a maximum of 200 model calls per problem, averaged over three runs. For Claude Code, Codex, and the DeepSeek harness, we use Docker containers to provide a consistent, isolated, and internet-free evaluation environment.

BenchmarkdaVinci-ProverClaude CodeCodexDeepSeek Harness
FATE-H93.085.086.093.0
FATE-X47.045.020.037.0
Putnam-20087.586.086.088.5

Conclusion: Our harness is competitive: it ties on FATE-H, scores highest on FATE-X, and trails DeepSeek on Putnam-200.

【W3】 Insufficient reporting of data synthesis costs.

The shared synthesis service used 48 H200 GPUs for approximately 20 days: 23,040 allocated GPU-hours. Lean verification used approximately 160 CPU cores over the same period: 76,800 allocated CPU core-hours. We have added these details to Appendix F, with the additions highlighted in blue.

【Q2】 Change a weak teacher model to show the effectiveness of the proposed method.

Our pipeline can also produce Lean-verified outputs with weak teachers. We conducted two targeted synthesis evaluations, each involving 1,000 examples.

(1) Decomposition synthesis. Our pipeline starts from existing verified proofs, prompts the teacher to rewrite them into lemma-based proofs, and verifies the resulting proofs with Lean.

TeacherParameterspass@1pass@2pass@3
Qwen3.8-27B27B50.5%65.7%73.7%
DeepSeek-V4-Flash-0731284B-A13B69.8%86.8%91.9%

(2) Agent-trajectory collection. We evaluate three teacher models using our agent rollout harness.

TeacherParameterspass@1pass@2Mean interaction turns
Qwen3.6-35B-A3B35B-A3B81.3%90.2%58.84
Qwen3.8-27B27B98.1%99.8%12.51
DeepSeek-V4.1-Flash-1012552B-A16B100.0%100.0%16.09
  1. These results demonstrate that verifier-guided rejection sampling enables the weak teacher to generate Lean-verified data. Although its success rate remains below that of the stronger teacher, this gap can be mitigated by scaling up the sampling budget.

  2. However, for the weak model like Qwen3.6-35B-A3B that exhibits repetitive and circuitous interactions, this may require additional rubric-based filtering beyond Lean-guided rejection sampling and increase the sampling cost per retained trajectory.

These experiments demonstrate the feasibility of collecting Lean-verified data with smaller teachers.


We hope to engage in further constructive discussion to help clarify our work and its contributions.

Reviewer_BHWcOpenReview ↗

Response to the rebuttal.

Thank the author's response to my concerns. Based on the rebuttal results, the improvements of the proposed agent and skill are minor compared with the base. I will keep my score unchanged.

We respectfully disagree that the improvements are minor, particularly given that your original review acknowledged our "broad empirical evaluation with consistent downstream gains."


1. Our midtraining data recipe combines skill and agent data.

As you acknowledged in Strength 1, our approach combines mathematical skill supervision with agent trajectories. We do not claim that each component individually yields substantial gains. Instead, our ablations show that the skill–agent mixture outperforms either component alone, including skill-only training at the same token budget. Furthermore, Section 6.2 shows that skill data also influence agent behavior.


2. The improvements are not minor.

For Qwen3-8B, our complete training recipe achieves an average improvement of 13.7 percentage points across three benchmarks over the SFT-only baseline (The SFT data are collected using the harness we developed).

Training configurationMT TokensMiniF2FFATE-MProofNet
Base + SFT (5 epochs)0B73.468.028.4
Base + SFT (7 epochs)0B80.972.732.6
Agent + SFT (5 epochs)7.1B84.279.337.1
Agent + Skill + SFT (5 epochs)33.7B85.385.340.4
Δ vs. 5-epoch SFT+11.9+17.3+12.0

Even compared with the stronger 7-epoch SFT baseline, the gains remain +4.4, +12.6, and +7.8 points, respectively.

Moreover, our recipe also improves the already strong Qwen3.5-9B backbone, despite its extensive prior agentic training. We include DeepSeek-Flash-0731 as a higher-performing reference model.

ModelSizeMiniF2FFATE-MProofNet
Qwen3.5-9B-Base + SFT9B87.488.442.9
Qwen3.5-9B-Midtrain + SFT9B90.193.344.5
Δ+2.7+4.9+1.6
DeepSeek-Flash-0731 (reference)286B-A13B92.798.153.4

We believe these results demonstrate meaningful and consistent improvements, rather than minor gains, and respectfully encourage you to reconsider this assessment.

Reviewer_6wgrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents daVinci-Prover, a training recipe for interactive theorem proving in Lean. It constructs a mid-training corpus combining mathematical skill priors, including decomposition, search, revision, and self-verification, with trajectories containing tool interactions and verifier feedback. A lightweight proving harness is used to collect successful agent trajectories for subsequent supervised fine-tuning. The paper evaluates the resulting models across several established theorem-proving benchmarks and studies data composition, training scale, agent behavior, and memory compaction. It reports consistent improvements over corresponding SFT-only baselines.

Critical strengths

The corpus explicitly represents intermediate proving activities rather than relying only on completed proofs. The distinction between reusable mathematical skills and executable interaction trajectories provides a useful way to organize training data for theorem-proving agents. Component ablations, data-scale experiments, and trajectory analyses offer evidence about the recipe’s behavior beyond the final benchmark scores.

Critical weaknesses

The broader performance comparisons do not establish the advantage of the harness itself.

Table 2 contrasts interactive pass@8 results with mostly externally reported pass@32 results under different protocols. The paper appropriately marks these results, but the comparisons vary both the underlying model and the inference procedure. They therefore provide limited support for attributing the observed advantage specifically to the daVinci-Prover harness, as done in Section 5.2. Similarly, the substantial gains from memory compaction are reported without total generation or verification costs, leaving their computational tradeoff unclear. These limitations affect the interpretation of the broader system claims, even though the internal mid-training comparisons remain useful.

Questions for authors

Were any comparisons performed with the total mid-training token budget held fixed while varying the proportion of skill data and agent trajectories? Such results would help distinguish complementarity between the data types from the benefit of additional training.

Were any same-model comparisons performed between the daVinci-Prover harness and an alternative proving setup under comparable inference budgets?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 6wgr

Thank you for acknowledging our useful data recipe and broad empirical evaluation.

【W1 & Q2 & W2.1】The broader performance comparisons do not establish the advantage of the harness itself. They therefore provide limited support for attributing the observed advantage specifically to the daVinci-Prover harness, as done in Section 5.2.

We additionally compare four harnesses using the same DeepSeek-V4.1-1012 checkpoint and sampling parameters, with a maximum of 200 model calls per problem, averaged over three runs. For Claude Code, Codex, and the DeepSeek harness, we use Docker containers to provide a consistent, isolated, and internet-free evaluation environment.

BenchmarkdaVinci-ProverClaude CodeCodexDeepSeek Harness
FATE-H93.085.086.093.0
FATE-X47.045.020.037.0
Putnam-20087.586.086.088.5

Conclusion: Our harness is competitive: it ties on FATE-H, scores highest on FATE-X, and trails DeepSeek on Putnam-200.

Revision: Our intended claim in Section 5.2 is that agentic proving can outperform direct whole-proof generation with repeated sampling. We have revised Section 5.2 to distinguish this from the contribution of the harness itself, with changes highlighted in blue.

【W2.2】 Computational cost of memory compaction

We now report generation and verification resource usage across 26,708 recorded turns. Token counts are in millions.

DatasetCompaction eventsInput tokensOutput tokens
MiniF2F104414.2922.44
FATE-M33140.677.89
FATE-H3651,646.5163.48
Total5022,201.4793.81

All 502 compactions use rule-based extraction without additional LLM calls for summarization. Recorded Lean verification durations total 11.02 hours, with a median of 0.734 seconds per call. We have added these statistics to Appendix H, with the additions highlighted in blue.

Conclusion: We include memory compaction as an additional analysis to illustrate the potential for further improving prover performance through inference-time techniques. Our primary contribution remains agent-native mid-training and its effectiveness in enhancing downstream agentic theorem proving.

【Q1】 Were any comparisons performed with the total mid-training token budget held fixed while varying the proportion of skill data and agent trajectories? Such results would help distinguish complementarity between the data types from the benefit of additional training.

We add a 33.7B-token skill-only baseline for Qwen3.5-9B to control for the total mid-training budget, keeping downstream SFT unchanged.

Qwen3.5-9B-baseTokensMiniF2FFate-MProofNet
Base--87.488.442.9
Skill-only25.3B87.889.843.6
Skill-only33.7B88.989.243.4
Skill-only (75%) + Agent (25%)33.7B90.193.344.5

Conclusion:

  1. Increasing skill-only mid-training from 25.3B to 33.7B tokens does not yield consistent gains across the three benchmarks.
  2. At the same 33.7B-token mid-training budget and with identical downstream SFT, the skill–agent mixture outperforms expanded skill-only training across all three benchmarks, supporting the value of agent trajectories beyond simply increasing training volume.

We hope to engage in further constructive discussion to help clarify our work and its contributions.

Dear Reviewer 6wgr,

As the discussion period draws to a close, we would like to kindly follow up on our response addressing:

  • Harness comparisons,
  • Computational costs of memory compaction (a supplementary analysis of an inference-time technique, rather than a core contribution),
  • The value of agent data under a fixed training budget.

We hope our additional experiments and clarifications have addressed your concerns, and we would greatly appreciate hearing whether any concerns remain.

Thank you for your time and consideration!

Thanks for the response and the new experiments. I’ll keep my score. My reasons:

  • Section 5.2: the claim was changed from “the harness helps” to “agentic proving beats whole-proof generation,” but it still rests on the same Table 2, where the two sides use different models, protocols and budgets. So the new claim has the same problem as the old one. Either describe the result without drawing this conclusion, or run one model in both settings with similar compute.
  • Harness comparison: the same-model experiment answers my question, and the harness looks on par with existing ones. Please add it to the paper; it is not in the revised version.
  • Compaction: Appendix H gives the costs with compaction, but not without it, so it is still hard to tell whether the gain is worth the extra compute.
  • Skill vs. agent data: the 33.7B skill-only run helps, but it is one data point with small gains and no variance, and it is also missing from the revised PDF. Other results in the rebuttal show that agent-only (8.4B) is close to the full mixture on MiniF2F and ProofNet, and that for Qwen3.5-9B, mid-training + SFT is about the same as longer SFT at equal tokens. So it is still unclear how much the skill data adds beyond more training.

The mid-training ablations are useful, but the broader claims are not well supported yet.

Thank you for your prompt responses. We provide further clarification below.


Agentic Proving vs. Whole-Proof Generation

  1. The ReAct paper, Synergizing Reasoning and Acting in Language Models, shows that combining reasoning with external actions can outperform reasoning alone on certain tasks. Given the time constraints, we present a preliminary experiment as follows:

    DeepSeek-V4.1-Flash-1012Total Input TokensTotal Output TokensFate-H (Pass @ 1)
    direct proof30,1165,771,7213.0
    agentic proof882,095,9054,101,23891.0

As shown, agentic proof substantially improves proof performance, despite requiring more input tokens, while generating fewer output tokens. Therefore, we polish our paper description as follows:

Agentic proving may benefit from multi-turn interaction and environment feedback, which enable the model to construct proofs incrementally, localize errors, and revise intermediate steps. Relative to direct whole-proof generation, this process provides additional opportunities for targeted correction, at the cost of additional model calls and Lean verification within each trajectory.

  1. 【Concern Solved】Thank you for your suggestion. We have incorporated it into our paper. We do not claim superior performance, but rather emphasize its lightweight design (as noted in your summary).

  1. 【Concern Solved】Thank you for this suggestion. We have updated a cost comparison in Appendix H, using a shared-prefix no-compaction baseline that stops before the first compaction. The additional reported cost comes from the continued generation and Lean verification enabled by compaction.

  1. We elaborate on the fourth point as follows.

4.1 Skill vs. agent data: the 33.7B skill-only run helps, but it is one data point with small gains and no variance, and it is also missing from the revised PDF.

Midtraining, followed by further SFT, is time-consuming. Given the limited remaining time, we cannot conduct more data mixture experiments. However, your original question was about the complementarity between data types through additional training, which we believe our initial response has addressed. We have revised the manuscript accordingly and included the variance.

4.2 Other results in the rebuttal show that agent-only (8.4B) is close to the full mixture on MiniF2F and ProofNet,

We presume you are referring to Reviewer BHWc. Indeed, this further demonstrates the effectiveness of our agent-only data. For reference, we also provide results on Qwen3-8B, which has less native agent training. The improvements are substantial, and incorporating skill data yields even better performance, suggesting that skill data also helps better align the training distribution.

Training configurationMT TokensMiniF2FFATE-MProofNet
Base + SFT (5 epochs)0B73.468.028.4
Agent + SFT (5 epochs)7.1B84.279.337.1
Δ vs. 5-epoch SFT+ 10.8+ 11.3+ 8.7
Agent + Skill + SFT (5 epochs)33.7B85.385.340.4
Δ vs. 5-epoch SFT+11.9+17.3+12.0
Δ vs. Agent + SFT (5 epochs)+ 1.1+ 6.0+ 3.3

4.3 and that for Qwen3.5-9B, mid-training + SFT is about the same as longer SFT at equal tokens.

We presume you are referring to Reviewer Vucw. As detailed in our experimental setup, the data includes both low-quality and high-quality SFT data, rather than merely controlling for equal token counts. We conclude that incorporating less-filtered agent trajectories during mid-training may help bridge the distribution gap between general pre-training and subsequent high-quality agentic SFT.

4.4 So it is still unclear how much the skill data adds beyond more training.

As shown in our previous response (4.2), skill data is particularly effective for models with limited Lean-related training, such as Qwen3-8B. Furthermore, Section 6.2 of our paper demonstrates that skill data also shapes agent behavior.

Reviewer_VucwOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

daVinci-Prover combines a 33.7B-token mid-training corpus of mathematical reasoning skills and interactive Lean trajectories with approximately 36K harness-generated SFT trajectories. The paper evaluates this recipe across base models and theorem-proving benchmarks, and studies data composition, training dynamics, tool use, and memory compaction.

Critical strengths

S1. The data construction targets concrete in-domain behaviors, including proof decomposition, tactic search, and error diagnosis and repair. This is a coherent, practically useful approach to preparing models for neural theorem proving.

S2. The matched downstream-SFT comparisons show consistent gains across backbones and benchmarks, including an average 13.7-point improvement on Qwen3-8B in Table 5.

S3. Appendix C describes benchmark decontamination, including removal of entire parent trajectories and replay copies.

Critical weaknesses

W1. Figure 5 varies SFT through five epochs, but does not match the total budget of mid-training plus SFT. A natural baseline is stronger SFT-only training, using additional epochs or in-domain trajectories under a comparable total training budget. Likewise, Table 4 compares 33.7B combined tokens with only 8.4B agent-only tokens. The OProver comparison, despite its larger 65B-token mid-training budget, provides useful supporting evidence but does not resolve how best to allocate a fixed budget between mid-training and SFT.

W2. Table 6 raises MiniF2F pass@1 from 90.1 to 97.1 through compaction, but does not include Base + SFT with compaction. It therefore does not establish the mid-training advantage when both models receive the same context-management support.

W3. Several improvements in Table 2 are below one percentage point, but paired uncertainty estimates are not reported, making their robustness difficult to assess.

W4. Figure 2 is missing a few strong open-source prover baselines, e.g. BFS-Prover-V2 and Leanstral

Questions for authors
  1. How does the method compare with stronger SFT-only and agent-only training under comparable total trained tokens?
  2. Does the advantage over Base + SFT persist when both models use compaction?
  3. Can you report paired uncertainty estimates for the smaller gains in Table 2?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Vucw

Thank you for recognizing the practical value of our data construction, the gains under matched downstream SFT, and the benchmark decontamination procedure.

【W1 & Q1】Stronger SFT-only and agent-only training under comparable total trained tokens?

We therefore compare seven epochs of high-quality SFT (HQ-SFT) with agent-only mid-training (MT), followed by 1.5 epochs of HQ-SFT, both using about 9.1B total training tokens. We also evaluate a mixed-SFT baseline that combines the less-filtered agent trajectories used for mid-training (denoted as LQ-SFT trajectories) with HQ-SFT data. All scores report mean pass@1 over three evaluation runs.

ModelTraining configurationTotal tokensMiniF2FFATE-MProofNet
Qwen3.5-9BHQ-SFT, 5 epochs6.5B87.488.442.9
Qwen3.5-9BHQ-SFT, 7 epochs9.1B88.191.343.1
Qwen3.5-9BAgent MT 7.1B + HQ-SFT 1.5 epochs9.1B88.891.143.3
Qwen3.5-9BLQ-SFT + HQ-SFT, 1 epoch9.7B85.688.239.8
Qwen3-8BHQ-SFT, 5 epochs6.5B73.468.028.4
Qwen3-8BHQ-SFT, 7 epochs9.1B80.972.732.6
Qwen3-8BAgent MT 7.1B + HQ-SFT 1.5 epochs9.1B81.775.633.4
Qwen3-8BAgent MT 7.1B + HQ-SFT 5 epochs13.6B84.279.337.1

Conclusion:

  1. Additional HQ-SFT (Stronger SFT) improves performance over the five-epoch baseline.
  2. At approximately 9.1B tokens, agent-only mid-training followed by lightweight HQ-SFT achieves higher scores on Qwen3-8B and comparable performance on Qwen3.5-9B relative to SFT-only.
  3. Mixed SFT underperforms HQ-SFT-only despite using more tokens, suggesting that data quality and composition matter beyond the total token budget during the SFT phase.
  4. Further scaling downstream HQ-SFT unlocks additional performance gains, as demonstrated by the mid-trained Qwen3-8B model.

These results show that incorporating less-filtered agent trajectories during mid-training may help bridge the distribution gap between general pre-training and subsequent high-quality agentic SFT.

【W2 & Q2】 Base + SFT with compaction

Model, with compactionMiniF2FFATE-MFATE-H
Qwen3.5-35B-A3B Base + SFT95.596.538.0
Qwen3.5-35B-A3B Midtrain + SFT97.197.342.0

Conclusion: Both models use the same compaction policy, sampling parameters, and inference limits. Scores are the mean pass@1 over three runs. The gains persist for Qwen3.5-35B-A3B when both models use compaction.

【W3 & Q3】robustness of results.

For Qwen3.5-9B, paired pass@8 results are:

Pass@8MiniF2FProofNetProverBench
Qwen3.5-9B-Base + SFT92.651.166.2
Qwen3.5-9B-Midtrain + SFT93.952.766.5
Paired bootstrap 95% CI for Δ (pp)[0.00, 2.05][-0.54, 4.30][-1.23, 1.85]
McNemar p0.500.37501.0

For Qwen3.5-9B, paired avg@8 results are:

Avg@8MiniF2FProofNetProverBench
Qwen3.5-9B-Base + SFT87.941.8060.38
Qwen3.5-9B-Midtrain + SFT89.145.3661.46
Paired bootstrap 95% CI for Δ (pp)[-0.15, 2.61][2.02, 5.24][0.27, 1.96]
Exact paired permutation test0.115.22×10⁻⁶0.0145

Δ = Midtrain + SFT − Base + SFT; pp denotes percentage points.

Conclusion: These results support the usefulness of mid-training for improving average per-attempt success, with significant Avg@8 gains on ProofNet and ProverBench under unadjusted paired permutation tests. Pass@8 point estimates also improve across all three benchmarks, although these gains and the MiniF2F Avg@8 gain are not statistically significant.

【W4】 BFS-Prover-V2 and Leanstral-1.5

We evaluate Leanstral-1.5 under our default evaluation protocol (30 interaction turns and a 200K-token context window). Published BFS-Prover-V2 results are included for reference.

ModelParamsMiniF2FFate-MProofNet
DeepSeek-V4-Flash-0731284B-A13B92.798.153.4
BFS-Prover-V2†32B dense86.1--41.4
Leanstral-1.5119B-A6B82.986.038.7
Ours9B dense90.193.344.5

† Published accumulative results (pass@8192 per configuration); evaluation protocols and budgets differ from ours.


We hope to engage in further constructive discussion to help clarify our work and its contributions.

The response addressed several of my concerns, so I raised my score to 3.

We sincerely appreciate your support and encouraging feedback!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Summary

Dear ACs/SACs/PCs

We sincerely thank you for your time and effort in managing the review process for our paper. We provide a summary of the current status of our paper and the progress of our rebuttal.

1. Summary of Contributions

  1. We construct a 33.7B-token mid-training corpus combining mathematical skill priors with environment-native agent trajectories to prepare models for interactive theorem proving.
  2. We develop daVinci-Prover, a lightweight prover harness that achieves performance comparable to more complex harnesses such as Codex, Claude Code, and DeepSeek harness.
  3. We curate approximately 36K high-quality, domain-diverse agent trajectories for supervised fine-tuning to elicit the theorem-proving capabilities acquired during mid-training.
  4. We show that incorporating less-filtered agent trajectories during mid-training can help bridge the distribution gap between general pre-training and subsequent high-quality agentic SFT.

2. Review Summary

We summarize the key points of the review below:

ReviewerStrengthsWeaknesses / Questions
VucwPractical and well-motivated approach; Strong empirical results; Rigorous benchmark decontaminationStrong SFT vs. Agent Midtraining + SFT; Base-model memory compaction; Result uncertainty; Additional baselines
6wgrUseful approach; Comprehensive experimens; consistent improvementsHarness comparison; Memory compaction cost; Skill vs. agent trajectory ratios
BHWcWell-motivated training recipe; Broad experimens; consistent downstream gainsHarness comparison; Teacher-model confounding; Data synthesis cost; Weak-model synthesis

3. Revision Summary

All three reviewers acknowledged our well-motivated skill–agent training recipe and extensive empirical evaluation demonstrating consistent performance gains. To address their concerns, we have made the following revisions:


(1) Added stronger SFT-only and agent-only baselines with matched total training tokens. (reviewer Vucw, Appendix K. )

(2) Added a Base + SFT baseline with compaction. (reviewer Vucw, Appendix N)

(3) Added paired uncertainty estimates for pass@8 and avg@8. (reviewer Vucw, Appendix M)

(4) Provided results for BFS-Prover-V2 and Leanstral. (reviewer Vucw)

(5) Added a harness comparison. (reviewer 6wgr, reviewer BHWc, Appendix J)

(6) Added comparisons of different skill-to-agent data ratios under a fixed budget (e.g., 100:0 and 75:25). (reviewer 6wgr, Appendix I)

(7) Provided a comparable data synthesis strategy using the same teacher model. (reviewer BHWc)

(8) Added estimates of data synthesis costs. (reviewer BHWc, Appendix F)

(9) Added data synthesis results with a weaker teacher model. (reviewer BHWc, Appendix L)

(10) Added the comparison of memory compaction costs (reviewer 6wgr, Appendix H)


The remaining concern is that Qwen3.5-9B-midtrain shows minor improvement over Qwen3.5-9B-Base after fine-tuning on our curated high-quality SFT data. We provide the following explanation as follows:

For Qwen3-8B, our complete training recipe achieves an average improvement of 13.7 percentage points across three benchmarks over the SFT-only baseline (The SFT data are collected using the harness we developed).

Training configurationMT TokensMiniF2FFATE-MProofNet
Base + SFT (5 epochs)0B73.468.028.4
Base + SFT (7 epochs)0B80.972.732.6
Agent + SFT (5 epochs)7.1B84.279.337.1
Agent + Skill + SFT (5 epochs)33.7B85.385.340.4
Δ vs. 5-epoch SFT+11.9+17.3+12.0

Even compared with the stronger 7-epoch SFT baseline, the gains remain +4.4, +12.6, and +7.8 points, respectively.

Moreover, our recipe also improves the already strong Qwen3.5-9B backbone, despite its extensive prior agentic training[1]. We include DeepSeek-Flash-0731 as a higher-performing reference model.

ModelSizeMiniF2FFATE-MProofNet
Qwen3.5-9B-Base + SFT9B87.488.442.9
Qwen3.5-9B-Midtrain + SFT9B90.193.344.5
Δ+2.7+4.9+1.6
DeepSeek-Flash-0731 (reference)286B-A13B92.798.153.4

We believe these results demonstrate meaningful and consistent improvements, rather than minor gains.

Reference:

[1] Qwen3.5: Towards Native Multimodal Agents, https://qwen.ai/blog?id=qwen3.5

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.