acceptodds
Under review as a conference paper at ICLR 2027

Target-Driven Automated Conjecturing for Neural Theorem Proving

Abstract

Specialised neural theorem proving systems are commonly trained through reinforcement learning on large collections of formal problems, often autoformalised from human-written sources. This dependence on human-curated data limits their scalability and prevents theorem-proving agents from improving autonomously through mathematical exploration, since each new training curriculum must ultimately be supplied by humans. To address this challenge, we introduce TACO (Target-Driven Automated Conjecturing), a framework that constructs theorem-proving curricula from model-generated conjectures. Since individual conjectures have small and noisy training effects, TACO evaluates them collectively: starting from a small seed set, it generates candidate conjecture sets, trains Prover variants on them, and scores each set by the improvement its corresponding Prover achieves on a fixed target set. This evolutionary procedure builds the curriculum stage by stage while iteratively updating the seeds and base Prover. Across four mathematical domains, using held-out problems selected from publicly available Lean datasets, TACO increases the number of held-out test problems solved within 32 attempts from 10/128 to 43/128 after ten iterations, while successful proof attempts increase from 16/4096 to 358/4096. These results demonstrate that model-generated conjectures can provide effective training curricula, enabling theorem proving systems to improve through targeted mathematical exploration while reducing reliance on human-written problems.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Formal theorem proving, Automated conjecturing, Curriculum learning, Reinforcement learning, Target-driven training

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_kNkcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TACO uses a fixed 27B conjecturer to generate training problems for a 2B prover. Separate prover copies are trained on candidate problem sets with GRPO, and their target-set performance determines prover updates and subsequent curriculum seeds. Across four domain-specific runs, held-out pass@32 coverage increases from 10/128 to 43/128 problems.

Critical strengths
  1. Selecting generated problems by their measured effect on prover performance is well motivated. It connects curriculum construction directly to the target learning objective rather than relying only on difficulty or relevance heuristics.
  2. The improvements extend to held-out problems in all four domains. The random-selection ablations also provide useful evidence about the contribution of curriculum selection.
Critical weaknesses
  1. The main improvement is measured against the initial SFT model rather than a comparable training baseline. In Table 2, random iterative seed selection with aligned initial seeds already solves 38/128 problems, compared with 43/128 for TACO. Every ablation retains target-based selection of the best prover. The experiments therefore do not isolate the value of the proposed selection mechanism from synthetic-data training and checkpoint selection.
  2. Appendix E evaluates each candidate curriculum after one training epoch. This favors immediate improvement, which need not identify the best longer-term curriculum. Ranking stability across training horizons is not tested, and one run per configuration leaves selection variability unresolved.
Questions for authors
  1. Does TACO outperform pooled or randomly selected synthetic curricula using the same 27B generator and total compute budget, without selecting among separately trained prover copies?
  2. How do curriculum rankings change with longer training horizons? Please report learning curves for fixed candidate sets and repeat the main comparison with independent runs.
3 comments

Thank you for the very thoughtful review! Your feedback was very helpful in improving our manuscript, and we are glad you found that selecting generated problems by their measured effect on prover performance is well motivated. We address your concerns and questions below and have updated our paper accordingly, marking the changes in blue for easy detection. If you have further concerns, please let us know. Otherwise, we would sincerely appreciate it if you would consider updating your review in light of our responses below.

Q1 / W1. How does TACO compare to training a single Prover on pooled conjectures?

We ran this comparison on the Radical Algebra domain, using the same initial Prover and fixed Conjecturer. At each iteration, we generate 64 conjectures from each of the 16 seeds and train a single Prover on all 16 x 64 = 1,024 conjectures using GRPO. We sample the 16 seeds for the next iteration uniformly from eligible conjectures: those proved in 1-7 of 8 attempts. We repeat this process for 10 iterations, retaining every update. The GRPO optimization settings were slightly adjusted for stable training.

MethodFinal pass@32Final pass rate
TACO16/3299/1,024
Single Prover on pooled conjectures3/3223/1,024

Both methods have a training budget of 10,240 conjectures and 81,920 proof attempts over 10 iterations. TACO solves more problems and produces more successful attempts. These results suggest that its selection process brings benefits beyond training a single Prover on all generated conjectures. One possible explanation is that the pooled Prover retains updates that may not benefit the target set, whereas TACO retains only updates that improve its measured target score.

These results and the experimental protocol have been added to Appendix F.1.

Q2a / W2. How do curriculum rankings change with longer training?

We are currently running an experiment on the Geometry: Area domain using all 16 candidate sets generated from the 16 seeds of the first iteration. Each set is used to train a copy of the same initial Prover for three epochs. All 16 candidate Provers are evaluated after epochs 1, 2 and 3 to compare the target-score curves and rank correlations. We will report the results as soon as they are available.

Q2b / W2. Do improvements recur across training runs?

We compare the original TACO run with one additional run on the Geometry: Area domain, using target-aligned seeds and the same initial models and target and test sets. We compare results after 10 iterations.

Training runFinal pass@32Final pass rate
Original run12/32170/1,024
Additional run14/3294/1,024

Both runs improve over their initial Prover evaluations, while their final coverage and number of successful attempts differ. This illustrates variability across runs: conjecture generation, GRPO training, and seed selection all involve randomness, and their effects can accumulate across iterations.

These results and the experimental protocol have been added to Appendix F.5.

Thank you for the new experiments and the revised manuscript. The pooled single-Prover baseline in Appendix F.1 addresses an important part of my original question. On Radical Algebra, the gap of 16/32 versus 3/32 is substantial and provides useful evidence for the full TACO procedure on that domain.

However, after considering the rebuttal, I am leaning toward keeping my original rating. The new baseline strengthens the paper, but I am still unsure how much of the gain comes from target-based Prover selection and whether it holds under a comparable end-to-end compute budget.

My main remaining question is what drives this gap and how much it costs. The comparison matches the number of training conjectures and rollouts, but the methods differ in seed selection, which Prover updates are retained, and some GRPO settings. Could you report the end-to-end compute for both runs, including target-set evaluation? A comparison that holds the candidate-training procedure and seed-selection rule fixed while varying only the Prover-update rule would help isolate the value of target-based Prover selection. Table 2 already gives a useful comparison of seed-selection rules under a shared Prover-update rule; the aggregate coverage difference there is 43/128 versus 38/128.

The additional Geometry: Area run is also helpful. Both runs improve on the initial Prover, although their final coverage and per-attempt success rates differ. The resampling analysis in Appendix F.7 addresses sensitivity to which target problems are included. I would be interested in the planned three-epoch results, particularly whether the candidates favored after one epoch remain competitive after three.

Thank you for this suggestion. We evaluated all 16 candidate Provers after one, two and three epochs in the first two iterations of an additional Geometry: Area run. More epochs do not consistently improve average target performance: the mean score (m=12) decreases from 0.792 to 0.654 in the first iteration, while increasing from 1.709 to 2.114 in the second. One possible explanation for the declines is that repeated updates on only 64 conjectures encourage over-specialisation to those problems, reducing transfer to the target set. The candidates ranked highest after one epoch do not consistently remain the strongest after three.

What is the end-to-end compute, including target-set evaluation? Appendix E estimates approximately 60 GPU-hours per ten-iteration TACO run, including conjecture generation, training all candidate Provers, their target-set evaluations, and seed and Prover selection on one NVIDIA GH200 GPU. For the pooled Radical Algebra baseline, Slurm accounting records 40.73 GPU-hours. This measured total includes initialisation, generation, training, target/test evaluations, and failed or restarted jobs, while excluding queue waiting.

Reviewer_Ljm1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a synthetic data generation method for neural prover training, in the spirit of evolutionary algorithm. A Qwen 32B iteratively augments data seeded on prior generated samples. The resulting augmented data is filtered by solvability, and scored by the actual performance improvement (on a validation set) when used to train the target Qwen 2B prover. The data that induces the largest improvements are preferred, and are more likely to be sampled as the seed sample for the next round of evolution. The data generator and the target model evolves simultaneously for 10 rounds, and are evaluated under a held-out test set within the same domains as the validation set used in training. Both pass@32 and pass@1 improves for the difficult problems which originally is challenging for the prover.

Critical strengths

Generalization. Even if the synthetic data is scored and selected on the performance of the validation set, the prover generalizes to a held-out test set. This demonstrates that the evolution algorithm does not overfit to the score on the val set, and the synthetic data is generalizable.

Critical weaknesses
  1. Compared to self-play style approaches, the paper's contribution is, by freezing the data generator, to trade the compute of target model training (and evaluation) for the compute of data generator training. However, it is unclear from experiments that the tradeoff is worthwhile. The self-play approaches are not set up as baselines, and are not compared with given equal compute.
  2. It is unclear whether the score selection is significantly better than uniform selection from the ablation studies. Pass@32 is close: 43/128 versus 38/128. There is no standard error estimation.
Questions for authors
  1. Can you provide equal compute (equal rollout/FLOPs) empirical results compared to self-play?
  2. Can you provide standard error estimation for the main results?
2 comments

Thank you for the thoughtful review. We appreciate your recognition of generalisation to held-out problems. We address the selection comparison and uncertainty estimates below; the corresponding manuscript changes are marked in blue. We would be grateful if you would reconsider your assessment in light of this evidence.

Q1 / W1. How does TACO compare with self-play?

We implemented a seed-selection rule inspired by Self-Play Theorem Prover (STP) on the Geometry: Area domain, with both target-aligned seeds and mixed seeds (8 target-aligned seeds, and 8 general seeds). The methods use the same initial models and target/test sets, with 16 candidate Provers per iteration. The proxy prioritises conjectures proved in 1–2 of 8 attempts and ranks them with an elegancy score, which corresponds to the ratio of their shortest correct proof over their statement length. The Conjecturer remains fixed and Prover selection remains target-based.

We compare results at iteration 5, which is the latest available test evaluation for these proxy runs.

Initial seedsMethodPass@32 at iteration 5Pass rate at iteration 5
AlignedTACO12/3292/1,024
AlignedSTP-inspired selection10/3246/1,024
MixedTACO11/3271/1,024
MixedSTP-inspired selection7/3255/1,024

TACO solves more problems and produces more successful attempts.

The STP-inspired selection run also takes longer, even though both methods train on the same number of conjectures and generate the same number of proof attempts during training. Averaged over the additional aligned and mixed runs, recorded processing time through five iterations is 20.6h for TACO versus 52.5h for the STP-inspired selection run (2.5 times longer). Evaluation logs show longer proof attempts for the proxy (307 output tokens on average, versus 24 output tokens for TACO) and substantially more time spent generating proofs.

These results and the selection protocol have been added to Appendix F.2.

Q2 / W2. What are the standard errors of the main results?

For the original pass@32 results, we estimate standard errors across test theorems within each domain and combine the four domains with equal weight.

MethodPass@32 (%)Approximate SE (percentage points)
Initial Prover7.812.39
TACO33.594.11
Random iterative selection, aligned seeds29.694.02

The gain over the initial Prover is large relative to these standard errors. The difference between selection rules is smaller. These estimates describe variation across test problems for the evaluated models.

These estimates and their calculation have been added to Appendix F.6.

Thank the author for their response.

My remaining concerns are:

  1. There's no comparison with self-play approaches in prior work where the conjecturer is not frozen.
  2. From the measured standard error, the score selection's advantage over uniform selection is not significant.

Therefore, I keep my original evaluation.

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

The paper introduces TACO (Target-Driven Automated Conjecturing), an iterative framework for constructing training curricula for neural theorem provers. Starting from a small seed set, a fixed Conjecturer generates candidate conjecture sets. Separate copies of the current Prover are trained on these sets using GRPO with Lean-verified rewards, then evaluated on a fixed target set. Their measured performance improvements guide subsequent seed selection and Prover updates. The central contribution is evaluating conjectures collectively through their observed downstream training utility.

Critical strengths
  1. The paper addresses an interesting problem: enabling LLM self-improvement in theorem proving. I find TACO’s focus on conjecture selection a particularly promising approach to this challenge.
  2. The experimental results demonstrate that TACO improves the prover’s ability to solve challenging problems.
Critical weaknesses
  1. Insufficient evidence for target-driven selection. The ablations do not establish a consistent benefit from target-driven selection. Filtering conjectures into the subset C^j(i)\hat{\mathcal{C}}_j^{(i)} with nonzero GRPO learning signals may itself contribute substantially to the overall gains. The contribution of this filtering should be distinguished from that of target-driven selection.
  2. Unclear generalization beyond the target distribution. Since the target and test sets share the same distribution, the evaluation does not establish whether the trained Prover preserves its performance in other mathematical domains. Cross-domain evaluation is needed to assess transfer and potential forgetting.
  3. Lack of strong baselines. Comparisons with closely related conjecture-based training methods are missing. A straightforward baseline would use the Conjecturer to augment the target/validation problems and train a single Prover on the resulting conjectures. Comparing against this baseline under a matched compute budget would help establish the framework’s added value.
  4. High computational cost. The framework trains a separate Prover for each candidate training set but retains only one, discarding most trained variants. This raises concerns about computational efficiency. Model merging could be explored to reuse the learning acquired by multiple candidates.
Questions for authors

See the weaknesses part.

2 comments

Thank you for the thoughtful review. We appreciate your view on the interestingness of the problem tackled and that conjecture selection is a promising approach to prover self-improvement. The changes described below are marked in blue in the paper. We would be grateful if you would reconsider your assessment in light of our responses.

W1. Does target-driven seed selection improve performance beyond the eligibility filter?

The eligibility filter determines which conjectures can become seeds for the next iteration, and each candidate Prover is trained on its full generated conjecture set. Requiring a successful proof avoids using false conjectures as seeds, while excluding seeds proved in all eight attempts keeps problems that remain challenging.

To examine the contribution of target-driven seed selection, we compare selection rules while keeping this filter fixed. We compare TACO with a seed-selection rule inspired by Self-Play Theorem Prover (STP): it prioritises conjectures proved 1–2 times, ranked by the ratio of shortest correct proof length to statement length. Both methods therefore keep the same eligibility filter, training procedure and target-based Prover selection.

On the Geometry: Area domain, with identical initial seeds (8 from the Geometry: Area domain, and 8 general seeds) and the same number of training and evaluation proof attempts, we obtain the following results after five iterations:

MethodPass@32 at iteration 5Pass rate at iteration 5
TACO11/3271/1,024
STP-inspired selection7/3255/1,024

This supports using measured training utility to choose among eligible seeds.

The STP-inspired selection also takes 2.5 times longer, despite both methods training on the same number of conjectures and generating the same number of proof attempts during training. Averaged over the additional aligned and mixed runs, recorded generation, training and evaluation time through five iterations is 52.5 h for the proxy versus 20.6 h for TACO. Evaluation logs show longer proof attempts for the proxy (307 output tokens on average, versus 24 for TACO). These observations suggest that prioritising difficulty can increase computational cost without improving test performance in this comparison.

This comparison has been added to Appendix F.2.

W2. Does specialization transfer to other domains?

We evaluate the final Prover (iteration 10) from our additional aligned Geometry: Area run, alongside the common initial Prover, on the four test sets. Each evaluation uses 32 proof attempts per problem.

Test domainInitial pass@32TACO pass@32Initial pass rateTACO pass rate
Radical Algebra2/325/324/1,02450/1,024
Geometry: Area3/3214/324/1,02494/1,024
Geometry: Volume3/3210/325/1,02455/1,024
Trigonometry0/320/320/1,0240/1,024

The Prover trained on the Geometry: Area domain also improves on Radical Algebra and Geometry: Volume, with no improvement on Trigonometry. Shared proof techniques and more effective use of Lean's automation may explain this transfer, while trigonometry may require more specialised identities and lemmas.

These results have been added to Appendix F.3.

W3. How does TACO compare with training on target variants?

We implemented the proposed baseline on the Geometry: Area domain with the same fixed Conjecturer: variants of 16 target problems are pooled into 1,024 conjectures per iteration, alternating two fixed groups of 16 target problems between iterations. One Prover is trained on the 1,024 pooled conjectures and retained at every iteration.

MethodFinal pass@32Final pass rate
TACO12/32170/1,024
Continued GRPO on target variants10/32157/1,024

Both have a budget of 81,920 training rollouts over ten iterations. TACO solves slightly more problems and produces more successful attempts. Its final Prover is trained on at most 10 x 64 = 640 generated conjectures, compared with 10,240 for the baseline, suggesting that TACO's selection can build a compact and effective curriculum.

We added this baseline, its optimisation settings, results and training budgets to Appendix F.1.

W4. Instead of selecting the best Prover variant, could we use model merging?

We had also considered this approach. Since the candidate Provers start from the same model, their parameter updates could be averaged with weights based on measured target improvement, allowing several candidates to contribute to the retained Prover. Doing this effectively would require tuning the merging weights and additional experiments to check whether combining the updates preserves or improves proving performance. We have added this direction to the limitations and leave model merging for future work.

Thank you for the additional experiments. The cross-domain evaluation and the single-Prover baselines substantially address parts of my concerns.

However, the incremental benefit of target-driven seed selection remains inconsistent relative to random selection in Table 2.

Therefore, I keep my original evaluation.

Reviewer_aGmgOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the curriculum bottleneck in neural theorem proving: specialised provers are trained by reinforcement learning on large collections of human-written formal problems, and every new curriculum must ultimately be supplied by humans. Recent work generates synthetic formal conjectures from a model, but selects which conjectures to train on using proxy criteria such as perceived difficulty, elegance, relevance to unsolved targets, or provability, none of which directly measures whether training on a conjecture improves the prover. The paper proposes TACO (Target-Driven Automated Conjecturing), which selects conjectures by measuring the downstream improvement they induce. Because the marginal effect of a single conjecture is small and noisy, TACO scores whole candidate sets: from a seed set, a fixed Conjecturer generates one candidate set of Lean conjectures per seed, a separate copy of the current Prover is trained on each set with GRPO using proof-checker reward, every variant and the current Prover are evaluated on a fixed target set, and each set receives the difference in a performance metric between its variant and the current Prover. The performance metric sums, over target problems, one minus the probability of zero successes in m attempts, interpolating between pass rate and coverage. Candidate sets scoring above the mean receive positions in the next seed set in proportion to their excess over the mean, with seeds drawn only from conjectures proved in some but not all rollouts; the highest-scoring variant becomes the next base Prover if its score is positive. The loop runs for ten iterations over four mathematical domains (radical algebra, geometry of areas, geometry of volumes, trigonometry), using Qwen3.5-2B as Prover and a fixed Qwen3.8-27B as Conjecturer. On held-out test sets, the number of problems solved at least once in 32 attempts rises from 10/128 to 43/128 and successful attempts from 16/4096 to 358/4096. A two-factor ablation crosses target-aligned versus random initial seeds with target-driven versus random iterative selection, and reports that target-driven selection beats random only in radical algebra, is mixed in trigonometry and geometry of areas, and is worse in geometry of volumes. Additional analyses report increased explicit references to domain lemmas and automation tactics, and per-domain examples of seeds, conjectures, and targets.

Critical strengths
  1. The paper correctly identifies that existing conjecture selectors optimise proxies — difficulty, elegance, relevance, provability — and that none of them measures whether training on the conjecture actually helps
  2. Set-level credit assignment is a principled response to a real statistical obstacle.
  3. The ablation reports inconvenient results instead of hiding them. Table 2 shows that target-driven selection is better than random only in radical algebra, mixed in trigonometry and geometry of areas, and worse in geometry of volumes, and the text says so directly
Critical weaknesses
  1. The central claim — that measured utility is a better selection signal than the proxy criteria used by prior work — is never tested against those criteria. The paper positions TACO against Minimo's difficulty criterion, STP's barely-provable-plus-elegance rule, and SGS's relevance/quality Guide, and states that "TACO instead keeps the Conjecturer fixed and scores whole conjecture sets by the measured target improvement" (page. 9; lines 439–447). But the experiments compare only target-driven selection against random selection within TACO's own loop. No experiment runs a proxy-criterion selector on the same seeds, target set, and compute. The paper's own Table 2 makes this gap consequential: target-driven selection beats random in radical algebra, is mixed in two domains, and is worse in geometry of volumes (Table 2), so the internal evidence for the selection rule is weak in exactly the domains where a comparison to STP/SGS would be most informative. This is a fundamental missing comparison for the paper's stated contribution, and it is not one of the limitations the authors acknowledge.
  2. There is no compute-matched human-data baseline, so the gains cannot be attributed to conjecture quality rather than to additional training. TACO spends about 940 GPU-hours across the reported runs (Appendix E), and the evaluation pool from which seeds, targets, and tests are drawn contains 26,418 unique statements (lines 269–283). A natural control — continue GRPO or SFT on additional human-written problems from that pool for the same compute, and evaluate on the same held-out test sets — is absent. Without it, the paper shows that iterative training on generated conjectures improves the Prover relative to the initial Prover, but not that model-generated conjectures are a better use of compute than the human-written data already available. Since reducing reliance on human-curated data is the paper's motivating claim, this control is needed to separate "more training" from "better data".
  3. The framing attributes the result to the target-driven component, while the paper's own ablation indicates that most of the gain comes from the iterative loop itself. The abstract and introduction nevertheless present the target-driven evaluation as the contribution that makes the approach work, and the Discussion concedes only that the ablations "show a domain-dependent benefit from target-driven seed selection" (lines 474–477). The result may well be real, but the paper currently over-attributes it, and the honest reading is that conjecture-based training is effective while the specific selection rule is a modest and inconsistent improvement.
Questions for authors
  1. Can you run one proxy-criterion selector under the same protocol? Concretely, on the same seeds, target set, number of Prover variants, and evaluation budget, compare target-driven set scoring against (a) a difficulty filter in the style of Minimo, (b) a barely-provable rule with an elegance filter in the style of STP, or (c) the relevance/quality Guide of SGS.
  2. What is the compute-matched human-data control? If the same 940 GPU-hours were spent on additional GRPO/SFT over the 26,418-statement pool, how does the resulting Prover compare on the same test sets?
  3. How much of the target-set gain is winner's curse? What is the distribution of measured target scores across variants that are statistically indistinguishable, and what happens to the reported held-out gain if the Prover update is chosen by a held-out slice of the target set rather than by the full target set?
2 comments

Thank you for the very thoughtful review! Your feedback was very helpful in improving our paper, and we particularly appreciate your comment that set-level credit assignment is a principled response to a real statistical obstacle. We address your concerns and questions below and have updated our paper accordingly, marking changes in blue for easy detection. If you have further concerns, please let us know. Otherwise, we would sincerely appreciate it if you would consider updating your review in light of our responses below.

Q1 / W1. How does measured utility compare with a proxy selector?

We implemented a seed-selection rule inspired by Self-Play Theorem Prover (STP) on the Geometry: Area domain, with both target-aligned seeds and mixed seeds (8 target-aligned seeds, and 8 general seeds). The methods use the same initial models and target/test sets, with 16 candidate Provers per iteration. The proxy prioritises conjectures proved in 1-2 of 8 attempts and ranks them with an elegancy score, which corresponds to the ratio of their shortest correct proof over their statement length. The Conjecturer remains fixed and Prover selection remains target-based.

We compare results at iteration 5, which is the latest available test evaluation for these proxy runs.

Initial seedsMethodPass@32 at iteration 5Pass rate at iteration 5Mean recorded time (h)
AlignedTACO12/3292/1,02420.6
AlignedSTP-inspired selection10/3246/1,02452.5
MixedTACO11/3271/1,02420.6
MixedSTP-inspired selection7/3255/1,02452.5

TACO solves more problems and produces more successful attempts. At iteration 5, 87.5% of the proxy's conjectures receive no correct proof in eight training attempts, versus 11.2% for TACO. This suggests that prioritising difficulty can produce a curriculum beyond the current Prover's capabilities, whereas measured training utility helps maintain useful training examples.

The STP-inspired selection run also takes longer, even though both methods train on the same number of conjectures and generate the same number of proof attempts during training. Evaluation logs show longer proof attempts for the proxy (307 output tokens on average, versus 24 output tokens for TACO) and substantially more time spent generating proofs.

These results and the selection protocol have been added to Appendix F.2.

Q2 / W2. How does TACO compare with training on human-written data at the same compute budget?

The 940 GPU-hours covered all 16 runs presented in the paper's Table 2, the approximative budget for one run is about 60 GPU-hours. We are currently running a continued SFT run that trains the initial Prover for 60 additional GPU-hours on the original human-written training split of 16,673 statement-proof pairs. We will report the results as soon as they are available.

Q3. How stable are candidate rankings, and could target selection favor sampling noise?

We analysed the 16 candidates at each of ten iterations in one additional aligned Geometry: Area run. At each iteration, we resampled the 32 target problems 20,000 times, using the same sampled problems for all candidates. The median rank correlation with the original ranking is 0.80 (Spearman), and the original top candidate remains first in 50.1% of resamples and among the top three in 79.6%. The overall ranking therefore retains substantial structure.

These diagnostics have been added to Appendix F.7.

W3. What does target-driven selection contribute?

Target performance guides both seed selection and Prover selection. Table 2 changes the seed-selection rule while retaining target-based Prover selection. These ablations therefore assess seed selection within a training loop that remains guided by target performance.

We ran an additional comparison on the Radical Algebra domain, using the same initial Prover and fixed Conjecturer. At each iteration, we generate 64 conjectures from each of the 16 seeds and train a single Prover on all 1,024 conjectures, retaining every update. We sample the 16 seeds for the next iteration uniformly from conjectures proved in 1–7 of 8 attempts.

MethodFinal pass@32Final pass rate
TACO16/3299/1,024
Single Prover on pooled conjectures3/3223/1,024

Both methods have a training budget of 10,240 conjectures and 81,920 proof attempts over 10 iterations. These results support the value of TACO's full selection procedure compared with training a single Prover on all generated conjectures. One possible explanation is that the pooled Prover retains updates that may not benefit the target set, whereas TACO retains only updates that improve its measured target score.

These results and the experimental protocol have been added to Appendix F.1.

Q2 / W2. How does TACO compare with training on human-written data at the same compute budget?

We have now completed the continued-SFT experiment described in our previous response. Starting from the initial Prover, we trained for 58.7 additional GPU-hours on its original training set of 16,673 human-written statement-proof pairs. We evaluated the final checkpoint on all four test sets, using the same evaluation settings and 32 proof attempts per problem. The table compares the re-evaluated initial Prover, continued SFT and the TACO Prover trained on each corresponding domain, using the original results from Table 1. Each TACO run uses approximately 60 GPU-hours. Each cell reports Pass@32; pass rate:

Test domainInitial ProverContinued SFTTACO (per domain)
Radical Algebra2/32; 4/1,0243/32; 7/1,02416/32; 99/1,024
Geometry: Area3/32; 4/1,0245/32; 11/1,02412/32; 170/1,024
Geometry: Volume3/32; 5/1,0245/32; 6/1,0249/32; 75/1,024
Trigonometry0/32; 0/1,0240/32; 0/1,0246/32; 14/1,024

Each domain-specific TACO Prover outperforms continued SFT on its target domain at a comparable GPU-hour budget. This supports using TACO to improve performance on a specific domain beyond further SFT on the same human-written data.

The additional Prover trained on the Geometry: Area domain and evaluated across all four domains in Appendix F.3 also outperforms continued SFT on Radical Algebra (5/32 versus 3/32) and Geometry: Volume (10/32 versus 5/32), providing evidence of transfer beyond its target domain.

These results and the training protocol have been added to Appendix F.1 (Table 9).

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

The paper proposes TACO, a framework for constructing theorem-proving curricula from generated conjectures. It trains separate prover copies on candidate conjecture sets and uses their improvement on a fixed target set to select the next prover and training problems. Experiments across four mathematical domains increase held-out theorem coverage from 10/128 to 43/128 within 32 attempts per problem. The main idea is to assess conjecture sets through their observed training utility.

Critical strengths
  1. Selecting conjectures according to their actual effect on prover performance is a useful direction. Evaluating sets of conjectures provides a concrete way to implement this idea.
  2. The method improves both theorem coverage and proof success rates across all four domains relative to the initial prover. The test set is kept outside the iterative selection loop.
  3. The paper is clearly written and candid about its findings. In particular, it acknowledges the mixed selection results and the limitations of interpreting tactic and lemma frequencies.
Critical weaknesses
  1. The experiments do not establish the benefit of the central curriculum-search procedure. All variants in Table 2 retain target-based selection among separately trained prover copies. With the same initial theorems, random iterative selection already solves 38/128 test problems, compared with 43/128 for TACO. The difference is concentrated in radical algebra, where TACO solves 16 rather than 9 problems; coverage is lower in both geometry domains. Given up to 16 candidate training runs per iteration and about 60 GPU-hours per complete run, a comparison with directly continuing GRPO on generated conjectures under a matched compute budget is important. The authors acknowledge the attribution issue, but it remains the main gap in the evidence.

  2. The separation between generated training data and evaluation data is not fully checked. Section 4.1 states that generated conjectures are not systematically compared against target or test statements. Deduplicating the original source datasets does not cover these newly generated examples. An explicit overlap audit would strengthen the interpretation of the held-out improvements.

Questions for authors
  1. How does TACO compare with directly continuing GRPO on conjectures from the same generation pipeline, under a matched total compute budget, without training and selecting among separate prover copies?
  2. Can you check the generated training conjectures against the target and test statements and report any overlap? Normalized exact matching and a check for simple template variants would be useful.
1 comment

Thank you for the very thoughtful review! Your feedback was very helpful in improving our manuscript, and we are glad you found that selecting conjectures by their effect on prover performance is a useful direction, and that the paper is clearly written. We address your concerns and questions below and have updated our paper accordingly, marking the changes in blue for easy detection. If you have further concerns, please let us know. Otherwise, we would sincerely appreciate it if you would consider updating your review in light of our responses below.

Q1 / W1. How does TACO compare to training a single Prover on pooled conjectures?

We ran this comparison on the Radical Algebra domain, using the same initial Prover and fixed Conjecturer. At each iteration, we generate 64 conjectures from each of the 16 seeds and train a single Prover on all 16 x 64 = 1,024 conjectures using GRPO. We sample the 16 seeds for the next iteration uniformly from conjectures proved in 1-7 of 8 attempts. We repeat this process for 10 iterations, retaining every update. The GRPO optimization settings were slightly adjusted for stable training.

MethodFinal pass@32Final pass rate
TACO16/3299/1,024
Single Prover on pooled conjectures3/3223/1,024

Both methods have the same training budget of 10,240 conjectures and 81,920 proof attempts over 10 iterations. TACO solves more problems and produces more successful attempts. These results suggest that its selection process brings benefits beyond training a single Prover on all generated conjectures. One possible explanation is that the pooled Prover retains updates that may not benefit the target set, whereas TACO retains only updates that improve its measured target score.

These results and the experimental protocol have been added to Appendix F.1.

Q2 / W2. Do generated conjectures overlap with evaluation problems?

In the original experiments, our normalised exact-match check found no matches between generated training conjectures and target or test problems.

To investigate overlap more thoroughly, we extended the analysis to template variants. We audited 20,480 training conjectures from two further TACO runs on Geometry: Area, using the same target and test sets as in the paper and including unretained candidates.

We again found zero normalised exact matches. Checks for variable renaming and numerical variants identified one numerical variant of a test problem. The corresponding Prover was not retained, and that candidate set supplied no seeds to the next iteration. This overlap therefore did not contribute to the final Prover's training.

We have clarified the original check in Section 4.1 and added the extended audit to Appendix F.4.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal summary: additional baselines and cross-domain transfer

We thank all reviewers and Area Chairs for their thoughtful feedback and constructive discussion. TACO's central idea is to evaluate generated conjecture sets through their measured effect on Prover performance after training on them, using this signal to guide both seed selection and Prover updates. The reviews primarily asked for comparisons with baselines under comparable computational budgets and further evaluations of transfer across mathematical domains. Our new experiments show that TACO compares favourably with the tested baselines and that the improvements extend to other domains. We summarise these findings below. The corresponding manuscript changes are marked in blue, and our individual responses provide further details.

1. How does TACO compare with training a single Prover on all conjectures generated from seeds or on variants of target problems?

We added two comparisons using the same initial models and the same number of training conjectures and proof attempts over ten iterations. One trains a single Prover on all conjectures generated from seeds; the other trains it on variants of target set problems.

Radical Algebra

MethodPass@32Pass rate
TACO16/3299/1,024
Single Prover on pooled conjectures3/3223/1,024

Geometry: Area

MethodPass@32Pass rate
TACO12/32170/1,024
Single Prover on target variants10/32157/1,024

These results support the value of TACO's full selection procedure. A possible explanation is that selecting updates by measured target improvement avoids retaining updates that do not benefit the target set. These comparisons are presented in Appendix F.1.

2. How does measured training utility compare with proxy-based selection?

We implemented a seed-selection rule inspired by Self-Play Theorem Prover (STP), while leaving the rest of the TACO pipeline unchanged. Results on the Geometry: Area domain after five iterations are:

Initial seedsMethodPass@32Pass rate
AlignedTACO12/3292/1,024
AlignedSTP-inspired selection10/3246/1,024
MixedTACO11/3271/1,024
MixedSTP-inspired selection7/3255/1,024

TACO solves more problems and produces more successful attempts in both comparisons. The proxy also takes 2.5 times longer, averaging recorded generation, training and evaluation times over the additional aligned and mixed runs. Further analysis suggests that prioritising difficulty can generate conjectures beyond the current Prover's capabilities, supporting the use of measured training utility to build useful curricula. These results are presented in Appendix F.2.

3. Do the improvements transfer beyond the target domain?

We evaluated the final Prover from our additional aligned Geometry: Area run on all four test sets, alongside the initial Prover. The table shows initial → final performance:

Test domainPass@32Pass rate
Radical Algebra2/32 → 5/324/1,024 → 50/1,024
Geometry: Area3/32 → 14/324/1,024 → 94/1,024
Geometry: Volume3/32 → 10/325/1,024 → 55/1,024
Trigonometry0/32 → 0/320/1,024 → 0/1,024

The improvement extends to two domains beyond the one used for target-based selection, providing evidence of transfer. These evaluations are presented in Appendix F.3.

4. How does continued training on human-written data compare?

We completed 58.7 additional GPU-hours of SFT (matching a TACO run compute) on the initial Prover's original 16,673 human-written statement-proof pairs. Each cell reports Pass@32; pass rate:

Test domainInitial ProverContinued SFTTACO (per domain)
Radical Algebra2/32; 4/1,0243/32; 7/1,02416/32; 99/1,024
Geometry: Area3/32; 4/1,0245/32; 11/1,02412/32; 170/1,024
Geometry: Volume3/32; 5/1,0245/32; 6/1,0249/32; 75/1,024
Trigonometry0/32; 0/1,0240/32; 0/1,0246/32; 14/1,024

Each domain-specific TACO Prover outperforms continued SFT on its target domain at a comparable GPU-hour budget. The single additional Area-trained Prover also outperforms continued SFT on Radical Algebra and Geometry: Volume, showing transfer beyond its target domain. These results and the protocol are presented in Appendix F.1.

Thank you again for helping us improve the paper!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.