acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Supervision Granularity: Segment-Level Learning for LLM-Based Theorem Proving

Abstract

LLM-based automated theorem proving in Lean 4 commonly uses step-level tactic prediction with tree search or whole-proof generation. These two paradigms represent opposite granularities for constructing supervised training data: the former provides dense local signals but may fragment coherent proof processes, while the latter preserves global structure but requires complex end-to-end generation. In this paper, we revisit supervision granularity as a training-set construction problem and propose segment-level supervision, which extracts locally coherent proof segments to train policy models. We further reuse the same strategy at inference time to trigger short rollouts for existing step-level models. When trained with segment-level supervision on STP, LeanWorkbook, and NuminaMath-LEAN, the resulting policy models achieve proof success rates of 64.84%, 60.90%, and 66.31% on miniF2F and 32.13%, 21.73%, and 17.20% on FATE-M, respectively, consistently outperforming both step-level and whole-proof baselines. Goal-aware rollout further improves existing step-level provers while reducing inference costs: on miniF2F, it increases the proof success rate of BFS-Prover-V2-7B from 68.77% to 70.74% and that of InternLM2.5-StepProver from 59.59% to 60.33%; on FATE-M, the corresponding success rates increase from 45.60% to 50.67% and from 33.47% to 36.13%, showing that appropriate supervision granularity better aligns model learning with proof structure and search. Code and models are available at https://anonymous.4open.science/r/SEG-ATP.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
large language models, automated theorem proving

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_um2DOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes segment-level supervision, which splits verified Lean 4 proof trajectories whenever the number of open goals changes and uses the resulting multi-tactic segments as supervision targets. They report that a Qwen2.5-Math-7B policy trained this way on NuminaMath-LEAN reaches 66.31% on miniF2F, compared with 63.11% for step-level and 55.90% for whole-proof supervision on the same data. The paper also proposes goal-aware rollout, which reuses the same signal at inference time and lets an existing step-level prover continue along the current branch for a few steps whenever a tactic changes the goal count. Without retraining, this raises BFS-Prover-V2-7B from 68.77% to 70.74% on miniF2F and from 45.60% to 50.67% on FATE-M.

Critical strengths
  • The comparison of supervision granularities is clean, with step-level, whole-proof and segment-level targets are built from the same verified trajectories and trained on the same base model. Segment-level training acheives the best in all setting (66.31% vs. 63.11% and 55.90% on miniF2F with NuminaMath-LEAN).
  • Goal-aware rollout, by only changing the search loop, an easy-adopt method, improves the step provers on both benchmarks while using fewer tokens and less time (on FATE-M, BFS-Prover-V2-7B goes from 45.60% to 50.67% with about 25% fewer tokens),
Critical weaknesses
  1. The gains are consistent in direction but small, and the evaluation cannot show that they are significant.
    • Most gains in Tables 1 and 3 are under 2%. Only NuminaMath-LEAN on miniF2F (+3.2%) and the two rollout results on FATE-M (+5.1% and +2.7%) are larger, and with Qwen3-14B (Appendix A.3) the gain over step-level falls below the standard deviation (53.36% vs. 52.46% on miniF2F, 17.87% vs. 17.20% on FATE-M).
    • The reported standard deviation comes from five beam-search runs of one checkpoint, so it does not cover training randomness. With differences of 1–2 points and a single training seed, the ranking between granularities is not reliable, and the +0.74 gain for InternLM2.5-StepProver on miniF2F is within the run-to-run noise.
    • The step-level baselines in Table 3 are also weaker than in their own papers. BFS-Prover-V2-7B gets 68.77% on miniF2F, below the 70.8% its paper reports for the older V1-7B, and InternLM2.5-StepProver gets 59.59%, which matches its no-critic result (59.4%) rather than the 65.9% with the critic. Both are run with beam search instead of their original decoding and on Lean 4.11 instead of 4.10.0 and 4.7.0, so small gains on these baselines may not hold under the original settings.
  2. The paper credits the gains to the goal-count boundaries, but none of the experiments separates this effect from simply taking several steps per expansion.
    • Table 3 compares goal-aware rollout only with plain best-first search. A rollout triggered at random positions with the same frequency could conceivably give a similar gain, and this would be easy to test.
    • I would also like to see the contributions of training and inference separated. The authors' own step-level policy is never combined with goal-aware rollout, so the extra value of segment-level training on top of rollout is unknown.
Questions for authors
  1. In Table 3, BFS-Prover-V2-7B and InternLM2.5-StepProver both score below the performance in their own papers. Have you tried their original inference settings (sampling at temperature 1.3 for BFS-Prover-V2, and the critic for InternLM2.5-StepProver)? I would like to see the results of goal-aware rollout under these settings.
  2. Did you try goal-aware rollout on your own step-level model? A rollout triggered at random positions with the same frequency would also be a useful control for the goal-change trigger.
  3. If I understand correctly, DeepSeek-Prover-V1.5 already uses a truncate-and-resume search, which generates several tactics at once, cuts the output at the first error and continues the search from the last valid state. Whole-proof-seg in Table 1 follows the same procedure, but the paper cites V1.5 only as a whole-proof method. Apart from the stopping rule, I do not see a clear difference between the two. Could you clarify this?
  4. I notice the paper only reveals a per-theorem timeout of 1800 seconds. Is there also a per-tactic timeout or heartbeat limit, and how much CPU and memory does each Lean process get? Are the final proofs also recompiled as complete files, with sorry, admit, native_decide and new axioms rejected? In my experience older Lean and LeanDojo setups let a proof pass while still relying on such shortcuts, and this matters a lot once the pipeline is used for RL. Since time cost is one of the main metrics, the reported time also depends directly on this environment.

Suggestions for future work

  1. It would be interesting to see segment-level supervision on top of a strong prover or how well it bridges the performance via RL.
  2. If the goal count is later used as a step boundary or reward in RL, the policy might add constructor or have steps just to change the number of goal count it self to hack, which is worth a short note in the limitations.
10 comments
AuthorsOpenReview ↗

Response to W1-1 && W1-2 (Part 1)

W1-1 && W1-2 (Part 1)

We thank the reviewer for their attention to the statistical reliability of the experiments. First, we clarify that the five independent runs reported in the original paper were all inference runs using the same fixed model checkpoint. Therefore, the reported mean ± std reflects randomness during inference rather than randomness during training. To further examine the effect of training randomness, we retrained the models using a different PyTorch random seed (seed 123, whereas the original experiments used seed 42) and conducted three independent inference evaluations on each of miniF2F and FATE-M. As shown in Table F, across all six combinations of the three training datasets and two evaluation benchmarks, Segment-level achieves the highest mean proof success rate, consistent with the original experimental results.

Table F: Results with an Additional Training Seed (PyTorch Seed 123; Mean ± Std over Three Independent Runs)

Training SetTest SetStep-levelWhole-proofWhole-proof-segSegment-level (Ours)
LeanWorkbookminiF2F58.74 ± 0.6354.10 ± 0.4155.19 ± 1.2560.66 ± 0.41
LeanWorkbookFATE-M20.00 ± 0.6719.11 ± 0.3817.11 ± 0.3822.67 ± 0.67
NuminaMath-LEANminiF2F64.75 ± 0.7155.87 ± 0.2453.55 ± 0.2466.94 ± 0.24
NuminaMath-LEANFATE-M16.67 ± 0.6716.44 ± 1.6816.00 ± 0.6716.89 ± 1.02
STPminiF2F62.70 ± 1.4861.61 ± 0.8563.66 ± 0.2464.75 ± 0.71
STPFATE-M30.00 ± 1.3329.56 ± 2.6930.22 ± 0.7732.44 ± 1.02

In addition, we added statistical significance tests. For the training-side comparisons, we combine the five inference results from the original experiments with the three additional inference results under seed 123, for a total of eight evaluations. For the inference-side GAR comparisons, we use the five inference results from the original experiments. miniF2F and FATE-M contain 244 and 150 theorems, respectively. Because the different methods are evaluated on the same test set, we use a two-sided theorem-stratified exact conditional randomization test, preserving theorem-level pairing and testing differences in mean Pass@1 between methods [1]. We also apply Holm correction separately to the multiple training-side and inference-side comparisons [2]. The results are shown in Tables G and H.

Table G: Statistical Significance of Segment-level Supervision (Training-side Comparisons; Holm-adjusted p-values)

Training SetTest SetSegment-level vs. Step-level Δ Pass@1 (pp)Holm-adjusted pSignificantSegment-level vs. Whole-proof Δ Pass@1 (pp)Holm-adjusted pSignificantSegment-level vs. Whole-proof-seg Δ Pass@1 (pp)Holm-adjusted pSignificant
LeanWorkbookminiF2F+1.903.524×10−43.524\times10^{-4}Yes+6.354.469×10−364.469\times10^{-36}Yes+6.059.644×10−289.644\times10^{-28}Yes
LeanWorkbookFATE-M+1.580.04318Yes+2.750.006240Yes+5.421.958×10−121.958\times10^{-12}Yes
NuminaMath-LEANminiF2F+2.822.514×10−62.514\times10^{-6}Yes+10.664.857×10−494.857\times10^{-49}Yes+13.521.533×10−751.533\times10^{-75}Yes
NuminaMath-LEANFATE-M+1.000.615641No+0.830.615641No+1.420.316880No
STPminiF2F+1.840.0005690Yes+3.183.044×10−113.044\times10^{-11}Yes+1.230.01443Yes
STPFATE-M+2.080.035677Yes+1.750.052150No+1.830.045823Yes

Table H: Statistical Significance of Goal-aware Rollout (Inference-side Comparisons; Holm-adjusted p-values)

ModelTest SetGAR Avg. Pass@1 (%)Baseline Avg. Pass@1 (%)Δ Pass@1 (pp)Raw pHolm-adjusted pSignificant
InternLM2.5-StepProverminiF2F60.3359.59+0.740.1936740.193674No
BFS-Prover-V2-7BminiF2F70.7468.77+1.970.0015240.003048Yes
InternLM2.5-StepProverFATE-M36.1333.47+2.670.0433620.043362Yes
BFS-Prover-V2-7BFATE-M50.6745.60+5.070.0001510.000302Yes
AuthorsOpenReview ↗

Response to W1-1 && W1-2 (Part 2)

W1-1 && W1-2 (Part 2)

The tests show that all training-side comparisons on miniF2F are statistically significant, whereas some comparisons on FATE-M do not reach significance. Among the four inference-side GAR comparisons, three are significant after Holm correction. For the nonsignificant result of InternLM2.5-StepProver on miniF2F, the GAR trigger rate is only 3.65%; this relatively low trigger frequency may be one reason for the limited improvement. Although some comparisons are not statistically significant, Segment-level achieves the highest mean accuracy across all six training-set and test-set combinations under the additional training seed, with overall improvements in accuracy.

We also agree that on formal theorem-proving benchmarks of limited size, such as miniF2F and FATE-M, an accuracy improvement of approximately one percentage point corresponds to only a small number of additional successfully proved theorems. However, each additional success represents the discovery of a valid path through a vast proof search space, culminating in a complete proof script that passes Lean verification. These improvements therefore remain practically meaningful. More importantly, Segment-level consistently improves performance across different training-data sources and both evaluation benchmarks. This overall pattern across data sources and evaluation benchmarks further supports the effectiveness of segment-level supervision.

References [1] S. H. Jung. Stratified Fisher's exact test and its sample size calculation. Biometrical Journal, 56(1):129–140, 2014. [2] S. Holm. A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics, 6(2):65–70, 1979.

AuthorsOpenReview ↗

Response to W1-3 && Q1

W1-3 && Q1

We thank the reviewer for the suggestion. To further validate the effectiveness of goal-aware rollout (GAR) under the original inference configurations, we add controlled experiments using the official inference settings of the two models. Specifically, for BFS-Prover-V2-7B, we use the official prompt template, sampling with temperature = 1.3, and path-length-normalized scoring (α=2.0\alpha = 2.0). For InternLM2.5-StepProver, we use the official prompt template, temperature = 0.7, and the original 1.8B critic model for state scoring. We retain the original numbers of generated candidates for the two models (3 and 32, respectively) and set the GAR rollout horizon to H=5H = 5 for both.

For a fair comparison, Base and GAR use the same model configurations and search budgets, with at most 600 regular expansions and a search timeout of 600 seconds per theorem. The additional generation and execution costs of GAR are also included in the shared budget. We conduct three independent evaluations on miniF2F (244 theorems) and FATE-M (150 theorems), respectively, with the following results:

Table M: Results of Goal-aware Rollout under the Original Inference Configurations (Mean ± Std over Three Independent Runs)

ModelDatasetBase Pass@1+ GAR Pass@1Gain
BFS-Prover-V2-7BminiF2F61.48 ± 1.0862.43 ± 1.03+0.96
BFS-Prover-V2-7BFATE-M22.22 ± 1.9226.22 ± 0.77+4.00
InternLM2.5-StepProverminiF2F51.64 ± 1.4853.83 ± 0.24+2.19
InternLM2.5-StepProverFATE-M30.00 ± 2.4031.33 ± 3.71+1.33

The results show that, with the original sampling strategies and scoring mechanisms, GAR still improves performance in all four settings involving the two models and two datasets. Specifically, BFS-Prover-V2-7B improves by 0.96 and 4.00 percentage points on miniF2F and FATE-M, respectively, while InternLM2.5-StepProver improves by 2.19 and 1.33 percentage points, respectively. These results further demonstrate that the effectiveness of GAR is not limited to the unified inference configuration in Table 3; it also improves proving performance under different sampling strategies and critic-guided search.

We thank the reviewer again for the suggestion. We will include these experimental results in the revised manuscript and further clarify the evaluation settings under different inference configurations to provide a more comprehensive account of the applicability of GAR to existing step-level provers.

Reviewer_Lz1UOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes segment-level supervision for LLM-based theorem proving in Lean, introducing an intermediate granularity between step-level and whole-proof generation. Specifically, the authors divide verified proof trajectories into segments based on changes in the number of open goals and train the model to generate multiple consecutive tactics within each segment. Additionally, they propose Goal-aware rollout, an inference-time strategy applicable to existing step-level provers, which performs several consecutive tactic generations when the number of open goals changes. Experimental results demonstrate improvements in theorem-proving performance and inference efficiency over existing baselines.

Critical strengths
  1. Simple and practical methodology: The proposed approach improves performance by restructuring existing proof trajectories without requiring architectural modifications or complex training algorithms. In particular, segment-level supervision can directly leverage existing verified proof data.

  2. Empirical validation across multiple settings: The authors evaluate segment-level supervision on three different training datasets and conduct ablation studies comparing alternative segmentation strategies. They also demonstrate that Goal-aware rollout can be applied to existing step-level provers, further supporting the practical applicability of their approach.

Critical weaknesses

1.Insufficient motivation and justification for the segmentation criterion: The paper defines segment boundaries based on changes in the number of open goals, but does not sufficiently explain why such changes correspond to meaningful reasoning units or appropriate supervision boundaries. Importantly, substantial changes in proof states can occur even when the number of open goals remains unchanged. For example, though the number of remaining subgoals are unchanged, the subgoal itself might change. Conversely, a change in the number of open goals does not necessarily indicate the completion of a meaningful reasoning step. Therefore, it remains unclear whether the proposed criterion adequately captures the underlying structure of mathematical proofs.

2.Insufficient justification for Goal-aware rollout: Goal-aware rollout performs consecutive tactic generations after a change in the number of open goals. However, the paper does not clearly explain why such changes should serve as appropriate triggers for additional rollouts. Furthermore, it would be valuable to investigate whether this goal-aware strategy offers meaningful advantages over a simpler baseline that performs fixed-length consecutive rollouts regardless of goal changes.

3.Limited analysis of the sources of performance improvement: It remains unclear whether the improvements from segment-level supervision arise from learning meaningful local proof structures or simply from increasing the generation granularity and improving search efficiency. Further analysis is needed to clarify what specific reasoning capabilities are enhanced by the proposed supervision strategy.

Questions for authors

My main questions are closely related to the weaknesses mentioned above.

6 comments
AuthorsOpenReview ↗

Response to W1

W1

We thank the reviewer for the comments on the rationale for open-goal boundary selection. Existing Step-level and Whole-proof supervision organize proof supervision at the local and global levels, respectively: the former uses a single tactic as the learning unit and provides fine-grained local supervision, whereas the latter uses a complete proof as the learning target and preserves the overall proof structure. However, a proof process typically contains multiple locally coherent tactics that jointly expand, transform, or resolve a subgoal. We therefore explore supervision units at an intermediate granularity between these two extremes to account for both the coherence of local proof processes and the overall proof structure.

Lean proofs have an explicit goal-driven structure, and their progression typically involves the expansion and closure of subgoals. We therefore use changes in the open-goal count as an approximate structural proxy signal to identify structural transitions in the proof process and determine segment boundaries. This signal is simple, stable, and readily available directly from Lean proof states, without additional semantic modeling. We do not claim that changes in the open-goal count fully characterize all semantic changes in a proof process, nor do we regard them as the only or optimal boundary-selection criterion.

In addition, we compare alternative boundary-selection strategies, including token-based, tactic-distance-based, and state-distance-based segmentation, through ablation experiments. Specifically, when training on NuminaMath-LEAN and evaluating on miniF2F (Appendix A.2.1, Table 4), open-goal change achieves a mean proof success rate of 66.31%, whereas the best results of the three alternative strategies across their respective tested configurations are 59.92%, 61.23%, and 59.51%, respectively. Compared with the strongest alternative, tactic-distance-based segmentation (threshold 0.4), open-goal change improves the proof success rate by 5.08 percentage points. Furthermore, we conduct an additional boundary-selection ablation on FATE-M using NuminaMath-LEAN for training, with results reported in Table C. Open-goal change achieves a mean proof success rate of 17.20%, whereas the best results of the three alternative strategies across their respective tested configurations are 14.44%, 16.22%, and 15.56%, respectively. Compared with the strongest alternative, tactic-distance-based segmentation (threshold 0.4), open-goal change improves the proof success rate by 0.98 percentage points. These results show that open-goal change outperforms all tested alternative boundary-selection strategies on both evaluation benchmarks, further supporting its effectiveness as a boundary-selection signal for intermediate-granularity supervision. In future work, we will further explore other signals that reflect structural changes in proof states to identify more suitable boundary-selection criteria and construct intermediate-granularity supervision units that better align with local proof progress.

Table C: Boundary-selection Ablation on FATE-M (Mean ± Std)

Boundary SelectionSettingFATE-M Acc. (%) ↑
Token-based32 tokens14.44 ± 1.92
Token-based64 tokens11.78 ± 0.38
Token-based96 tokens11.56 ± 2.78
Tactic-distance-basedThreshold 0.416.22 ± 2.04
Tactic-distance-basedThreshold 0.615.11 ± 0.77
Tactic-distance-basedThreshold 0.814.22 ± 1.68
State-distance-basedThreshold 0.415.56 ± 2.52
State-distance-basedThreshold 0.612.22 ± 0.38
State-distance-basedThreshold 0.812.67 ± 0.00
Open-goal (Ours)—17.20 ± 1.10
AuthorsOpenReview ↗

Response to W2

W2

We thank the reviewer for the suggestions regarding the motivation for triggering rollouts based on changes in the open-goal count and the related ablation experiments. Lean proofs have an explicit goal-driven structure, and their progression typically involves the expansion and closure of subgoals. Changes in the open-goal count can reflect transitions in the structure of proof goals to some extent. We therefore trigger short rollouts after such changes so that the model can continue along the current local proof direction: when a goal expands into multiple subgoals, the model can continue attempting to resolve the newly generated subgoals; when some subgoals are closed, the model can proceed with the remaining goals under the reduced set of proof obligations. This approach avoids the frequent branch expansions caused by resuming explicit tree search prematurely, thereby improving search efficiency.

To further validate the role of the open-goal trigger mechanism, we add a Random rollout baseline. Rather than unconditionally triggering a fixed-length rollout after every successfully executed tactic, we adopt a random trigger strategy matched to the average trigger frequency of goal-aware rollout to avoid additional computational overhead from excessively frequent rollouts. Specifically, during inference, after each successful tactic execution, Random rollout randomly determines whether to trigger a subsequent rollout with a probability equal to the average trigger rate of goal-aware rollout under the corresponding setting. This setting retains consecutive tactic generation but no longer uses the open-goal signal to select trigger positions, thereby examining the role of the structural signal in determining when to trigger rollouts. The trigger statistics and experimental results are reported in Tables D and E, respectively.

Table D: Trigger Statistics of Goal-aware Rollout

ModelBenchmarkTrigger Rate (%)Avg. Triggers per Theorem
InternLM2.5-StepProverminiF2F3.6520.96
InternLM2.5-StepProverFATE-M8.8254.08
BFS-Prover-V2-7BminiF2F14.56101.29
BFS-Prover-V2-7BFATE-M29.76245.60

Table E: Comparison of Different Rollout Trigger Strategies (Mean ± Std over Five Independent Runs)

ModelBenchmarkBest-first (%)Random rollout (%)Goal-aware rollout (%)
InternLM2.5-StepProverminiF2F59.59 ± 0.3759.67 ± 0.4760.33 ± 0.45
InternLM2.5-StepProverFATE-M33.47 ± 2.0833.87 ± 1.7336.13 ± 0.73
BFS-Prover-V2-7BminiF2F68.77 ± 0.7969.10 ± 1.0370.74 ± 1.18
BFS-Prover-V2-7BFATE-M45.60 ± 1.3046.80 ± 1.3750.67 ± 0.82

The results show that Random Rollout improves over standard Best-first Search by 0.08%, 0.40%, 0.33%, and 1.20% across the four settings, indicating that consecutive tactic generation itself provides some benefits. However, Goal-aware Rollout achieves further improvements of 0.66%, 2.26%, 1.64%, and 3.87% over Random Rollout, obtaining the highest proof success rate in all four settings. These results suggest that, compared with randomly selecting trigger positions, using changes in the open-goal count to determine rollout trigger positions more effectively guides local search and further improves proof success rates.

AuthorsOpenReview ↗

Response to W3

W3

We thank the reviewer for the comments on the sources of the performance improvements from Segment-level and the reasoning capabilities it enhances.

First, to distinguish whether the improvements arise from proof-structure information encoded by boundary selection or merely from changes in generation granularity, we conduct further ablation experiments. The original ablation in Appendix A.2.1 (Table 4), together with the additional results reported in W1 (Table C), compares alternative segmentation strategies, including token-based, tactic-distance-based, and state-distance-based segmentation. The results show that different boundary-selection strategies lead to substantial performance differences. In addition, we add two length-distribution-matched random segmentation baselines, Random-Step and Random-Token, which strictly match the full tactic-step and token length distributions of the original open-goal segmentation, respectively. As shown in Tables A and B, open-goal segmentation outperforms both random baselines on miniF2F and FATE-M. This indicates that the improvements from Segment-level do not arise solely from increasing generation length or combining multiple tactics; the proof-structure information used for boundary selection also plays an important role.

In addition, we add two length-distribution-matched random segmentation baselines, Random-Step and Random-Token, which strictly match the full tactic-step and token length distributions of the original open-goal segmentation, respectively. As shown in Tables A and B, open-goal segmentation outperforms both random baselines on miniF2F and FATE-M. This indicates that the improvements from Segment-level do not arise solely from increasing generation length or combining multiple tactics; the proof-structure information used for boundary selection also plays an important role.

Table A: Segment Length Distribution of Open-goal Supervision

LengthMeanMedianStdP90
Steps4.241534.20559
Tokens255.1111135322.2803664

Table B: Proof Success Rates of Open-goal and Random-boundary Baselines (Mean ± Std over Five Independent Runs)

MethodminiF2F Acc. (%) ↑FATE-M Acc. (%) ↑
Open-goal (Ours)66.31 ± 0.7917.20 ± 1.10
Random-Step64.51 ± 0.5513.47 ± 0.99
Random-Token65.08 ± 0.9814.27 ± 1.61

Second, Segment-level improves proving capabilities at two levels: model generation and the search process. At the model-generation level, unlike Step-level, which supervises only the prediction of a single tactic, Segment-level organizes multiple locally coherent tactics into one supervision unit, allowing the model to learn dependencies between consecutive proof steps and thereby generate and advance a local proof process more coherently. At the search level, this multi-step generation allows search to expand an entire proof segment as a unit without resuming explicit branch expansion at every intermediate proof state, reducing frequent search expansions and repeated model calls and improving search efficiency. The training loss analysis in Figure 2 further shows that although Segment-level uses longer prediction targets than Step-level, its optimization process maintains good convergence behavior. This indicates that Segment-level can learn locally coherent multi-step proof processes while retaining relatively favorable optimization properties.

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

This paper studies supervision granularity for LLM-based theorem proving in Lean 4, formulating step-level and whole-proof supervision as two extremes of a boundary-selection problem over verified proof-state trajectories. The authors propose segment-level supervision: boundaries are placed at states reached after a tactic changes the open-goal count, and the tactics between adjacent boundaries form a multi-tactic SFT target, used as a macro-action in best-first search. The same signal is reused as goal-aware rollout (GAR) for existing step-level provers: when a candidate tactic changes the open-goal count, the prover performs a linear rollout of up to HH steps, and only the final valid state enters the frontier. Fine-tuning Qwen2.5-Math-7B on STP, LeanWorkbook, and NuminaMath-LEAN, segment-level models outperform step-level, whole-proof, and whole-proof-with-segment-search baselines on miniF2F and FATE-M; GAR improves BFS-Prover-V2-7B (68.77 → 70.74 on miniF2F, 45.60 → 50.67 on FATE-M) and InternLM2.5-StepProver while reducing token cost.

Critical strengths
  1. Clean and practical formulation. The boundary-selection view unifies step-level and whole-proof supervision and yields a method that needs no new architecture or proof data, only re-segmentation of existing trajectories.
  2. Consistent direction of gains. Segment-level is best in all 6 training-source × benchmark settings (Table 1), and GAR improves both off-the-shelf step-level provers without retraining (Table 3).
  3. Useful controls and cost analyses. The Whole-proof-seg control, the boundary ablation (Table 4), the horizon study (Table 5), and the time-/token-budget analyses (A.5) go beyond reporting final accuracy.
Critical weaknesses
  1. The causal role of the open-goal signal is not isolated.
  • Segmentation: The alternatives in Table 4 (token/edit-distance thresholds) are not matched to the segment-length distribution of the open-goal rule, and that distribution is never reported. It therefore remains unclear whether the gain comes from where boundaries are placed or simply from predicting a few tactics per action plus valid-prefix retention. A length-matched random-boundary baseline would settle this. The ablation also covers only one setting (NuminaMath-LEAN → miniF2F).
  • GAR: GAR is compared only with plain best-first search and with different HH. There is no trigger ablation, such as rolling out after every successful tactic or at random with the same trigger frequency. Single-sample lookahead naturally saves tokens and increases depth, so without such a baseline the gain cannot be attributed to the open-goal trigger.
  • The Whole-proof-seg control is also weak: a model trained only on s0→s_0 \to full proof is out of distribution when queried from intermediate states.
  1. Small margins with limited statistical support.
  • The "five independent runs" are inference runs from a single checkpoint with beam-search decoding (A.1), so the std reflects only inference nondeterminism, not training variance. No significance tests are reported.
  • Several margins are within about one std: NuminaMath → FATE-M 17.20±1.10 vs. 16.13±0.56; STP → FATE-M 32.13±0.56 vs. 31.07±0.89; InternLM2.5 + GAR on miniF2F 60.33±0.45 vs. 59.59±0.37.
  • Under matched token budgets (Table 7), Step-level is equal or better at small budgets in several settings (e.g., LeanWorkbook → FATE-M up to 5k tokens), so the efficiency claim should be qualified.
  • With Qwen3-14B (Table 6), all granularities drop 7–13 points below Qwen2.5-Math-7B (segment-level 53.36% vs. 66.31% on miniF2F). This is not discussed, which weakens the generality claim.
  1. Scope relative to current Lean provers.
  • The conclusions apply to CoT-free, step-level tree-search provers. The Whole-proof baseline (tactic scripts via beam search) does not represent modern whole-proof provers (DeepSeek-Prover-V2, Goedel-Prover-V2, Kimina-Prover), which use long reasoning, have-style subgoal sketches, and self-correction. Claims about "whole-proof supervision" should be scoped accordingly.
  • Agentic frameworks (e.g., Hilbert, and more recently LEAP and AlphaProof Nexus) choose granularity dynamically, as lemma/subgoal decomposition and compiler-feedback refinement rounds, rather than through a fixed segmentation rule. Discussing how open-goal boundaries relate to such decomposition would strengthen the paper's significance. (LEAP and Nexus may be contemporaneous, so I do not penalize the lack of an empirical comparison.)
Questions for authors
  1. What is the distribution of tactics per segment under the open-goal rule, and how does a random-boundary baseline with the same length distribution perform? How much gain remains without valid-prefix retention?
  2. How does GAR compare with unconditional rollout, or with a random trigger at the same frequency? How often is rollout triggered per theorem?
  3. Can you report results over multiple training seeds for at least one setting, along with paired significance tests? What causes run-to-run variance under beam search?
  4. Why does Qwen3-14B underperform Qwen2.5-Math-7B by 7–13 points across all granularities?
  5. What fraction of each corpus is discarded by parsing/replay, and what are the resulting dataset sizes per granularity?
  6. Training segments end at the next goal change, whereas GAR rolls out a fixed HH steps and ignores further goal changes. Have you tried terminating the rollout at the next goal change, which would match the training-time segment definition more closely?
  7. Could open-goal boundaries serve as a cheap decomposition or commit signal inside agentic provers, or do agentic designs make fixed segmentation rules unnecessary?
9 comments
AuthorsOpenReview ↗

Response to W1-1 && Q1

W1-1 && Q1

We thank the reviewer for the valuable comments on the independent contribution of the open-goal signal and the completeness of the ablation experiments.

First, we computed the distributions of the number of tactic steps and tokens per segment after segmenting the NuminaMath-LEAN dataset according to the open-goal signal. The results are shown in Table A.

Table A: Segment Length Distribution of Open-goal Supervision

LengthMeanMedianStdP90
Steps4.241534.20559
Tokens255.1111135322.2803664

Second, to isolate the effects of boundary positions and segment lengths, we constructed two randomly segmented training datasets from the same original proof trajectories: Random-Step and Random-Token, which respectively match the full tactic-step length distribution and token-length distribution of open-goal segmentation. We trained a model on each of these two datasets and evaluated the models on miniF2F and FATE-M under the same training configuration and evaluation protocol. The results are shown in Table B. The results show that Open-goal outperforms both random baselines on both evaluation benchmarks, indicating that the performance gains arise not only from segment length or multi-step tactic prediction but also from boundary selection based on open-goal changes.

Table B: Proof Success Rates of Open-goal and Random-boundary Baselines (Mean ± Std over Five Independent Runs)

MethodminiF2F Acc. (%) ↑FATE-M Acc. (%) ↑
Open-goal (Ours)66.31 ± 0.7917.20 ± 1.10
Random-Step64.51 ± 0.5513.47 ± 0.99
Random-Token65.08 ± 0.9814.27 ± 1.61

In addition, during multi-step tactic generation, we use valid-prefix retention to preserve valid prefixes that have passed Lean verification. Even if a subsequent tactic fails to execute, the local proof progress achieved by the preceding tactics can still be retained, and search can continue from the corresponding intermediate proof state. To ensure a consistent comparison, Open-goal, Random-Step, and Random-Token all use the same valid-prefix retention mechanism.

Finally, to address the concern that the ablation experiments covered only NuminaMath-LEAN → miniF2F, we kept the original model checkpoints fixed and followed the original evaluation protocol to additionally conduct three independent inference evaluations on FATE-M for each of the nine alternative segmentation settings in Table 4, reporting the mean ± sample standard deviation. For Open-goal, we retain the five evaluation results reported in the paper. As shown in Table C, Open-goal still achieves the highest mean success rate. This result is consistent with the observation on miniF2F and provides additional cross-benchmark evidence for the effectiveness of the open-goal signal.

Table C: Boundary-selection Ablation on FATE-M (Mean ± Std)

Boundary SelectionSettingFATE-M Acc. (%) ↑
Token-based32 tokens14.44 ± 1.92
Token-based64 tokens11.78 ± 0.38
Token-based96 tokens11.56 ± 2.78
Tactic-distance-basedThreshold 0.416.22 ± 2.04
Tactic-distance-basedThreshold 0.615.11 ± 0.77
Tactic-distance-basedThreshold 0.814.22 ± 1.68
State-distance-basedThreshold 0.415.56 ± 2.52
State-distance-basedThreshold 0.612.22 ± 0.38
State-distance-basedThreshold 0.812.67 ± 0.00
Open-goal (Ours)—17.20 ± 1.10
AuthorsOpenReview ↗

Response to W1-2 && Q2 && W1-3

W1-2 && Q2

We thank the reviewer for pointing out that the existing experiments did not sufficiently distinguish the contribution of the rollout mechanism from that of the open-goal trigger signal. We agree that comparison with standard best-first search alone is insufficient to verify the independent role of the open-goal trigger, and we have added random-trigger experiments with the same trigger rate. Table D reports the GAR trigger statistics for the two step-level provers in Table 3 of the paper on miniF2F and FATE-M, including the mean trigger rate and the mean number of triggers per theorem.

Table D: Trigger Statistics of Goal-aware Rollout

ModelBenchmarkTrigger Rate (%)Avg. Triggers per Theorem
InternLM2.5-StepProverminiF2F3.6520.96
InternLM2.5-StepProverFATE-M8.8254.08
BFS-Prover-V2-7BminiF2F14.56101.29
BFS-Prover-V2-7BFATE-M29.76245.60

To distinguish the gains from general lookahead from the contribution of the open-goal triggering strategy, we added a Random rollout baseline. Specifically, during inference, whenever a tactic executes successfully, Random rollout randomly decides whether to trigger a subsequent rollout with a probability corresponding to the mean GAR trigger rate reported in Table D.

The experimental results are shown in Table E. Random rollout yields some performance improvements over standard best-first search but underperforms goal-aware rollout in all four settings. This indicates that, compared with general lookahead rollout, triggering short rollouts based on changes in the number of open goals can guide search more effectively and further improve the proof success rate.

Table E: Comparison of Different Rollout Trigger Strategies (Mean ± Std over Five Independent Runs)

ModelBenchmarkBest-first (%)Random rollout (%)Goal-aware rollout (%)
InternLM2.5-StepProverminiF2F59.59 ± 0.3759.67 ± 0.4760.33 ± 0.45
InternLM2.5-StepProverFATE-M33.47 ± 2.0833.87 ± 1.7336.13 ± 0.73
BFS-Prover-V2-7BminiF2F68.77 ± 0.7969.10 ± 1.0370.74 ± 1.18
BFS-Prover-V2-7BFATE-M45.60 ± 1.3046.80 ± 1.3750.67 ± 0.82

W1-3

We thank the reviewer for pointing out the potential difference between the training and inference input distributions in the Whole-proof-seg experiments. This paper compares three supervision granularities: Step-level, Whole-proof, and Segment-level. The standard Whole-proof model takes the initial proof state as input and generates a complete proof during both training and inference. Whole-proof-seg is an additional inference setting that applies the model to segment-style search without changing the whole-proof training objective, to examine its performance under the same search interface. We acknowledge that this setting may involve some difference in the input distribution when inference is performed from intermediate proof states. Its results should therefore be understood as a supplementary comparison rather than a control experiment that strictly isolates the effect of supervision granularity.

AuthorsOpenReview ↗

Response to W2-1 && W2-2 && Q3 (Part 1)

W2-1 && W2-2 && Q3 (Part 1)

We thank the reviewer for their attention to the statistical reliability of the experiments. First, we clarify that the five independent runs reported in the original paper were all inference runs using the same fixed model checkpoint. Therefore, the reported mean ± std reflects randomness during inference rather than randomness during training. To further examine the effect of training randomness, we retrained the models using a different PyTorch random seed (seed 123, whereas the original experiments used seed 42) and conducted three independent inference evaluations on each of miniF2F and FATE-M. The results are shown in Table F. Across all six combinations of the three training datasets and two evaluation benchmarks, Segment-level achieves the highest mean proof success rate, consistent with the original experimental results.

Table F: Results with an Additional Training Seed (PyTorch Seed 123; Mean ± Std over Three Independent Runs)

Training SetTest SetStep-levelWhole-proofWhole-proof-segSegment-level (Ours)
LeanWorkbookminiF2F58.74 ± 0.6354.10 ± 0.4155.19 ± 1.2560.66 ± 0.41
LeanWorkbookFATE-M20.00 ± 0.6719.11 ± 0.3817.11 ± 0.3822.67 ± 0.67
NuminaMath-LEANminiF2F64.75 ± 0.7155.87 ± 0.2453.55 ± 0.2466.94 ± 0.24
NuminaMath-LEANFATE-M16.67 ± 0.6716.44 ± 1.6816.00 ± 0.6716.89 ± 1.02
STPminiF2F62.70 ± 1.4861.61 ± 0.8563.66 ± 0.2464.75 ± 0.71
STPFATE-M30.00 ± 1.3329.56 ± 2.6930.22 ± 0.7732.44 ± 1.02

In addition, we added statistical significance tests. For the training-side comparisons, we combine the five inference results from the original experiments with the three additional inference results under seed 123, for a total of eight evaluations. For the inference-side GAR comparisons, we use the five inference results from the original experiments. miniF2F and FATE-M contain 244 and 150 theorems, respectively. Because the different methods are evaluated on the same test set, we use a two-sided theorem-stratified exact conditional randomization test, preserving theorem-level pairing and testing differences in mean Pass@1 between methods [1]. We also apply Holm correction separately to the multiple training-side and inference-side comparisons [2]. Specifically, we treat miniF2F and FATE-M as separate families of tests and apply Holm correction to all relevant comparisons within each benchmark, rather than combining comparisons from the two benchmarks into a single family. The results are shown in Tables G and H.

Table G: Statistical Significance of Segment-level Supervision (Training-side Comparisons; Holm-adjusted p-values)

Training SetTest SetSegment-level vs. Step-level Δ Pass@1 (pp)Holm-adjusted pSignificantSegment-level vs. Whole-proof Δ Pass@1 (pp)Holm-adjusted pSignificantSegment-level vs. Whole-proof-seg Δ Pass@1 (pp)Holm-adjusted pSignificant
LeanWorkbookminiF2F+1.903.524×10−43.524\times10^{-4}Yes+6.354.469×10−364.469\times10^{-36}Yes+6.059.644×10−289.644\times10^{-28}Yes
LeanWorkbookFATE-M+1.580.04318Yes+2.750.006240Yes+5.421.958×10−121.958\times10^{-12}Yes
NuminaMath-LEANminiF2F+2.822.514×10−62.514\times10^{-6}Yes+10.664.857×10−494.857\times10^{-49}Yes+13.521.533×10−751.533\times10^{-75}Yes
NuminaMath-LEANFATE-M+1.000.615641No+0.830.615641No+1.420.316880No
STPminiF2F+1.840.0005690Yes+3.183.044×10−113.044\times10^{-11}Yes+1.230.01443Yes
STPFATE-M+2.080.035677Yes+1.750.052150No+1.830.045823Yes

Table H: Statistical Significance of Goal-aware Rollout (Inference-side Comparisons; Holm-adjusted p-values)

ModelTest SetGAR Avg. Pass@1 (%)Baseline Avg. Pass@1 (%)Δ Pass@1 (pp)Raw pHolm-adjusted pSignificant
InternLM2.5-StepProverminiF2F60.3359.59+0.740.1936740.193674No
BFS-Prover-V2-7BminiF2F70.7468.77+1.970.0015240.003048Yes
InternLM2.5-StepProverFATE-M36.1333.47+2.670.0433620.043362Yes
BFS-Prover-V2-7BFATE-M50.6745.60+5.070.0001510.000302Yes

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.