acceptodds
Under review as a conference paper at ICLR 2027

SEVA-Lyap: Solver-Aware Data Evolution and Verification-Guided Test-Time Adaptation for Symbolic Lyapunov Function Discovery

Abstract

Lyapunov functions provide fundamental certificates for the stability of dynamical systems, yet discovering explicit symbolic Lyapunov functions remains a long-standing computational challenge due to the vast search space and strict mathematical constraints. Recently, symbolic transformers have advanced this direction by formulating Lyapunov discovery as a sequence-to-sequence translation task, directly predicting interpretable Lyapunov functions from system dynamics. Despite this pioneering advance, these models face two critical bottlenecks: they rely on static synthetic datasets fixed prior to training, and they lack mechanisms to adapt at test time. Consequently, rare regions of the system distribution can remain underrepresented, making novel or hard systems difficult to solve. To address these limitations, we introduce SEVA-Lyap, a unified framework integrating solver-aware data evolution with verification-guided test-time adaptation. During training, SEVA-Lyap continuously evolves verified examples around the weaknesses of the current solver, creating a progressively challenging curriculum that strengthens reusable symbolic solving capability. At inference, SEVA-Lyap allocates additional test-time compute through targeted risk-seeking adaptation when direct prediction fails. Extensive evaluations demonstrate that SEVA-Lyap raises system-level verification success from 70.75% to 92.00% on polynomial systems and from 60.50% to 74.50% on non-polynomial systems, while reliably solving representative challenging dynamics and demonstrating notable transferability. Anonymous code and data are available at https://anonymous.4open.science/r/SEVA-review-DBD6-0921.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Lyapunov functions, Transformer, data evolution, test-time adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rHenOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents SEVA-Lyap, a framework for discovering symbolic Lyapunov functions by combining adaptive data generation with test-time refinement. Instead of relying on a fixed training dataset, the method uses feedback from the current solver to generate increasingly challenging training examples. At test time, it further uses unsuccessful predictions through verification-guided reinforcement learning. Experiments on polynomial and non-polynomial systems show that SEVA-Lyap improves verification success from 70.75% to 92.00% and from 60.50% to 74.50%, respectively, compared with the Symbolic Transformer baseline.

Critical strengths
  1. Well-motivated approach: Motivation is clear, stability is one of the fundamental problems in control theory.

  2. Strong empirical results: This method improves upon previous work by a good margin.

  3. Comprehensive evaluation: Authors do a great job at ablation.

Critical weaknesses
  1. Concerns about the verification: The most significant concern is that the SHGO verifier does not appear to enforce strict positive definiteness, which is essential for a valid Lyapunov function. Appendix A accepts candidates based on approximate nonnegativity, potentially allowing invalid solutions. For example, consider the unstable system x˙1=−x1, x˙2=x2\dot{x}_1=-x_1,\ \dot{x}_2=x_2 with V(x)=x12V(x)=x_1^2. Although V˙=−2x12≤0\dot V=-2x_1^2\leq0, the candidate is not positive definite because it vanishes away from the origin. Nevertheless, it satisfies the stated acceptance conditions. The authors should clarify whether their implementation includes additional checks (that is not mentioned in the paper, I did not see one) to rule out such cases.

Alfarano et al, NeurIPS 2024 (which is cited in this paper) explicitly mentions this. See Def 2.2 and Evaluation paragraph in Section 3.

  1. Fundamental problem: I do not understand how this method can be applied to real-world systems. Typically, unstable trajectories lead to disastrous consequences. I do not think it is reasonable to assume we can use such violations indefinitely.

  2. Limited generalization: Most of the benchmarks are generated using procedures similar to those used during training. While the authors also test on independently generated systems, the low success rates make it difficult to tell how well the method would work on a wider range of systems.

  3. Limited novelty: The method combines several existing ideas, such as adaptive data generation, policy-gradient learning, replay, and GRPO. Although these techniques work well together, it is not entirely clear what the paper adds beyond combining them.

Questions for authors

See weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer rHen (1/4)

Thank you for your careful review and constructive comments, which have helped us improve the manuscript. We have also revised the main text to clarify the novelty of our work, with the relevant changes highlighted in green. Below, we address each of your concerns in turn.

W1: Strict positive definiteness and verifier soundness

Thank you for raising this important concern and providing a concrete example. We agree that your example highlights the need for further explanation of our verification procedure. Below, we clarify how positive definiteness is ensured during data construction and how we have further checked the Solver's predictions.

For generated training pairs, strict positive definiteness is ensured by the rules used to assemble the symbolic components produced through adversarial generation, rather than inferred solely from SHGO acceptance. As described in Section 3 and Appendix C.3, the assembled target has the form

VGT(x)=Vproper(x)+∑k(qk(x)−qk(0))2,V_{\mathrm{GT}}(x) = V_{\mathrm{proper}}(x) + \sum_k \bigl(q_k(x)-q_k(0)\bigr)^2,

where VproperV_{\mathrm{proper}} is a positive-definite base function with Vproper(0)=0V_{\mathrm{proper}}(0)=0. The additional squared terms are nonnegative and vanish at the origin. Consequently,

VGT(0)=0,VGT(x)≥Vproper(x)>0for x≠0V_{\mathrm{GT}}(0)=0,\qquad V_{\mathrm{GT}}(x)\geq V_{\mathrm{proper}}(x)>0\quad\text{for }x\neq0

on the admissible domain. The dynamics are also assembled so that

LfVGT(x)=−∑ihi(x)2(∂iVGT(x))2≤0.L_fV_{\mathrm{GT}}(x) = -\sum_i h_i(x)^2\bigl(\partial_i V_{\mathrm{GT}}(x)\bigr)^2\leq0.

During adversarial data evolution, the Generator learns the symbolic components, while these validity-preserving assembly rules remain fixed. Thus, a target such as V(x)=x12V(x)=x_1^2 in the reviewer's two-dimensional example cannot arise from this construction, since it lacks the required positive-definite base. SHGO serves as an additional numerical consistency check for these constructed pairs.

The inference stage requires a separate explanation, because an unconstrained Solver prediction does not automatically inherit this construction guarantee. Our SHGO-based procedure follows the numerical verification approach used in prior symbolic Lyapunov work [1], with a numerical tolerance τ\tau in the acceptance rule. To describe this tolerance precisely, we explicitly write the numerical acceptance condition as m^V≥−τ\widehat m_V\geq-\tau. This formulation distinguishes the implemented numerical criterion from a mathematical proof of positive definiteness.

To examine whether the tolerance affected the reported results, we rechecked all 666 final SHGO-accepted predictions contributing to our reported success rates. In this recheck, we replaced the positivity acceptance threshold m^V≥−τ\widehat m_V\geq-\tau with the stricter requirement m^V>0\widehat m_V>0. All 666 predictions continued to pass. This provides evidence that their acceptance did not depend on allowing zero or slightly negative computed positivity minima. We also performed a dReal check using the same strict-positivity requirement and found no confirmed counterexamples, providing additional support for these recheck results.

We have added further explanations of the construction guarantee, the numerical acceptance criterion, and the stricter checks to the appendix, with the added text highlighted in orange. Thank you again for identifying this issue. Your comment has helped us describe the verification procedure and the scope of its guarantees more precisely. We hope these explanations and additional checks address your concern.

AuthorsOpenReview ↗

Response to Reviewer rHen (2/4)

W2: Real-world applicability and unsafe violations

Thank you for raising this important concern. We agree that unsafe trajectories cannot be tolerated in physical systems, and we would like to provide further clarification building on our response to W1.

During adversarial data generation, the generated symbolic components are assembled into system–Lyapunov pairs using fixed validity-preserving rules. These rules ensure that the target function is strictly positive away from the origin and satisfies the Lyapunov decrease condition on the admissible domain. Thus, the nonzero-state degeneracy identified in W1 is excluded from the constructed training targets. The Solver therefore learns from valid pairs rather than from examples whose validity is relaxed to accommodate violations. As reported in W1, all 666 final accepted outputs also passed the stricter SHGO positivity threshold, and the additional dReal check found no confirmed counterexamples. These results provide no evidence that our reported performance relies on what you describe as “using such violations indefinitely.” Instead, they suggest that pretraining on valid examples can help the model learn to avoid this type of failure, while explicit verification remains part of the procedure.

Our method further supports this assessment by generating Lyapunov functions as explicit symbolic expressions. Domain experts can inspect and simplify these expressions, understand their mathematical structure, and examine them using additional mathematical analysis and application-specific knowledge. The verification of these expressions can also be made more stringent without changing the underlying discovery framework. In our evaluation, all 666 reported successful outputs remained accepted under the stricter positivity check. Thus, the demonstrated effectiveness of these outputs was retained under a stronger acceptance requirement. This flexibility allows stricter checks to be applied to the generated expressions, reducing the risk of accepting an invalid certificate and drawing an incorrect stability conclusion about potentially unstable dynamics. It is this combination of explicit symbolic outputs and adjustable verification requirements that makes the approach useful for further stability assessment.

Thank you again for this concern and the example provided in W1. Your comments have helped us clarify the intended use of our method and strengthen its evaluation. We hope this explanation, together with our response to W1, addresses your concerns.

AuthorsOpenReview ↗

Response to Reviewer rHen (3/4)

W3: Generalization to a wider range of systems

Thank you for raising this important concern. Neural symbolic Lyapunov discovery is still a developing field. Our main benchmarks combine equal numbers of held-out systems from conventional backward construction and solver-aware generation to provide a balanced comparison with existing methods. While this design covers both construction procedures, we agree that stronger evidence of generalization requires evaluation beyond them. We have therefore conducted additional experiments to complement the existing results.

To interpret these results, it is useful to first consider why success rates on ITW are low. As detailed in Appendix F, ITW directly samples dynamics without constructing reference Lyapunov functions and applies only basic filtering. Its success rates are therefore measured over a pool with unknown stability composition, rather than a known-stable test set. The possible presence of unstable systems, together with the difficulty of certificate discovery, can contribute to low success rates across methods. Nevertheless, this setting has practical relevance because a Lyapunov certificate is often sought precisely to establish stability when it is not already known. Such a search can support further system analysis and inform controller redesign when stabilization is needed. To better assess solving capability within this challenging setting, we extensively evaluated Alfarano et al.'s Symbolic Transformer [1] and SEVA-Lyap on a large pool of independently generated ITW polynomial systems, collected 200 systems per method with candidates satisfying our verification criterion, and cross-tested each method on the other method's subset to address your concern about the low aggregate success rates. The results are shown below.

Subset selection (200 systems)Evaluated methodSuccess rate
Solved by Symbolic TransformerSEVA-Lyap91.00%
Solved by SEVA-LyapSymbolic Transformer9.00%

By focusing on systems for which at least one method finds an accepted candidate, this experiment makes the difference in their solving capabilities clearer. SEVA-Lyap solves 91% of the sampled systems solved by Symbolic Transformer, whereas Symbolic Transformer solves only 9% of those sampled from SEVA-Lyap's successful cases. Thus, SEVA-Lyap covers most of the baseline's sampled successes while also solving many instances that the baseline does not solve. We believe these results provide additional evidence of our method's broader solving capability on independently generated systems.

The manuscript appendices complement this comparison by examining generalization from two further perspectives. Appendix G.2 evaluates transfer across different symbolic families without additional training or test-time adaptation on the target family. The results indicate that the Solver learns reusable structural relationships and can apply them beyond the symbolic operator distribution encountered during training. Appendix I then examines applicability to systems drawn from prior Lyapunov studies and established benchmark sources. On these systems, SEVA-Lyap discovers explicit symbolic candidates that pass verification, including instances not solved by Symbolic Transformer under the reported settings. Together, these evaluations extend the evidence beyond ITW by demonstrating both cross-family transfer and applicability to literature systems selected independently of our training-data construction.

We appreciate your suggestion, which has helped us strengthen the evaluation and explain its implications more clearly. We hope that the additional cross-testing results and the further explanation of the appendix experiments address your concern about generalization to a wider range of systems.

Reviewer_uA9JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes SEVA-Lyap, a framework for symbolic Lyapunov function discovery that combines two main ideas: (i) solver-aware data evolution, where a learnable generator produces verified system–Lyapunov pairs that are increasingly difficult for the current solver, and (ii) verification-guided test-time adaptation, where a pretrained symbolic Transformer is locally adapted using verifier-derived rewards and risk-seeking GRPO when direct prediction fails. The paper reports strong improvements over the Symbolic Transformer baseline: direct inference improves from 70.75% to 76.75% on Poly and from 60.50% to 65.50% on NonPoly, while test-time adaptation further improves performance to 92.00% and 74.50%, respectively.

Critical strengths
  • The idea of adapting the training-data distribution to the current solver’s weaknesses is well-motivated and useful for Lyapunov function discovery.
  • The paper presents a fairly complete pipeline combining data generation, symbolic prediction, verification, and test-time adaptation.
  • The authors also provide additional verification for the identified Lyapunov functions using dReal and SOS rather than relying only on the main SHGO-based verification.
Critical weaknesses
  1. The novelty of the test-time adaptation component relative to Zou et al. (2025) should be clarified much more carefully. The RL-based approach of Zou et al. (2025) already performs symbolic Lyapunov search using numerical Lyapunov condition verification, verifier/counterexample feedback, and risk-seeking RL. SEVA-Lyap's test-time procedure similarly samples symbolic candidates, evaluates them using verifier-derived rewards, and locally optimizes the symbolic generator/solver using a risk-seeking RL objective. The paper describes the main distinction as reusing a pretrained cross-system solver instead of optimizing from scratch, which is meaningful, but the underlying test-time search mechanism appears conceptually quite close. Specifically, I would like the authors to provide a more explicit comparison between the two algorithms:

pretrained symbolic Transformer model + risk-seeking RL method used in Zou et al. (2025)

v.s.

pretrained symbolic Transformer model + the proposed risk-seeking GRPO

This would make it much clearer whether the improvement comes from the proposed test-time algorithm itself or primarily from having learned a strong pre-trained model solver before inference.

  1. Like the first point, the large performance gap in Table 1 is difficult to interpret because the methods start from very different levels of prior learning. SEVA-Lyap is trained on up to 500K system–Lyapunov pairs and already achieves 76.75% / 65.50% before test-time adaptation, whereas the prior ICML paper Zou et al. (2025) searches from scratch for each system. Thus, 92.00% vs. 30.50% should not be viewed simply as a comparison of RL search quality. More controlled baselines are needed to separate the effects of cross-system pretraining, adaptive data generation, and test-time RL.

  2. Generalization beyond the synthetic data-generation pipeline remains unclear. The main Poly/NonPoly test sets are still closely tied to the construction families used during training. The ITW experiment is more informative, but success drops to only 18.0% and 21.5%, and the benchmark may contain unstable systems. A stronger evaluation on independently generated, known-stable systems or established control benchmarks would better support the generalization claim.

  3. The benefit of solver-aware data evolution could be isolated more clearly. On Poly, solver-aware training improves 70.75% to 76.75%, while test-time adaptation provides a much larger gain to 92%. The paper shows that evolved systems become harder, but harder examples are not necessarily more useful for generalization. Comparisons against simpler alternatives—such as more backward-construction data, hard-example mining from a fixed pool, or curriculum selection based directly on solver failures—would clarify whether the proposed generator–solver co-evolution is truly necessary.

Questions for authors

The following questions directly summarized the main concerns I raised above:

  1. Novelty of test-time adaptation: Could the authors clarify the key difference from Zou et al. (2025), and ideally compare pre-trained symbolic Transformer + Zou et al.’s risk-seeking RL with pretrained symbolic Transformer + your risk-seeking GRPO?
  2. Interpretation of Table 1 and fair baselines: Could you provide a more controlled comparison baselines that separates the effects of cross-system pretraining, solver-aware data evolution, and test-time RL?
  3. Generalization: Could you evaluate SEVA-Lyap on independently generated, known-stable systems or broader established control benchmarks to better assess out-of-distribution performance?
  4. Necessity of solver-aware data evolution: Could you compare against simpler alternatives such as more backward-construction data, hard-example mining from a fixed pool, or curriculum selection based directly on solver failures?
5 comments
AuthorsOpenReview ↗

Response to Reviewer uA9J (1/4)

Thank you for your careful review and constructive comments, which have helped us clarify our contributions and strengthen the experimental evaluation. We have added experiments and revised the main text, with the revisions marked in different colors. Below, we address each weakness together with its corresponding question.

W1 & Q1: Novelty of test-time adaptation

Thank you for this helpful suggestion. We agree that the novelty of our test-time fine-tuning scheme and its advantage over Zou et al.'s RL-based Approach [1] should be explained more clearly.

Drawing on the successful full training workflow of large language models, we introduce the first complete pipeline from cross-system pretraining to test-time RL fine-tuning for symbolic Lyapunov function discovery. Within this pipeline, our GRPO-based scheme is tailored to further develop the capabilities of a pretrained solver, whereas RL-based Approach was designed to search from scratch for each system [1]. Our fine-tuning design therefore aims to improve performance on the target system while retaining useful knowledge acquired during pretraining. To support this goal, KL regularization penalizes excessive departure from the reference policy, helping preserve pretrained capabilities. Meanwhile, group-relative advantages guide learning by comparing candidate functions for the same system, and clipping discourages overly large policy updates. By combining these mechanisms with feedback on the Lyapunov conditions, our scheme supports stable, targeted refinement that builds on the pretrained solver's existing capabilities.

To assess whether this design improves fine-tuning performance, we have added an experiment that applies the RL update from RL-based Approach to SEVA-Direct. This baseline and SEVA-Full use the same pretrained checkpoint, the same number of candidates per round, and the same maximum number of fine-tuning rounds. This allows us to compare the two fine-tuning procedures from an identical starting point under matched adaptation budgets. The results are shown below.

MethodSuccess rate
SEVA-Direct + RL-based Approach87.25%
SEVA-Full92.00%

Our GRPO-based fine-tuning scheme achieves a success rate of 92.00%, exceeding the added baseline by 4.75 percentage points. Since both procedures start from the same pretrained model and use matched adaptation budgets, this improvement cannot be attributed to a stronger pretrained initialization. Instead, it provides direct evidence that our fine-tuning scheme makes more effective use of the same pretrained model in this evaluation.

We appreciate your suggestion, which helped us clarify this distinction through a controlled comparison. We hope the explanation and additional results make the novelty and effectiveness of our fine-tuning scheme clearer.

AuthorsOpenReview ↗

Response to Reviewer uA9J (2/4)

W2 & Q2: Interpretation of Table 1 and prior learning

Thank you for this helpful suggestion. We agree that interpreting Table 1 requires clarifying the methods' different training strategies and the contributions of individual components.

We would first like to clarify that our work is the first to introduce adversarial training-data generation into Lyapunov function discovery. This design is motivated by the decisive role of training-data quality in pretrained neural model performance. The Generator learns to produce harder valid examples, while the Solver learns to solve them. As the Solver improves, the Generator adapts, keeping examples challenging and directing training toward the Solver's remaining weaknesses. As explained in our response to W1&Q1, we also draw on the training workflow of large language models. We introduce the first complete pipeline from cross-system pretraining to test-time RL fine-tuning for symbolic Lyapunov function discovery, addressing the gap between pretraining and fine-tuning in existing approaches. Our adaptation of GRPO is designed to refine the pretrained model. The comparisons below show that it is more effective for this purpose than applying the RL update originally designed for search from scratch.

Reusing a pretrained solver also avoids repeating the entire training process for every system. Zou et al.'s RL-based Approach [1] searches from scratch for each system, so its training cost accumulates as more systems are evaluated. Across the test systems in our evaluation, its cumulative training time far exceeds our one-time pretraining time. This highlights the practical value of cross-system reuse and motivates controlled comparisons of data evolution and fine-tuning.

Building on the experiment in our response to W1&Q1, we have added two configurations based on Alfarano et al.'s Symbolic Transformer [2]: Symbolic Transformer + SEVA-Finetuning and Symbolic Transformer + RL-based Approach, alongside SEVA-Direct + RL-based Approach. Together with the original configurations, these experiments compare pretrained models under the same fine-tuning procedure and fine-tuning procedures from the same pretrained model. Both models use 500,000 pretraining examples with the same model scale and training settings. All fine-tuning variants use the same number of candidates per round and the same maximum number of rounds.

MethodSuccess rate
Symbolic Transformer70.75%
SEVA-Direct76.75%
Symbolic Transformer + SEVA-Finetuning84.00%
Symbolic Transformer + RL-based Approach79.75%
SEVA-Direct + RL-based Approach87.25%
SEVA-Full92.00%

First, comparing the pretrained models shows a consistent benefit from solver-aware data evolution. SEVA-Direct outperforms Symbolic Transformer without fine-tuning (76.75% versus 70.75%). The advantage remains when both models use our fine-tuning scheme (92.00% versus 84.00%) or the update from RL-based Approach (87.25% versus 79.75%). Thus, the benefit of data evolution is present before fine-tuning and persists under either fine-tuning procedure. These gains under matched pretraining settings support the value of adversarial, feedback-driven evolution that makes the training distribution responsive to the Solver's changing capabilities, improving its ability to discover Lyapunov functions for difficult systems.

Second, fixing the pretrained model reveals the additional contribution of test-time fine-tuning. Our scheme raises the success rate of Symbolic Transformer from 70.75% to 84.00% and that of SEVA-Direct from 76.75% to 92.00%. To assess the importance of the fine-tuning procedure itself, we compare these results with those obtained using RL-based Approach. Our scheme performs better with both pretrained models, by 4.25 and 4.75 percentage points, respectively, under the same candidate and round budgets. These gains support the value of designing the update procedure specifically for pretrained-model refinement. Our GRPO-based scheme uses Lyapunov-condition feedback to improve the solver while helping retain its pretrained capabilities. KL regularization discourages excessive departure from the reference policy, while group-relative advantages and clipping guide controlled updates. The improvements with both pretrained models show that this fine-tuning design contributes beyond the stronger initialization provided by data evolution.

We appreciate your suggestion, which helped us clarify both the novelty of our framework and the interpretation of Table 1. We hope these controlled comparisons make the respective contributions of data evolution and test-time fine-tuning clearer.

AuthorsOpenReview ↗

Response to Reviewer uA9J (3/4)

W3 & Q3: Generalization beyond the synthetic pipeline

Thank you for this constructive suggestion. Using neural networks to discover Lyapunov functions as explicit symbolic expressions remains an emerging research direction. Our main benchmarks are designed to support comparison with existing methods. We also agree that stronger evaluation on independently generated, known-stable systems or established control benchmarks would further support our claims. We have therefore added experimental evidence to address this concern.

In response to your request for evaluation on independently generated systems with evidence of stability, we conducted a supplementary cross-testing experiment. We extensively evaluated Alfarano et al.'s Symbolic Transformer [2] and SEVA-Lyap on a large pool of independently generated ITW polynomial systems. For each method, we collected 200 systems for which it found a symbolic Lyapunov candidate satisfying our verification criterion. Each method was then evaluated on the subset selected by the other. The results are summarized below.

Subset selection (200 systems)Evaluated methodSuccess rate
Solved by Symbolic TransformerSEVA-Lyap91.00%
Solved by SEVA-LyapSymbolic Transformer9.00%

On these subsets with accepted candidates, SEVA-Lyap solves 91% of the systems sampled from Symbolic Transformer's successes, while Symbolic Transformer solves only 9% of those sampled from SEVA-Lyap's successes. This shows that SEVA-Lyap covers most of the baseline's sampled successes while also solving many instances that the baseline does not solve. We believe these results provide additional evidence of its broader solving capability on independently generated systems.

The manuscript appendices complement these findings by examining generalization from two additional angles. Appendix G.2 tests transfer between symbolic families without further training or test-time adaptation on the target family. The results indicate that the Solver acquires reusable structural relationships that can be applied beyond the symbolic operator distribution encountered during training. Appendix I examines systems drawn from prior Lyapunov studies and established benchmark sources. SEVA-Lyap discovers explicit symbolic candidates that pass verification on these systems, including instances not solved by Symbolic Transformer under the reported settings. These evaluations demonstrate cross-family transfer and applicability to literature systems selected independently of our training-data construction, extending the experimental evidence beyond ITW.

Thank you again for your suggestion. It has helped us strengthen the evaluation and explain the implications of the results more clearly. We hope that the additional cross-testing experiment and the further explanation of the appendix evaluations address your concern about generalization to a wider range of systems.

Reviewer_Kc3fOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SEVA-Lyap, a neural-symbolic framework for discovering symbolic Lyapunov functions for dynamical systems. The framework introduces solver-aware data evolution, where a learnable Generator produces verified system–Lyapunov pairs based on feedback from a reusable Solver, progressively improving its ability to discover Lyapunov functions. It further incorporates verification-guided test-time adaptation using risk-seeking GRPO to refine the Solver on challenging unseen systems. Experimental results on polynomial and non-polynomial benchmarks demonstrate improved verification success rates over existing baselines.

Critical strengths

-The paper introduces two interesting ideas, solver-aware data evolution and verification-guided test-time adaptation, and integrates them into a unified framework for symbolic Lyapunov function discovery. The overall approach is promising, and the paper includes comparisons with several recent and relevant studies.

-The work also addresses two practical limitations of existing learning-based approaches, namely their reliance on static training datasets and their limited ability to adapt to challenging unseen systems. The proposed framework provides a potentially useful way to improve both the training process and instance-specific solving capability.

-The experimental results are encouraging, with improved verification success rates on both polynomial and non-polynomial benchmarks compared with the evaluated baselines. The additional ablation studies and generalization experiments also help demonstrate the potential effectiveness of the proposed approach.

Critical weaknesses

-Although this work may offer some useful contributions to the field, the current manuscript suffers from substantial presentation issues. The overall organization and logical flow are not sufficiently clear, making the paper difficult to follow. In addition, several important concepts and technical terms are introduced without adequate explanation, which makes the paper particularly challenging for readers who are not already familiar with this research area.

-Several statements throughout the paper are vague or insufficiently explained. For example, the abstract uses expressions such as "around the weaknesses of the current solver" and "through targeted risk-seeking adaptation," but it is unclear what specific weaknesses are being addressed or what "targeted" refers to. Similarly, the claim of "demonstrating notable transferability" does not clearly specify what kind of transferability is being evaluated. Such ambiguities also appear in the experimental section. For instance, the authors refer to "representative challenging systems reported in established Lyapunov studies" without explaining what makes these systems challenging or how their difficulty is characterized. These descriptions should be made more precise.

-The background and motivation substantially overlap with those of Alfarano et al. (2024), while several key components of the proposed framework build upon existing studies, particularly Alfarano et al. (2024) and Zou et al. (2025). However, the manuscript does not sufficiently distinguish its contributions from these works. The authors should place greater emphasis on explaining what is genuinely new, how the proposed mechanisms differ from existing approaches, and why these differences are technically meaningful.

-The Introduction lacks a coherent narrative connecting neural and symbolic Lyapunov functions. For example, the first paragraph discusses the limited automation capabilities of traditional symbolic approaches, while the second paragraph introduces the flexibility of neural certificates before returning to the limitations of symbolic approaches relative to neural methods. This back-and-forth discussion makes the motivation difficult to follow. I suggest reorganizing these paragraphs to establish a clearer progression from traditional Lyapunov synthesis to neural approaches and finally to symbolic Lyapunov discovery. The discussion of symbolic Lyapunov functions in the first paragraph could also be moved or integrated into the second paragraph.

-Figure 1 does not clearly communicate the overall pipeline of the proposed framework. As the main framework illustration, it should provide an intuitive overview of the key components and their interactions. However, the current figure contains too much text and is visually complicated, making it difficult to understand the workflow at a glance. A more concise and structured diagram would substantially improve the presentation.

Minor Comments:

i. The Introduction repeats much of the abstract, particularly the discussion of the experimental improvements. Detailed numerical results would be more appropriate in the experimental section. Instead, the Introduction should provide a concise and explicit summary of the paper's main contributions, which are currently not sufficiently highlighted.

ii. In Line 147, the verification region is fixed with (R=5) for all experiments. What motivates this choice? Is there any theoretical or empirical justification, and how sensitive are the results to different values of (R)?

iii. Section 3.2 mainly presents the backward construction procedure adopted from prior work. Since much of this material is not original to the proposed method, I suggest incorporating it into the preliminaries or background rather than presenting it as a separate subsection that may appear to describe a methodological contribution.

iv. The abbreviation GRPO in Figure 1 should be explicitly defined when it is first introduced.

v. Sections 5 and 6, covering the experimental setup and results, could be combined to improve readability and avoid unnecessary fragmentation.

vi. The distinctions between the baseline methods and the proposed approach are not sufficiently explained. For example, around Line 345, LyZNet is described as learning Lyapunov functions from sampled states and verifying them using SMT procedures, whereas FOSSIL 2.0 is categorized as iteratively refining candidates through falsification or verifier feedback. However, FOSSIL 2.0 also employs learning and formal verification within an iterative framework. The authors should clarify the methodological differences rather than presenting these approaches as if their characteristics were mutually exclusive.

vii. The experimental evaluation is not sufficiently comprehensive. In particular, more detailed comparisons are needed to isolate the contributions of solver-aware data evolution and test-time adaptation. A baseline combining the original Symbolic Transformer with the same test-time adaptation procedure would be especially valuable. Additional comparisons under matched computational budgets would also strengthen the experimental conclusions.

Questions for authors
  1. In the Introduction, the authors suggest that neural Lyapunov functions introduce additional verification complexity. However, neural networks are primarily used to efficiently discover promising candidates that approximately satisfy the Lyapunov conditions on sampled states, while the final verification still requires checking the original mathematical constraints. Therefore, using neural networks for candidate generation does not necessarily increase verification complexity. Moreover, existing studies, such as FOSSIL, have demonstrated the computational benefits of learning-based certificate synthesis compared with traditional approaches such as SOS-based methods. Could the authors clarify the motivation and evidence behind their claim that neural Lyapunov functions are more difficult to verify?

  2. The paper introduces SHGO-based verification in Section 3.1. From its mathematical formulation, this approach appears similar to worst-case optimization, where the verifier searches for states that maximally violate the Lyapunov conditions. Could the authors clarify the differences between SHGO-based verification and conventional worst-case optimization methods? In particular, what are the advantages or distinctive features of SHGO in this setting?

  3. What are the fundamental advantages of symbolic Lyapunov functions over neural Lyapunov functions in the context of this work? The current manuscript does not clearly establish why symbolic representations should be preferred, particularly from the perspective of verification efficiency. Since symbolic expressions are not necessarily easier to verify than neural representations, could the authors briefly explain the practical or theoretical benefits of symbolic Lyapunov discovery and provide evidence supporting these advantages?

10 comments
AuthorsOpenReview ↗

Response to Reviewer Kc3f (1/9)

Thank you for your careful review and detailed comments on both the writing and the methodology of our paper. We acknowledge that some parts of the manuscript were not explained clearly enough. This made it difficult to understand the motivation behind our method and its novel contributions. Below, we respond to each weakness and question you raised. We have also revised the corresponding parts of the manuscript to make our explanations clearer. We hope these responses and revisions address your concerns and provide a clearer basis for assessing our work.

W1: Presentation and conceptual explanations

Our work applies artificial intelligence to Lyapunov function discovery in control theory. This requires us to consider how familiar researchers in both fields are with the relevant concepts, as well as their reading habits. We therefore greatly appreciate your feedback on the presentation. We recognize that our original explanations did not adequately account for these differences in background and reading expectations.

To address these issues, we have revised the manuscript. We have reorganized the Introduction and Related Work to make the motivation and the connections between approaches easier to follow. We have also added plain-language explanations of key concepts and technical terms. The revised passages are highlighted in color for ease of reference. Details of these changes are provided in our responses to your specific comments below.

We would also like to clarify the main contributions of our method. Symbolic dynamical systems and Lyapunov functions can be represented as token sequences, making them well suited to Transformer-based learning. The performance of these models depends strongly on the amount and quality of training data. To address this need, we are the first to introduce adversarial data construction for symbolic equations in Lyapunov function discovery. The Generator learns to produce harder valid examples, while the Solver learns to solve them. This competition keeps the training data challenging for the Solver as its capabilities improve. It substantially strengthens the Solver's ability to discover Lyapunov functions for difficult systems. In addition, we draw on the pretraining-and-fine-tuning paradigm used in large language models. We adapt GRPO, an established algorithm for fine-tuning these models, to use a risk-seeking objective guided by verification feedback. Following large-scale training, we use this adapted GRPO algorithm to fine-tune the Solver's parameters for individual test systems that remain unresolved. This further improves its ability to solve these specific systems. Our work is the first to bring this complete training-to-fine-tuning paradigm to Lyapunov function discovery. We explain the differences from prior work and the supporting evidence in the responses below.

AuthorsOpenReview ↗

Response to Reviewer Kc3f (2/9)

W2: Precision of claims and terminology

Thank you for pointing out these unclear expressions. We agree that the original wording did not explain the intended meaning precisely enough. We clarify the four examples below.

(1) “Around the weaknesses of the current solver.” This refers to systems for which the current Solver produces candidates with relatively large empirical violations of the Lyapunov conditions. We quantify this feedback using the difference between the candidate's empirical violation score and that of the verified reference, evaluated on sampled states. The resulting feedback guides the Generator toward systems on which the current Solver needs improvement.

(2) “Through targeted risk-seeking adaptation.” Here, “targeted” means that fine-tuning is performed separately for each test system that remains unresolved after direct inference. During this process, the risk-seeking objective uses rewards derived from verification feedback to focus updates on candidates in the high-reward tail. Before adapting to the next system, the Solver is reset to the shared trained model.

(3) “Demonstrating notable transferability.” The intended meaning is generalization to independently constructed systems, as evaluated on the ITW benchmark in Section 5.5. ITW systems are generated by directly sampling their dynamics, without first constructing a reference Lyapunov function. This evaluation therefore examines whether SEVA-Lyap can solve systems generated independently of the training data-construction procedures.

(4) “Representative challenging systems reported in established Lyapunov studies.” The systems evaluated in Appendix I are benchmark systems and published examples from prior Lyapunov studies. They supplement our generated benchmarks with cases drawn from published sources. We agree that their origin and evaluation purpose should be stated more clearly, and have revised the wording accordingly.

These clarifications are reflected in the revised manuscript. In the abstract, we have replaced “around the weaknesses of the current solver” with “using feedback on the current Solver's empirical violations of the Lyapunov conditions.” The phrase “through targeted risk-seeking adaptation” now reads “through risk-seeking fine-tuning on unresolved systems,” and “notable transferability” has been revised to “generalization to independently constructed systems.” To support this clarification, Section 5.1 explains the ITW construction, and Section 5.5 reports the corresponding generalization results. In the experimental section, “representative challenging systems reported in established Lyapunov studies” is now described as “benchmark systems and published examples from prior Lyapunov studies.”

We also reviewed similar expressions elsewhere in the manuscript. For example, the heading “Transfer and challenging systems” in Section 5.1 has been changed to “Additional evaluation settings,” and “Representative Challenging-System Evaluation” in Appendix I to “Evaluation on Systems from Prior Lyapunov Studies.” Related descriptions in the Introduction, method, results, and appendices have been revised for consistency. Please refer to the green text in the revised manuscript for these changes.

AuthorsOpenReview ↗

Response to Reviewer Kc3f (3/9)

W3: Novelty relative to prior work

Thank you for this important suggestion. We agree that the original manuscript did not clearly highlight our main contributions. The explanations were scattered across several sections, making the differences from prior work difficult to identify. We apologize for this lack of clarity.

As discussed in our response to W1, our motivation follows from two limitations of existing approaches. Alfarano et al. (2024) [1] train a reusable symbolic Transformer on synthetic data. For neural models pretrained in this way, training-data quality plays a decisive role in model performance. However, their fixed generation process cannot respond to the Solver's learning progress. Zou et al. (2025) [2] use risk-seeking reinforcement learning to search for symbolic Lyapunov functions from scratch for each system. Repeatedly exploring the symbolic search space without reusing knowledge learned across systems can incur substantial search costs.

To address these limitations, we are the first to introduce the principle of adversarial learning into the construction of system–Lyapunov training pairs for a reusable symbolic Solver. The Generator learns to construct valid systems that challenge the Solver, while the Solver learns to overcome these challenges. This adversarial interaction keeps the training data challenging as the Solver improves and strengthens its ability to discover Lyapunov functions for difficult systems. Inspired by the success of large language models, we also introduce the first training pipeline in this setting that connects cross-system pretraining with verification-guided, risk-seeking fine-tuning on individual test systems. An adapted GRPO algorithm further develops the pretrained Solver's ability to solve unresolved cases.

Within the data-evolution stage, backward construction from Alfarano et al. (2024) [1] provides an effective foundation that we explicitly acknowledge. Our contribution makes the symbolic construction choices learnable and guides them through adversarial feedback. Verification checks whether generated pairs meet the acceptance conditions. The empirical violation gap between the Solver's candidate and the reference function then measures how challenging a system is for the current Solver. Using this feedback to update the Generator allows the training distribution to evolve with the model.

At test time, the optimization objective is to find at least one candidate that passes verification. Accordingly, our adaptation of GRPO uses verifier-derived rewards and focuses policy updates on candidates in the high-reward tail. Through update clipping and a KL penalty, the fine-tuning procedure discourages overly aggressive changes and encourages the model to remain close to its pretrained behavior. This helps the Solver explore new solutions while preserving what it has already learned. Comparisons within each sampled group provide the update signal without requiring a separate value network.

The additional controlled comparisons reported in our response to Minor Comment vii further examine these mechanisms. Under matched pretraining settings and candidate and round budgets for fine-tuning, these comparisons help distinguish the contributions of solver-aware data evolution and test-time adaptation.

These contributions should have been presented together and explained more directly. Accordingly, we have revised the Introduction to bring our motivation and contributions into clearer focus. These changes are highlighted in purple. We hope this clearer presentation helps explain the reasons behind our design and makes its contributions beyond prior work easier to assess.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.