acceptodds
Under review as a conference paper at ICLR 2027

Veri-GPS: Learning to Coordinate Geometric Proof and Calculation with Verification and Solver Feedback

Abstract

Large language models (LLMs) can solve geometric proof problems with the support of symbolic solvers. However, many geometry problems require more than correct geometric relations, as proof and calculation must also support each other. We introduce Veri-GPS, an interaction learning framework that learns subgoal policies from symbolic solver feedback to coordinate proof and calculation within a given interaction budget. An extended and accelerated Newclid backend enables geometric proof and exact calculation to share verified results. When execution is blocked, goal-related diagnostic feedback identifies missing premises and quantities along candidate rule or formula paths to limit LLM trial and error. To train the LLM to use solver feedback, we construct a dual-source pipeline combining teacher–solver interaction trajectories with Goal Augmentation trajectories. Both sources are split into turn-level training data for supervised fine-tuning (SFT) of Qwen2.5-VL-7B-Instruct, providing a cold start for the interaction policy. We introduce turn-level Dynamic-Difficulty Reinforcement Learning (DDRL), combining adaptive sampling across fixed difficulty buckets with derivation-use rewards for intermediate results used in subsequent reasoning. Across 3,082 problems from Geometry3K, PGPS9K, the planar geometry subset of MathVerse, GeoQA+, and GeoQA, Veri-GPS achieves a problem-weighted pass@1 of 82.12%. Compared with standalone symbolic execution and Qwen3-VL-32B-Instruct direct answering, Veri-GPS improves performance by 10.25 and 17.39 percentage points, respectively. Within Veri-GPS, full feedback yields a 1.27-point gain over reduced feedback at inference with the trained policy fixed, while DDRL improves pass@1 by 3.57 points over the SFT checkpoint.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Large Language Model, Reinforcement Learning, Interactive Geometry Solving

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_S53TOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Veri-GPS to train LLMs from symbolic solver feedback to organize subgoals across geometric proof and calculation within a bounded interaction budget. The authors extend the Newclid symbolic solver and introduce Goal-related Diagnostics to identify missing premises. They combine Teacher–Solver Interaction trajectories with Goal Augmentation for supervised fine-tuning (SFT), followed by Dynamic-Difficulty Reinforcement Learning (DDRL) to further optimize the subgoal-selection policy. Experiments on 3,082 problems across five geometry benchmarks achieve an overall pass@1 of 82.12%.

Critical strengths
  1. A relatively complete neural-symbolic collaborative solving framework. This paper expands the Newclid backend by combining geometric proofs with exact calculations and introducing goal-related diagnostics to provide feedback for subgoal selection and revision in the model. This design allows the system to make use of verified intermediate results over multiple interactions, offering a useful framework for geometric reasoning.
  2. Strong experimental performance across multiple benchmarks. This paper achieved an overall pass@1 of 82.12% on 3,082 problems from five geometry datasets, which is an improvement compared to the standalone symbolic solver.
Critical weaknesses

There's a lack of a simple baseline for diagnostic-guided planning. While the paper shows improvements over a standalone symbolic solver, it does not compare the learned subgoal-selection policy with a simple heuristic planner using the same Newclid backend and Goal-related Diagnostics. This makes it hard to judge how much extra benefit the learned subgoal-selection policy actually provides compared to just using diagnostic guidance directly.

Questions for authors

What's the difference between Veri-GPS and RAVS (Agentic Geometry Problem Solving via Human-like Parallel Bidirectional Reasoning), especially in terms of subgoal planning, solver feedback, and proof search strategies? Figuring out these differences helps evaluate the novelty of the proposed method.

2 comments

Dear Reviewer S53T,

Thank you again for your careful review and helpful suggestions. We have added further experiments and clarifications in our responses and revised manuscript.

As the discussion period is nearing its end, we hope to hear your thoughts. If our responses have addressed your concerns, we would be grateful for your confirmation. If any concerns remain, we would welcome the opportunity to discuss them with you before the discussion closes.

Regards from the authors.

Thank you for the  rebuttal. The authors have addressed my concerns and clarified the issues raised in my  review. I therefore have no further major concerns and will maintain my original score of 3.

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

This paper proposes Veri-GPS, a framework that learns to coordinate geometric proof and calculation through interactions with a symbolic solver. The authors extend Newclid to share verified relations and quantities, then train a subgoal policy using supervised fine-tuning and reinforcement learning with solver feedback. Experiments on five geometry datasets show that the trained policy improves problem-solving performance over the standalone backend and the SFT model.

Critical strengths
  1. The framework allows proof and calculation to support each other through a shared verified state. The description of the solver interface, formula registry, and dependency tracking makes the system reasonably clear and concrete.

  2. The experiments support the effectiveness of policy learning. DDRL improves overall pass@1 by 3.57 points over SFT and by 3.05 points over uniform-sampling RL. The reward ablation also supports the contribution of rewarding intermediate results used in later reasoning.

Critical weaknesses
  1. The backend comparison in Table 2 uses each system's own saved formalizations. Differences in the input formalizations may therefore contribute to the reported accuracy gains, making it difficult to determine how much improvement comes from the extended solver. A comparison on problems that both backends can represent, using equivalent givens, would help clarify this point.

  2. Some component improvements are small. Combined supervision solves only seven more problems than Teacher-only, and full-feedback SFT solves nine more than reduced-feedback SFT, while Appendix C.5 reports only one trained model per data-source setting. I would like to see whether these gains are stable across runs. The efficiency comparison also averages requests over each model's own solved problems, so the decrease from 6.752 to 5.464 does not establish a reduction in cost on the same problems.

Questions for authors
  1. How are the saved formalizations generated and checked? Are problems with incorrect formalizations included in the end-to-end evaluation?
  2. Are results across training seeds or paired per-problem outcomes available for the policy and feedback comparisons?
  3. What are the request counts and solver times on problems solved by both models? What is the total cost when unsuccessful attempts are included?
2 comments

Dear Reviewer bMYW,

Thank you again for your careful review and helpful suggestions. We have added further experiments and clarifications in our responses and revised manuscript.

As the discussion period is nearing its end, we hope to hear your thoughts. If our responses have addressed your concerns, we would be grateful for your confirmation. If any concerns remain, we would welcome the opportunity to discuss them with you before the discussion closes.

Regards from the authors.

Thank you for the detailed response. The common source formalizations address my concern about the backend comparison. The three-seed SFT results and the paired comparison on jointly solved problems also provide useful support for the reported gains.

I agree with the distinction you make between fewer proposals and higher total symbolic execution cost. This tradeoff should remain clear in the paper, since lower latency on jointly solved problems does not imply lower cost over the full test set. My main concerns have been addressed, and I remain at weak acceptance.

Reviewer_7nzXOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents Veri-GPS, a framework that trains an LLM to select and revise geometric subgoals through interaction with a symbolic solver. An extended Newclid backend supports proof and exact calculation through shared verified results and diagnostic feedback. The policy is initialized using teacher–solver and goal-augmentation trajectories, then optimized with Dynamic-Difficulty Reinforcement Learning (DDRL), which combines adaptive sampling across fixed difficulty buckets with rewards for useful intermediate derivations. The system achieves 82.12% overall pass@1 across five benchmarks.

Critical strengths
  1. The framework provides a coherent approach to coordinating proof and calculation. Verified intermediate results can support later operations, while diagnostic feedback helps the policy revise blocked subgoals. This offers a concrete setting for studying how LLMs learn to use tools over multiple steps.
  2. The internal comparisons provide meaningful evidence for the training approach. Table 1 reports 82.12% completion for DDRL, compared with 78.55% for SFT and 79.07% for uniform-sampling RL under the same interaction limits. Removing the derivation-use reward reduces completion to 80.18%, supporting its contribution.
  3. The analysis extends beyond final accuracy to feedback-guided revision, reuse of intermediate results, and solver execution costs. The backend, data-source, feedback, and training comparisons make the empirical study reasonably comprehensive.
Critical weaknesses
  1. The benefit of the learned interaction policy over prompting a strong model to use the same solver remains unclear. DeepSeek Flash is included in Table 1, but only in direct-answering mode. A comparison with this model using the same formalized inputs, solver interface, feedback, and interaction budget would help establish the practical value of the learned policy.
  2. The difficulty measure needs further validation. Appendix E.3 defines problem difficulty through the 70th percentile of successful-turn demonstration losses under a frozen SFT model. Such losses may reflect demonstration wording or model familiarity as well as actual solving difficulty. The comparison with uniform sampling supports the complete sampling strategy, but does not establish that this particular ordering captures the intended notion of difficulty.
  3. The benefits are less clear on problems requiring auxiliary constructions. On GeoLaux, Table C5 reports 42.82% pass@1 for Veri-GPS, compared with 63.85% for DeepSeek Flash direct answering. The authors acknowledge this limitation, but the results leave unclear how much of the gap comes from solver coverage versus subgoal and construction selection.
Questions for authors
  1. Could you evaluate the DeepSeek Flash model used in Table 1, or a recent strong model, as a prompted controller on a representative subset, with the same formalized inputs and access to solver feedback under the same interaction budget as Veri-GPS? This would clarify the benefit of the learned policy relative to a strong prompted baseline.
  2. Using existing rollout logs, could you show how the fixed difficulty buckets relate to observed success rates before the monotonic calibration described in Appendix E.3? This would help establish whether demonstration likelihood provides a meaningful curriculum signal. Other analyses clarifying the difficulty measure would also be welcome.
2 comments

Dear Reviewer 7nzX,

Thank you again for your careful review and helpful suggestions. We have added further experiments and clarifications in our responses and revised manuscript.

As the discussion period is nearing its end, we hope to hear your thoughts. If our responses have addressed your concerns, we would be grateful if you could reconsider your assessment and rating. If any concerns remain, we would welcome the opportunity to discuss them with you before the discussion closes.

Regards from the authors.

Thank you for the additional experiments. They address my concerns, and I have no further questions. I will raise my rating.

Comments on the paper

2 comments
AuthorsOpenReview ↗

A note of thanks before the discussion ends.

Dear Area Chair and Reviewers,

With about two days left in the author–reviewer discussion period, we would like to thank you again for your time and for reading our work so carefully. Your suggestions helped us strengthen the controller comparisons and explain more clearly what the 7B policy learns.

We have uploaded the revised manuscript, with changes highlighted in blue, and posted our point-by-point responses under each review. We would appreciate hearing whether the revisions and responses address your concerns, and we would welcome the opportunity to discuss any remaining questions with you.

Regards from the authors.

AuthorsOpenReview ↗

A Heartfelt Thank-You to Veri-GPS's Reviewers

Dear Reviewers,

Thank you for taking the time to read our paper so carefully and share your thoughts. Your feedback has truly helped us explain Veri-GPS more clearly and improve our work.

Before the discussion ends, we wanted to thank you once more and send you our warmest wishes. Although we have never met, ICLR has brought us together through this exchange. Your time, care, and encouragement have made our research journey feel a little warmer.

We wish you all the best with your research and work, and lots of happiness in everyday life.

Regards from the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.