acceptodds
Under review as a conference paper at ICLR 2027

Peppy: An AI-Assisted Workflow for Tight Convergence Analysis of Optimization Algorithms

Abstract

This paper presents Peppy, an AI-assisted workflow for discovering tight, analytic convergence proofs for first-order optimization algorithms. Generic approaches to using LLMs to conduct mathematical research target an unspecified, broad spectrum of problems and sometimes use the Lean 4 proof assistant for formalization. On the other hand, Peppy leverages domain-specific knowledge more heavily and is thereby capable of constructing the proofs in a more structured manner, which allow a minimal and accessible verification through SymPy. We experimentally demonstrate through examples that Peppy provides a rigorous, practical, and reproducible paradigm for AI-assisted theorem synthesis in optimization. We further highlight its capability of closing several open problems on tight convergence analysis of first-order optimization algorithms, including conjectures for Nesterov’s FGM. Overall, Peppy is designed to turn the art of optimization algorithm analysis into a science.

Primary area:
optimization
Keywords:
Optimization, AI-Assisted Workflow, Agent, Proof, Convergence Analysis

Reviews

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

This paper presents Peppy, an AI-assisted workflow for discovering tight analytic convergence proofs for first-order optimization algorithms. Peppy combines numerical PEP certificates, the PEP-to-Lyapunov framework, specialized structural-recovery utilities, and a five-stage agent workflow that progresses from PEP implementation. The paper evaluates Peppy on several known optimization results and reports that the generated proofs pass mathematical verification. More importantly, the framework is used to derive new tight convergence results for several previously unresolved cases, including conjectures for FGM, OGM, rational FGM, and FISTA. The paper contributes both a domain-specific workflow for AI-assisted proof discovery and a set of substantive new convergence analyses obtained through this workflow.

Critical strengths
  1. The paper demonstrates substantive research-level mathematical contributions. Peppy is used to derive tight analytic convergence results for several previously unresolved settings. In particular, the FGM results provide explicit finite-horizon Lyapunov analyses and are shown to match the corresponding lower bounds. The rational FGM and FISTA results are similarly shown to be tight. This goes beyond recovering standard convergence proofs, as Peppy is also applied to problems with apparently novel analytic results.

  2. The workflow is well structured and inspectable. Peppy decomposes proof discovery into explicit stages, including PEP implementation, certificate extraction, Lyapunov construction, structural recovery, and final symbolic completion. The paper clearly distinguishes numerical evidence from a valid analytic proof and is transparent about which parts of the workflow remain heuristic. In particular, the Stage-4 utilities operate on fixed numerical instances, while the authors explicitly acknowledge their dependence on numerical tolerances and on the candidate vector family supplied to the basis search.

Critical weaknesses
  1. The evaluation does not clearly isolate the benefit of the proposed workflow. The full Peppy pipeline is shown to work well, but it is unclear how much of the gain comes from the five-stage workflow itself versus the strong model, PEP machinery, and domain-specific tools.
  2. The contribution of Stage 4 is not directly validated. The structural-recovery step is presented as a key novel component. However, the paper does not show how much it improves proof discovery compared with using the same PEP certificate and Gram information without the specialized recovery utilities.
Questions for authors
  1. Could the authors compare full Peppy with the same model, tools, mathematical information, and computational budget, but using only a high-level prompt for proof generation? This would help isolate the value of the staged workflow.

  2. Could the authors provide an ablation of the Stage-4 structural-recovery mechanism? For example, how does proof success or efficiency change when the agent receives the same PEP certificate and numerical structural information but without the dedicated recovery procedure?

4 comments

We are happy to hear that the reviewer recognized our "substantive research-level mathematical contributions" and found our workflow "well structured and inspectable." We particularly appreciate the reviewer's attention to the "PEP-to-Lyapunov framework" and the reviewer's observation that our paper "clearly distinguishes numerical evidence from a valid analytic proof." We thank the reviewer for their detailed and constructive feedback and address their questions below.

W1, Q1

W1. The evaluation does not clearly isolate the benefit of the proposed workflow. The full Peppy pipeline is shown to work well, but it is unclear how much of the gain comes from the five-stage workflow itself versus the strong model, PEP machinery, and domain-specific tools.

Q1. Could the authors compare full Peppy with the same model, tools, mathematical information, and computational budget, but using only a high-level prompt for proof generation? This would help isolate the value of the staged workflow.

As discussed in our common response, Peppy contributes both the specification of the desired Lyapunov proof structure and concrete verifiable intermediate steps for obtaining it. The staged workflow also enables researchers to inspect and interact with the discovery process through Jupyter notebooks. Since supporting this interactive environment also requires computational resources, we did not aim to match the total computational budget, focusing instead on the resulting proof structures.

To examine whether a high-level prompt alone could produce the desired Lyapunov representations, we conducted one experiment for each of the five cases in Section 5 using the same Astra Ultra model. We provided the problem statements and target convergence bounds, requesting Lyapunov analyses without Peppy's instructions or numerical PEP inputs. All five runs proposed Lyapunov analyses, but four failed to obtain the desired simple, summation-free form, while FISTA-rational yielded such a representation. The following table presents representative terms from the Lyapunov functions proposed by the base model.

CaseRepresentative terms in the base model's outputSimple, summation-free form presented?
FGM, Conjecture 4∥r−D∑i=0kτiqi∥2,∑i,j=0kEij⟨qi,qj⟩\displaystyle \left\lVert r-D\sum_{i=0}^k\tau_i q_i\right\rVert^2,\quad \sum_{i,j=0}^k E_{ij}\langle q_i,q_j\rangleNo
FGM, Conjecture 5∥r−a∑i=0kτiqi∥2,∑i,j=0kSij⟨qi,qj⟩\displaystyle \left\lVert r-a\sum_{i=0}^k\tau_i q_i\right\rVert^2,\quad \sum_{i,j=0}^k S_{ij}\langle q_i,q_j\rangleNo
OGM, Conjecture 4∑i=0k−1λiCi+∑i=0kτiSi\displaystyle \sum_{i=0}^{k-1}\lambda_i\mathcal C_i+\sum_{i=0}^k\tau_i\mathcal S_iNo
FGM-rational∑i=0k−1λiCi+∑i=0kμiSi\displaystyle \sum_{i=0}^{k-1}\lambda_i\mathcal C_i+\sum_{i=0}^k\mu_i\mathcal S_iNo
FISTA-rationalBiδi,∥ui∥2,⟨ui,ui−1⟩,∥ui−1∥2\displaystyle B_i\delta_i,\quad \lVert u_i\rVert^2,\quad \langle u_i,u_{i-1}\rangle,\quad \lVert u_{i-1}\rVert^2Yes

The full expressions and notation are provided in lyapunov_functions.md within conjectures_high_level_prompt_baseline.zip, along with the original outputs. These results highlight the value of Peppy's structured guidance in obtaining the desired simple, summation-free Lyapunov representations.

W2, Q2

W2. The contribution of Stage 4 is not directly validated. The structural-recovery step is presented as a key novel component. However, the paper does not show how much it improves proof discovery compared with using the same PEP certificate and Gram information without the specialized recovery utilities.

Q2. Could the authors provide an ablation of the Stage-4 structural-recovery mechanism? For example, how does proof success or efficiency change when the agent receives the same PEP certificate and numerical structural information but without the dedicated recovery procedure?

We would first like to clarify that, while the outcomes of individual blocks can be verified separately, the role of Block 4 in the overall proof discovery process is not fully separable from Blocks 2 and 3. As described in Appendix D.2.2, failed basis or reconstruction checks can lead back to Block 3 for alternative groupings and, if needed, Block 2 for different certificates.

In this context, we provided the base model with PEP certificates from successful Peppy runs, but not the resulting partial sums or Peppy's instructions. We conducted one fresh trial for each of the five cases in Section 5 using GPT-6 Astra with Ultra reasoning and access to generic PEPFlow tools. The following table presents representative terms from the Lyapunov functions proposed by the base model.

CaseRepresentative terms in the base model's outputSimple, summation-free form presented?
FGM, Conjecture 4∑i=0kwiq⋆,i,∑i=0k−1Aiqi,i+1\displaystyle \sum_{i=0}^{k}w_iq_{\star,i},\quad \sum_{i=0}^{k-1}A_iq_{i,i+1}No
FGM, Conjecture 5−∑i=0kdi⟨gi,xi−x⋆⟩,12L∑i=0k−1Ai∥gi−gi+1∥2\displaystyle -\sum_{i=0}^{k}d_i\langle g_i,x_i-x_\star\rangle,\quad \frac{1}{2L}\sum_{i=0}^{k-1}A_i\lVert g_i-g_{i+1}\rVert^2No
OGM, Conjecture 4∑i=0kτiEi,∑i=0k−1λiRi\displaystyle \sum_{i=0}^{k}\tau_iE_i,\quad \sum_{i=0}^{k-1}\lambda_iR_iNo
FGM-rational−2t∑i=0k−1αi⟨ui,wi−x⋆⟩,t∑i=0k−1Ai∥ui−ui+1∥2\displaystyle -2t\sum_{i=0}^{k-1}\alpha_i\langle u_i,w_i-x_\star\rangle,\quad t\sum_{i=0}^{k-1}A_i\lVert u_i-u_{i+1}\rVert^2No
FISTA-rationalqk(F(yk)−F(x⋆)),∥zk−zk−1∥2,∥zk−1∥2,∥zk∥2\displaystyle q_k\bigl(F(y_k)-F(x_\star)\bigr),\quad \lVert z_k-z_{k-1}\rVert^2,\quad \lVert z_{k-1}\rVert^2,\quad \lVert z_k\rVert^2Yes

The full expressions and notation are provided in the case-specific submission/lyapunov.md files under trials/ in conjectures_b3_baseline.zip, along with the original outputs. These results provide concrete evidence of the value of Peppy's structured guidance in recovering the desired simple, summation-free Lyapunov representations from numerical PEP certificates.

For reference, we also include an experiment that additionally provides the partial sums from a successful Peppy run, but without the dedicated recovery procedure. Interestingly, the base model succeeded in numerically reconstructing the supplied partial sums in all five cases, suggesting that the mathematical structure obtained through Peppy's iterative workflow may play an important role in structural recovery. The corresponding experimental records are provided in conjectures_b4_baseline.zip.


We thank the reviewer again for their thoughtful and constructive feedback, which has helped us clarify the contributions of Peppy and provide additional experimental evidence. We hope that our explanations and new experiments address the reviewer's concerns regarding the value of the staged workflow and its structural-recovery mechanism. We would be happy to discuss any remaining questions.

Thank you for your detailed reply. I have no further questions. I will keep my original score.

Thank you again for your positive review and constructive suggestions. We appreciate your feedback.

Reviewer_D4txOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors present Peppy, an AI-assisted workflow that leverages the well-established Performance Estimation Problem (PEP) framework to derive analytic convergence proofs for first-order optimization algorithms. The authors demonstrate their approach by (i) recovering known convergence bounds for several optimization algorithms and (ii) obtaining novel mathematical results, including analytic proofs of tight convergence bounds that resolve several long-standing conjectures in the optimization literature.

Critical strengths

Novel mathematical results

The paper presents potentially significant mathematical contributions, including the first analytic proofs of tight convergence bounds for Nesterov's Fast Gradient Method (FGM), resolving conjectures that had remained open since 2017. The authors also establish tight bounds for OGM, FGM with rational coefficients, and FISTA with rational coefficients. These results represent meaningful advances in the theoretical understanding of classical optimization algorithms and demonstrate the potential of AI-assisted mathematical research.

Critical weaknesses

Limited AI novelty and unclear value of the proposed pipeline

The paper presents potentially significant mathematical results. However, Peppy is positioned as the primary contribution, and its significance for the AI community remains questionable. The proposed pipeline largely combines existing mathematical tools and established agentic techniques, while its claimed advantages are discussed primarily qualitatively, without empirical evidence demonstrating their incremental value.

AI-assisted mathematical discovery is already well established, with systems such as AlphaEvolve demonstrating advances across a broad range of mathematical problems [1]. More recently, a general-purpose OpenAI reasoning model autonomously disproved a long-standing Erdős conjecture concerning the planar unit-distance problem, without specialized mathematical proof-search scaffolding [2].

At the same time, recent empirical studies have begun to question the added value of sophisticated agentic pipelines. Gideoni et al. [3] show that simple sampling baselines can match or outperform more elaborate code-evolution methods. Wang et al. [4] demonstrate that harness evolution does not consistently outperform simpler test-time scaling approaches under comparable budgets. Brilliantov et al. [5] find that a minimal, single-session coding agent can match sophisticated autonomous ML-engineering harnesses under controlled comparisons. Further evidence comes from studies showing that harness choice mostly affects inference costs rather the task success rates [6, 7]. Although much of this emerging literature has not yet undergone peer review, these findings provide meaningful evidence that the value of a proposed agentic harness should not be assumed, but rather demonstrated through controlled comparisons with simple yet surprisingly powerful baselines.

In light of these developments, it is particularly important to establish what Peppy's specialized workflow contributes beyond the capabilities of the underlying coding agent. A natural baseline would be a general-purpose coding agent using the same underlying model, with access to the same mathematical tools and comparable computational budgets, but without Peppy's prescribed pipeline, domain-specific language, or stage-by-stage prompting. Such a comparison would help determine whether the proposed workflow provides meaningful benefits beyond the capabilities of the underlying model and existing mathematical software.

Consequently, although the mathematical discoveries are valuable, the paper does not convincingly establish that the proposed pipeline offers a meaningful methodological contribution to AI research.

References

[1] Georgiev et al. (2025). Mathematical Exploration and Discovery at Scale. arXiv:2511.02864.

[2] OpenAI (2026). An OpenAI Model Has Disproved a Central Conjecture in Discrete Geometry. OpenAI.

[3] Gideoni et al. (2026). Simple Baselines Are Competitive with Code Evolution. arXiv:2602.16805.

[4] Wang et al. (2026). Rethinking the Evaluation of Harness Evolution for Agents. arXiv:2607.12227.

[5] Brilliantov et al. (2026). How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering? arXiv:2609.40303.

[6] Pan et al. (2026). HarnessTax: How Much Does Harness Matter for Coding Agents? Project and study.

[7] Liu & Han (2026). What Does a Harness Buy? Tokens, Mostly. arXiv:2610.04433.

Questions for authors

Question on the added value of Peppy

Could the authors demonstrate the incremental value of Peppy by comparing it against a general-purpose coding agent using the same underlying model, with access to the same mathematical tools and comparable computational budgets, but without Peppy's prescribed workflow, domain-specific language, and stage-by-stage prompting? In particular, does Peppy improve proof discovery success rates, computational efficiency, or the amount of human intervention required compared to this simpler baseline?

5 comments

We thank the reviewer for recognizing the value of our mathematical discoveries, including the resolution of conjectures that had remained open since 2017. We are glad that the reviewer views these results as “meaningful advances in the theoretical understanding of classical optimization algorithms” that “demonstrate the potential of AI-assisted mathematical research.” Below, we clarify the goals and contributions of Peppy and address the reviewer’s concerns regarding its added value.

The reviewer has pointed out a large body of work which appears to indicate that agentic workflows do not provide much additional value on top of using the base model’s capabilities. We would like to highlight two points.

  • The reviewer’s criticism relies largely on general arguments and cited studies questioning agentic workflows, without explaining how their findings apply to Peppy’s specific methodological contribution.
  • As mentioned in our common response, our focus is on obtaining the desired Lyapunov analyses, rather than establishing general superiority in proof-discovery success rates or token efficiency. Peppy aims to make the AI-assisted discovery of research-level convergence proofs in first-order optimization more transparent, understandable, and reproducible. We explain this in more detail below and kindly ask the reviewer to evaluate our work based on this clarified viewpoint and the specific design principles of Peppy.

What is the added value of Peppy? This is two-fold. First, Peppy specifies the desired Lyapunov-function-based proof structure and provides concrete mathematical instructions and verifiable success criteria for obtaining it, making the resulting proofs easier for researchers to understand and verify. Second, Peppy has five separable blocks, each representing a sub-task whose success can be assessed. This makes the proof discovery process transparent, allowing users to easily observe how each step is performed, and interact with the overall procedure if needed.

Because we view Peppy's main contribution as obtaining the desired Lyapunov analyses with transparency in the discovery process and simplicity of proof structure, we did not focus on designing evaluation benchmarks or extensive ablation studies demonstrating that Peppy outperforms a base LLM. The experiments in our Section 4 were designed to highlight reproducibility, in the sense that Peppy is capable of strictly adhering to our specifically designed proof-searching strategy over repeated runs. Section 5 applies our framework to more challenging, open research questions, which are precisely what Peppy is designed to handle, and obtains relatively simple Lyapunov-style resolutions as desired.

Nevertheless, as discussed in our common response, we tested the base model on the same five open problems without Peppy's instructions. As reported in conjectures_high_level_prompt_baseline.zip in the supplementary material, the base model failed to obtain the desired form of Lyapunov analyses in four of the five cases. We believe this provides concrete evidence of Peppy's incremental value over the base model.

We will revise our paper to clarify the goal and scope of Peppy, with more emphasis on the desired Lyapunov proof structure and the inspectable, interactive discovery process. We hope that the clarifications and additional materials address the reviewer’s concerns. If so, we would appreciate your reconsideration of our work. If there are other questions or points to discuss, please let us know and we will be happy to respond.

Thank you for the clarification and additional experiments. I acknowledge that my original criticism should be reframed in light of the authors' explanation that Peppy primarily aims to produce transparent and interpretable Lyapunov analyses, rather than merely solving convergence conjectures.

However, my central concern remains the same as in the original review: the incremental value of Peppy's specialized workflow over a general-purpose coding agent has not been convincingly established. In particular, I remain unconvinced about its added value beyond providing a general-purpose coding agent with a prompt along the following lines:

Here is an open problem: [problem statement].

Here is the mathematical framework I would like you to use: Yoon et al. (2026).

You can use the PEPFlow library and other mathematical tools as needed.

I want the final result to be a compact, summation-free Lyapunov analysis using a fixed number of interpretable algorithmic vectors, with analytically specified coefficients.

The additional baseline experiments do not resolve this concern. From the supplementary materials, the general-purpose coding agent was not given either the Yoon et al. framework or an explicit requirement to produce a compact, summation-free representation. Nevertheless, the absence of such a representation was subsequently used to classify four of the five outputs as unsuccessful.

Consequently, these experiments demonstrate that the coding agent does not necessarily produce the authors' preferred proof structure without being explicitly asked to do so, but they do not establish that Peppy's specialized workflow provides meaningful advantages over a general-purpose agent given the same mathematical framework and desired output requirements. I recognize that formulating the desired mathematical objective can itself be a contribution, but the paper also positions the five-stage workflow as a central contribution, whose incremental value remains unclear.

An additional question concerns the value of the desired proof structure itself. I acknowledge that my expertise in optimization theory is limited, and I do not question the established theoretical value of Lyapunov analysis. However, I would appreciate a more concrete explanation of what additional mathematical understanding the particular Lyapunov representations produced by Peppy provide. As far as I understand, if the authors had obtained rigorous proofs of the same convergence bounds in an arbitrary form, the main mathematical conclusions of the paper would remain unchanged. What additional insights would have been lost? For example, do the discovered Lyapunov functions reveal previously unknown structural properties of the algorithms, facilitate generalization to related settings, or suggest new algorithmic designs? While these are plausible benefits of Lyapunov analysis in general, I do not see them clearly demonstrated for the specific results in this paper. Since obtaining the desired proof structure is now emphasized as Peppy's primary objective, I believe it is important to clarify what additional value this structure provides beyond establishing the conjectured bounds.

I continue to appreciate the mathematical contributions and the authors' additional efforts to address the reviewers' concerns. Nevertheless, my original concern about the incremental AI methodological contribution remains unresolved.

We thank the reviewer for the thoughtful follow-up and for acknowledging our explanation of Peppy's objectives and the mathematical contributions of our work. We appreciate the opportunity to further discuss the role of Peppy's specialized workflow and the mathematical value of the resulting Lyapunov analyses.

Incremental value of Peppy's specialized workflow

We appreciate the reviewer's concrete example of a high-level prompt. As the reviewer also acknowledges, formulating the desired mathematical objective can itself be a contribution. We would like to emphasize that constructing the proposed prompt already incorporates an important part of Peppy's contribution, as it requires identifying an appropriate mathematical framework and specifying the desired structure of the Lyapunov analysis. These requirements reflect domain-specific mathematical knowledge that Peppy aims to make accessible to general-purpose coding agents.

Regarding the incremental value of the five-stage workflow, we would like to emphasize that, in addition to guiding the agent with an appropriate domain-specific mathematical framework, the workflow provides two complementary benefits:

  • Decomposition into verifiable intermediate steps: The workflow breaks down the construction of the Lyapunov function into verifiable intermediate steps. The quantities, derivations, and corresponding verification checks at each stage are organized in a single Jupyter notebook. This makes the entire proof construction process, not just the final result, transparent and inspectable.
  • Human interaction interface: The notebook is updated at each stage, providing human researchers with an interactive interface to monitor progress and provide additional guidance.

Peppy's main goal is therefore not one-shot autonomous proof generation, but rather to support an iterative mathematical discovery process in which human researchers can inspect and guide the agent's progress. Although this paper focuses on successful cases, the same structure could also be useful in failure cases, helping researchers identify where the process went wrong.

For some of the open problems in Section 5, we interacted with Peppy to inspect intermediate analyses and provide additional guidance, as illustrated in Appendices F.2 and F.3. Such guidance further simplified the analyses and improved their interpretability. For reference, we have also attached a complicated proof of FGM Conjecture 5 obtained without such additional guidance, to give a sense of the complexity involved. In fgm_conjecture5_complicated.pdf, establishing the positivity of the Lyapunov coefficients requires algebraic verification of the polynomial coefficients listed in Tables 1 and 2.

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

Peppy is an agentic workflow of five prompt-specified stages built on PEPFlow. It turns numerical PEP certificates into analytic, horizon-independent Lyapunov proofs, with SymPy checking the identities. The authors evaluate it on five known algorithms (30/30 runs succeed) and use it to close Conjectures 4–5 of Taylor et al. (2017c) for FGM, the finite-horizon gap for OGM at y_N, and the FGM-rational and FISTA-rational conjectures of Taylor et al. (2017a). Each bound comes with a matching lower-bound instance.

Critical strengths
  1. The results are real and checkable. These are decade-old conjectures, the Lyapunov certificates are explicit, and tightness is closed by Huber or linear instances (Lemmas F.12–F.13).
  2. I really like this idea of domain-specific proof shape. Fixing the proof to "weighted interpolation residuals minus sum of squares" reduces verification to rational identities plus coefficient signs.
Critical weaknesses

(1) Unfortunately, I feel that tThe method's contribution is hard to separate from the model's. There is no baseline: the same model with PEPFlow but without the Peppy commands, or without lyapunov_utils.py. Without that ablation, 30/30 on known algorithms could just reflect the LLMs capability, rather than their tools' . (2) Sadly, the training data is lilkely contaminated.

  • The "clean room" is enforced only by instructions: no filesystem isolation, three protocol flags, and access logs the authors admit are not exhaustive.
  • The convincing evidence is therefore the open problems, but those are n=1 each, with unreported failed attempts, cost, and number of human interventions. Table 2 should exist for Section 5. (4) The positivity lemmas (F.2, F.4, F.6, F.10) are prose arguments with induction on ρ_k bounds and chains of numeric constants (e.g. the β_k bound in F.6). None of it is machine-checked. The Lean skill is mentioned but apparently unused. Why didn't they follow through with Lean check here?
Questions for authors
  • How many runs did the FGM conjectures take, including failures, and at what token and time cost?
  • On the five known algorithms, does a no-Peppy baseline with the same model and PEPFlow access succeed?
  • Why didn't you perform lean-verification?
4 comments

We thank the reviewer for recognizing the capabilities of our workflow in producing “real and checkable” results that resolve “decade-old conjectures,” and for providing constructive feedback. We also appreciate the reviewer’s detailed recognition that “the Lyapunov certificates are explicit” and “tightness is closed,” as well as the reviewer’s positive assessment of our “domain-specific proof shape.” Below, we provide responses to specific points raised by the reviewer.

W1

W1. Unfortunately, I feel that the method's contribution is hard to separate from the model's. There is no baseline: the same model with PEPFlow but without the Peppy commands, or without lyapunov_utils.py. Without that ablation, 30/30 on known algorithms could just reflect the LLMs capability, rather than their tools'.

We understand the reviewer’s concern, and we agree that distinguishing the contribution of our workflow from the model's capability is indeed important. We would like to clarify that we do not claim that Peppy outperforms the base model in terms of proof-discovery success rates or end-to-end computational efficiency. Peppy aims at providing more understandable proofs and making the proof search procedure more tractable and inspectable. The 30/30 verified proofs in Section 4 demonstrate that Peppy can consistently follow these guidelines to produce admissible Lyapunov-style proofs. As discussed in our common response, specifying the desired proof structure and verifiable success criteria is itself part of our workflow, making a direct comparison with the base model difficult. Nevertheless, we tested the base model on the same five open problems without Peppy's instructions, and it failed to obtain the desired form of Lyapunov analyses in four of the five cases, as detailed in the supplementary material.

We kindly ask the reviewer to assess our work based on the above perspective. Our goal is to find human-friendly solutions to research-level problems in optimization by leveraging the base LLM's capabilities through principled instructions, not to claim that the LLM itself is incapable of doing so.

W3, Q1

W3. The convincing evidence is therefore the open problems, but those are n=1 each, with unreported failed attempts, cost, and number of human interventions. Table 2 should exist for Section 5.

Q1. How many runs did the FGM conjectures take, including failures, and at what token and time cost?

We would first like to remind the reviewer that the conjectures addressed in Section 5 were long-standing open problems, reflecting their substantial difficulty. Our goal was to demonstrate that Peppy can be used to successfully tackle such challenging research problems, rather than to present examples for which repeated success without human involvement can be easily demonstrated. Therefore, we would appreciate it if the reviewer could understand that Section 5’s purpose is to (i) demonstrate that Peppy is applicable to challenging, open research-level problems, and (ii) to present the outcomes from it, which are mathematically significant results in their own right, and are arguably straightforward for human researchers to follow and verify.

Table 2 is suitable for Section 4, which tests whether Peppy can consistently and faithfully execute the prescribed five-stage workflow without human interaction for algorithms with known convergence bounds. However, Section 5 addresses open research problems, where we also used human guidance to simplify the SOS decompositions (as illustrated in Appendix F.2 and F.3). Therefore, we believe extensive ablation or benchmarking related to token usage and execution time is not necessary, as it does not align with the purpose of Section 5.

Nevertheless, for reference, we provide the resource usage for the initial five-stage workflows and the subsequent human-guided refinement work.

CaseInitial workflow task time (min)Input + output tokens (millions)
FGM, Conjecture 4128.3651.186
FGM, Conjecture 5131.1053.885
OGM, Conjecture 4123.5738.214
Rational FGM140.7574.573
Rational FISTA80.8111.505

The following table reports the additional resource usage for the human-guided refinements described in Appendix F.

CaseAdditional refinement task time (min)Additional input + output tokens (millions)
FGM, Conjecture 458.4829.461
FGM, Conjecture 562.5441.404

In our experience, the five core stages consistently produced valid Lyapunov analyses with Astra Ultra, although proof details such as SOS decompositions varied. We selected simpler proofs for presentation to facilitate human understanding.

W2, Q2

W2. Sadly, the training data is likely contaminated. The "clean room" is enforced only by instructions …

Q2. On the five known algorithms, does a no-Peppy baseline with the same model and PEPFlow access succeed?

Based on the instructions, each run used a separate working directory and a fresh context, flagging any access to files outside the directory. There were three flags concerning generic Python/Jupyter environment dependencies, but as we clarified in the paper, these flags did not necessarily indicate a violation of the clean-room protocol’s primary purpose: preventing the model from accessing previous runs’ intermediate files. Each run was required to independently generate its own intermediate artifacts as evidence that Peppy’s block-wise steps were performed as intended, and we believe this was not violated.

Furthermore, again, the purpose of Section 4 was to verify that the coding agent consistently and faithfully implemented what was instructed by Peppy. If what the reviewer meant by “training data is likely contaminated” is that the LLMs we used were likely already trained on proofs of the convergence rates we asked them to re-derive here, this possibility cannot be ruled out. However, it does not invalidate the purpose of these reproducibility experiments, which was to assess whether Peppy could construct a Lyapunov-style proof by following the required steps of Blocks 1–5, rather than to establish proof discovery independent of prior model knowledge. Therefore, the notion of success evaluated in Section 4 is not directly applicable to a model without Peppy's instructions.

Furthermore, as reported in Appendix E.2.2, the five PGM runs without a supplied SOS decomposition produced three distinct Lyapunov functions. Two explicit examples are presented in Appendices E.3.3 and E.3.5. This variation provides additional evidence that Peppy is neither simply repeating previous trials nor reproducing a fixed known proof.

W4, Q3

W4. The positivity lemmas (F.2, F.4, F.6, F.10) are prose arguments with induction on ρ_k bounds and chains of numeric constants (e.g. the β_k bound in F.6). None of it is machine-checked. The Lean skill is mentioned but apparently unused. Why didn't they follow through with Lean check here?

Q3. Why didn't you perform lean-verification?

We first clarify that we did machine-check the key algebraic identities and polynomial positivity certificates used in the coefficient-sign arguments of these lemmas. The accompanying Jupyter notebooks provide all necessary SymPy checks (e.g., Cells 33–35 for Lemma F.2, Cell 10 for F.4, Cells 31–32 for F.6, and Cell 19 for F.10, in the respective notebooks). Our intention was to highlight that our proof style reduces much of the verification to straightforward symbolic checks.

Nevertheless, we did perform Lean verification internally on all five convergence results in Section 5 using our lean-from-ipynb skill. We did not initially provide these files to avoid giving the impression that Lean verification is an essential component of our workflow. Our aim is to integrate verification into the same Jupyter notebooks where researchers monitor and interact with the discovery process. However, in response to the reviewer's question, we now provide the Lean verification files as conjectures_lean_files.zip in the supplementary material, including the required coefficient-sign arguments from the cited lemmas. For reference, we also report the time and token usage for this work.

ScopeElapsed timeInput + output tokens
Lean formalization and verification of all five results2 h 21 min 42 s129,499,648

We will revise our paper to clarify the goal and scope of Peppy, with more emphasis on the desired Lyapunov proof structure and the inspectable, interactive discovery process. We hope that the clarifications above address the reviewer’s concerns. If so, we would appreciate your reconsideration of our work. If there are other questions or points to discuss, please let us know and we will be happy to respond.

Thank you again for your constructive feedback. We would greatly appreciate it if you could review our responses and consider our clarifications in your overall assessment. Thank you very much for your time and consideration.

Reviewer_KMsgOpenReview ↗

Good math work, AI work is incremental

After having reviewed the response from the authors, I will retain my original score because I feel that the math work is strong, but the AI aspect remains incremental.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common response

We thank the reviewers for their feedback, and we are happy that all reviewers recognized the significance of our contribution to resolving the open conjectures. While we understand that the reviewers raised questions about comparisons with the base model, we appreciate the opportunity to clarify the aim of our workflow and its contribution in more detail. Our aim and contributions are to obtain Lyapunov analyses, and we believe this goes beyond simply resolving the conjectures.

Lyapunov analysis is often regarded as "human-readable," and obtaining a Lyapunov function even from an existing proof has often been considered an independent research-level problem. For example, the convergence proofs for the Optimized Gradient Method (OGM) [1] and OGM-G [2] were discovered when the methods were first introduced, while their Lyapunov analyses were developed in subsequent works [3–5]. Furthermore, finding proper basis vectors that constitute a Lyapunov function is a nontrivial research task, and [5] focused on the special vector z_k that appears in the basis.

However, clarifying the task of finding a Lyapunov function is nontrivial and requires domain-specific knowledge. While the concept of a Lyapunov function is well established in the literature, its general definition does not specify a simple structure of being representable by a small number of local vectors, which would make it useful for a human-understandable convergence analysis. We have observed that baseline LLM often struggles to identify such a desired form. For example, the Lyapunov function used in our proof of Theorem 5.1 involves only f(xk)−f(x⋆)f(x_k) - f(x_\star) and the vectors x0−x⋆x_0 - x_\star, ∇f(xk)\nabla f(x_k), zk+1−x⋆z_{k+1} - x_\star and yk+1−zk+1y_{k+1} - z_{k+1}, while the baseline models often delivered Lyapunov function involving undesirable double summations like ∑i,j=0kaij⟨∇f(xi),∇f(xj)⟩\sum_{i,j=0}^{k} a_{ij}\langle \nabla f(x_i),\nabla f(x_j)\rangle. In this context, providing AI with instructions specifying concrete tasks and success criteria for obtaining a desired form of proof is a central contribution of our workflow.

Since the task definition itself is part of our workflow, a direct quantitative comparison with the base model presents a fundamental difficulty. Instructing the base model in full detail to provide our desired proof structure would already partially transfer the design principle of Peppy, while completely omitting the Lyapunov-related instructions would make the tasks different. Therefore, we focused on evaluating whether Peppy consistently meets our success criteria, and provided the corresponding results for open questions, imposing stricter requirements than merely obtaining an arbitrary convergence proof. We believe that obtaining Lyapunov analyses for the open problems, which we consider an achievement beyond simply resolving the conjectures, provides clear evidence of the capabilities of our workflow. We also believe that the accompanying Jupyter notebook files provide clear evidence that the proofs were indeed generated through our workflow.

Also, while Peppy aims to streamline proof search by testing candidate structures numerically before symbolic derivation and verification, it also devotes computational resources to Jupyter notebooks for human monitoring and interaction. Therefore, an end-to-end computational cost comparison is inadequate for assessing Peppy's intended contribution.

In short, Peppy focuses on finding a desirable form of Lyapunov analyses and leverages AI agents through concrete five-stage instructions with verifiable outcomes. Despite the difficulty of a comprehensive comparison discussed above, we tested the base model on the same five open problems without Peppy's instructions, following the reviewers' suggestions. As reported in conjectures_high_level_prompt_baseline.zip in the supplementary material, the base model failed to obtain the desired form of Lyapunov analyses in four of the five cases. We believe this result provides concrete evidence of Peppy's incremental value over the base model in obtaining the desired Lyapunov analyses.

References

  • [1] Kim and Fessler. Optimized First-Order Methods for Smooth Convex Minimization. Mathematical Programming, 2016.
  • [2] Kim and Fessler. Optimizing the Efficiency of First-Order Methods for Decreasing the Gradient of Smooth Convex Functions. Journal of Optimization Theory and Applications, 2021.
  • [3] d'Aspremont, Scieur, and Taylor. Acceleration Methods. Foundations and Trends in Optimization, 2021.
  • [4] Park, Park, and Ryu. Factor-√2 Acceleration of Accelerated Gradient Methods. Applied Mathematics & Optimization, 2023.
  • [5] Lee, Park, and Ryu. A Geometric Structure of Acceleration and Its Role in Making Gradients Small Fast. NeurIPS, 2021.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.