acceptodds
Under review as a conference paper at ICLR 2027

STRIDE: A Self-Reflective Agent Framework for Reliable Automatic Equation Discovery

Abstract

LLM-based equation discovery offers a promising route to recovering symbolic laws from data, but many systems still rely on generation-centered loops that propose candidates, fit parameters, score results, and reuse selected examples. Such loops can misjudge useful skeletons under unreliable fitting, discard near-correct equations that require repair, and accumulate redundant memories that provide limited guidance. We propose STRIDE, a self-reflective agent framework that improves reliability by coordinating data-aware generation, mixed-fitting evaluation, critic–executor repair, and diversity-preserving semantic memory. STRIDE turns fitting scores and candidate behavior into shared feedback to guide targeted structural repair, while retaining diverse, high-quality equations to inform subsequent generation, thereby closing the loop between evaluation, refinement, and sampling. Experiments on representative symbolic-regression benchmarks and LSR-Synth suites show that STRIDE improves accuracy, OOD robustness, and structural recovery across multiple LLM backbones, with ablations and analyses confirming the contribution of its core components.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
symbolic regression, equation discovery, large language models, AI for science, program synthesis, scientific machine learning, agentic AI

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vNVnOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes STRIDE for discovering interpretable equations from observational data. The system uses training-data hints and high-performing equations from previous iterations to guide candidate generation, separates linear from nonlinear parameters during fitting, revises equations through a critic and an executor, and retains diverse high-scoring candidates based on structural similarity. The authors evaluate prediction error, out-of-domain performance, and ground-truth term coverage using two language-model backbones on the original LLM-SR tasks and a selected subset of LSR-Synth, with ablations examining the role of each component. The paper also includes a supplementary experiment on training the critic.

Critical strengths

The full search pipeline demonstrates practical effectiveness across multiple tasks and two language-model backbones. Revision suggestions are implemented as executable equations, which are then refitted and scored; retained candidates subsequently guide further generation. The main experiments show favorable predictive and out-of-domain performance on oscillator, chemical-reaction, and biological population tasks, while the ablations and augmented LLM-SR controls also support the value of combining these components. Although the evidence is limited to the selected benchmarks, these results indicate practical value that warrants further investigation.

Critical weaknesses
  1. The paper’s contributions beyond prior work remain insufficiently clear. Many reflective agent methods across different fields already have models propose candidates, critique them, and revise them accordingly; equation-discovery methods have also combined data analysis with repeated evaluation, as in DoLQ[1]: a model proposes equations → a program fits their parameters → feedback is provided based on the results → revision and search continue. STRIDE brings these steps together into one system, which has value; however, while the existing experiments support the usefulness of these designs, the authors should clarify how they differ substantively from prior methods and how much of the improvement comes from targeted repair versus more thorough parameter fitting and additional computation.

  2. The central motivation lacks direct experimental validation. Figure 1 presents “promising structures being discarded because parameter fitting fails,” “nearly correct equations not being repaired,” and “redundancy in the experience library limiting search” as the main problems, but the figure is only schematic. Comparisons of fitting methods and ablation experiments show that the relevant modules are useful, but they do not directly quantify how often these failures occur or how many of them the corresponding modules actually correct. The LLM-SR controls augmented with mixed fitting and data hints in Figure 5 are helpful; further comparisons are needed to determine whether critic-guided repair outperforms simply generating more candidates under the same computational budget. Otherwise, it is difficult to distinguish the benefits of targeted feedback from those of additional search.

  3. The evidence for structural recovery is insufficient to support strong claims about scientific discovery. STC (Symbolic Term Coverage) in Table 2 measures only the fraction of ground-truth terms recovered; it does not adequately reflect whether the predicted expression contains additional incorrect terms, nor does it establish that the entire equation structure is correct. The paper does not report whole-equation symbolic accuracy and does not sufficiently specify the rules for splitting equations into terms, matching equivalent terms, or aggregating results across tasks. Moreover, the experiments use only a small number of tasks from the original LLM-SR work and a subset of LSR-Synth selected using the authors’ difficulty ranking, whose weights and implementation are not fully specified. The current results should be described more cautiously as improvements in prediction and partial structural recovery. Symbolic Accuracy (SA) is itself an important metric in LLM-SRBench.

  4. The structural judgments in the data hints lack sufficient justification. Statistics such as the mean and standard deviation can be calculated directly, but they do not by themselves establish whether the true equation requires a constant term or exhibits odd or even symmetry. For multivariate data in particular, testing symmetry requires holding the other variables fixed and explaining how the output after reversing the sign of the variable under examination is obtained or estimated. The paper does not sufficiently specify the calculations or thresholds underlying these judgments, making it difficult to assess when the hints are trustworthy. I am also concerned about the reliability of odd/even symmetry hints for complex multivariate relationships. Such relationships do not necessarily exhibit clear symmetry with respect to individual variables, and approximate symmetry observed in the data may hold only within the sampled range. The authors should explain how these properties are tested using observational data and analyze how inaccurate hints affect the search results.

  5. My concern regarding “dominant terms” is that the contribution of one term depends on the other candidate terms: different features or terms may substitute for one another, and their coefficients may change depending on the combination, making their effects difficult to disentangle. Indeed, I believe this coupling is difficult to disentangle, and sparse regression alone does not automatically eliminate its effects.

Refs: [1] S ong, S. K., Shin, B. G., & Lee, J. Y. (2026). Discovering Ordinary Differential Equations with LLM-Based Qualitative and Quantitative Evaluation. Proceedings of ICML 2026, PMLR, 306, 115029–115067.

Questions for authors

Same with the Weaknesses.

2 comments
AuthorsOpenReview ↗

Follow-up on the Discussion

Dear Reviewer,

As the discussion period comes to a close, we wanted to follow up on the evidence you requested for the failure modes in Figure 1. Our replies report a comparison with DoLQ, a fixed-skeleton fitting analysis, a matched-budget comparison of repair strategies, and a memory-coverage ablation. We also clarify term precision and whole-equation equivalence, how the benchmark tasks were selected, and how to interpret the data-derived hints when features are correlated.

Thank you for taking another look at our responses. We would be glad to discuss any points that remain open. If the additional evidence resolves your main concerns, we would appreciate your reconsidering your rating in light of it.

Best regards,

The authors of Submission 7853

Thank you for clarifying the differences from DoLQ and adding the repair comparison. In my view, the overall methodological advance is still modest. My overall rating remains unchanged.

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

The paper proposes STRIDE, a self-reflective agent framework for reliable automatic equation discovery, aiming to improve the performance of LLM-based symbolic regression in prediction accuracy, structure recovery, and out-of-domain generalization. STRIDE consists of four main modules: data-aware generation, hybrid parameter fitting, Critic-Executor reflective repair, and semantic memory. Specifically, the data-aware generation module uses statistical features of the training data and historically good expressions to guide candidate generation; the hybrid fitting module divides parameters into linear coefficients and nonlinear parameters, and estimates them using linear least squares and nonlinear optimization, respectively; the reflective repair module proposes local modifications based on fitting results and the structure of candidate expressions, and screens revision results through re-execution, fitting, and scoring; the semantic memory module clusters candidates according to the structural features of expressions and retains excellent representatives in each structural cluster to support subsequent search.

Critical strengths

This paper focuses on parameter fitting, structure repair, and search efficiency in LLM-driven symbolic regression, and has certain research significance. STRIDE integrates candidate generation, hybrid parameter fitting, Critic-Executor reflective repair, and semantic memory into a complete search framework. Each module has a clear functional division, and revision results are screened through execution feedback. Its hybrid parameter fitting improves parameter estimation by combining linear least squares with nonlinear optimization, which has reasonable technical motivation. In addition, the paper conducts experiments in multiple scientific domains, examines out-of-domain generalization and noise robustness, and analyzes the role of each module through ablation experiments, providing some experimental basis for the effectiveness of the method.

Critical weaknesses
  1. Insufficient methodological innovation and similarity to existing LLM-based symbolic regression methods. This paper takes fitting-feedback-driven expression reflection, structural correction, and utilization of historical experience as its core innovations, but this idea has already been adopted by various LLM symbolic regression methods. For example, LLM-SR has implemented iterative search through expression generation, parameter fitting, performance evaluation, and excellent candidate memory; LaSR uses high-quality expressions to extract abstract concepts to guide subsequent search; DrSR combines data analysis and search feedback for reflection and dynamically adjusts expression generation; SR-Scientist also implements iterative optimization through agent-based data analysis, expression evaluation, and feedback. In contrast, STRIDE mainly differs in the specific implementations of data-aware generation, hybrid parameter fitting, Critic-Executor correction, and semantic memory, but has not fully embodied a new search idea. Therefore, the authors are suggested to clarify the substantive innovation differences from existing methods and verify the independent contributions of these differences through targeted experiments.

  2. Evaluation metrics are insufficient to verify the structural recovery ability of true expressions. This paper mainly uses NMSE, ACC, and ACCMax to evaluate algorithm performance. These metrics reflect numerical fitting accuracy, rather than the structural recovery ability of the true expression. However, one of the core challenges of symbolic regression is identifying the true mathematical structure from a large number of candidate expressions with similar fitting accuracy, rather than merely reducing prediction error. Although the paper introduces SymbolicTermCoverage (STC), this metric only measures the coverage of true function terms and cannot determine whether the predicted expression is completely equivalent to the target expression. Therefore, the existing results are insufficient to prove that STRIDE improves the recovery ability of complete expressions. The authors are suggested to add a strict symbolic equivalence recovery rate and compare the structural recovery performance of different methods under the same search budget.

  3. The dataset scale and coverage are limited, making it difficult to fully verify the generality of the algorithm. This paper only uses 4 scientific equation discovery tasks and 5 high-difficulty tasks selected from each of the four domains of LSR-Synth, totaling 24 tasks. Although different scientific domains are involved, the dataset scale and the coverage of expression structures are still limited, making it difficult to fully verify the generality of the method. The authors are suggested to add commonly used symbolic regression benchmarks such as AlFeynman and Livermore, and systematically evaluate prediction accuracy and structural recovery rate under different numbers of variables, operator combinations, and expression complexity, so as to more fully verify the effectiveness and generalization ability of the algorithm.

Questions for authors

See Weaknesses

1 comment
AuthorsOpenReview ↗

Follow-up on the Discussion

Dear Reviewer,

We wanted to follow up on your questions about STRIDE’s methodological contribution, structural recovery, and benchmark coverage. Our replies clarify how mixed fitting, targeted repair, and structural memory differ from prior methods and point to experiments that test their contributions. We also report term precision and whole-equation equivalence judgments, and extend the evaluation to six LSR-Transform and Livermore tasks with the same full-fit budget across methods.

Thank you for considering the additional results. We welcome any follow-up questions before the discussion closes. If you find that the new analyses address your concerns, we would be grateful if you considered updating your rating.

Best regards,

The authors of Submission 7853

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

STRIDE is an LLM-based symbolic regression framework that combines data-informed equation generation, separate fitting of linear and nonlinear parameters, critic-executor edits, and memory organized by structural similarity. The paper evaluates two LLM backbones on four LLM-SR tasks and twenty selected LSR-Synth problems. It reports predictive accuracy, extrapolation, term coverage, and component ablations, along with an experiment in which the critic is trained from execution outcomes.

Critical strengths
  1. The framework addresses a useful and concrete problem: a promising equation structure can be discarded simply because its parameters have not been fitted well. Combining conditional linear fitting with local repairs that are fitted and evaluated again is well motivated. The mixed-fitting controls in Figure 5 help clarify this contribution beyond the comparison with the original LLM-SR.

  2. The reported gains on the tested oscillator, CRK, and BPG tasks are substantial under both backbones. For example, Table 7 reports 100% OOD [email protected] on CRK for STRIDE, versus 65.40%/41.56% for LLM-SR. The component ablations, Oscillator 1 noise experiment, and estimates of token use and runtime also make the results easier to assess.

Critical weaknesses
  1. It is difficult to tell how consistently STRIDE recovers the underlying equation structure from the current STC results. Section 4.2 defines STC as the fraction of independent ground-truth terms recovered but leaves term matching, coefficient tolerances, algebraic equivalence, and aggregation unspecified. As a recall-like measure, it also does not penalize spurious terms. Figure 15a illustrates why this matters: the STRIDE equation fits the data closely but omits the ground-truth sin(x) term and includes a linear x term. A reproducible scoring rule, together with term precision and whole-equation recovery for the existing outputs, would help distinguish accurate prediction from recovery of the underlying law.

  2. The experiments leave some uncertainty about how much of the improvement comes from critic guidance and how much comes from the additional search and fitting. The paper specifies 2000 STRIDE candidates versus 2500 for LLM-SR, but it is unclear whether screened and fully fitted revisions count toward that limit. Table 4 reports the estimated cost difference, at 3.5M versus 2.6M tokens. The remaining question is how the budget is used: the critic-removal ablation does not explain whether the repair budget is reallocated to ordinary generation. Figure 7 shows improvements following repair, but a comparison at a similar budget would help establish the benefit of critic guidance over ordinary proposals or unguided local edits.

Questions for authors
  1. Could the authors explain how STC is computed and whether its gains persist when spurious terms and algebraic equivalence are taken into account? The scoring rule and variability across the existing runs would be helpful. Do Figure 15a and Table 1 describe the same selected equation or different runs?

  2. What counts toward the candidate budget, including screening, refinement, and restarts? A comparison of one existing oscillator setting could help clarify the role of the critic: with the same mixed fitter and memory and comparable token/evaluation budgets, how does critic-guided repair compare with spending that budget on ordinary proposals or local edits without critic guidance?

  3. Deliberate Evolution (ICML 2026) would also be worth discussing. A comparison of its adaptive search operators, analytical diagnostics, and trajectory memory with STRIDE's fit-conditioned repair and structurally clustered memory could help explain where the two approaches differ.

3 comments
AuthorsOpenReview ↗

Follow-up on the Discussion

Dear Reviewer,

As the discussion period draws to a close, we wanted to follow up on your questions about structural recovery, budget accounting, and Deliberate Evolution. Our replies distinguish term-level recovery from whole-equation equivalence, explain how to read Figure 15a alongside Table 1, and compare critic-guided repair with ordinary generation and unguided editing at the same full-fit budget. We also clarified how STRIDE differs from Deliberate Evolution in search, feedback, and memory.

Thank you for taking the time to consider these additions. We would be happy to discuss any remaining questions. If the new evidence addresses your concerns, we would appreciate your reconsidering your rating.

Best regards,

The authors of Submission 7853

Thank you for the careful clarifications and the additional evaluations. The term-precision results and the comparison at 800 full fits are useful additions. My remaining reservations concern the evidence for consistent whole-equation recovery and the robustness of the critic's gains, which are currently supported by a single controlled run with higher token usage. I will therefore maintain my current assessment. I encourage the authors to build on these results with further validation of structural recovery and the accuracy-cost trade-off; the combination of fitting and targeted repair remains a promising direction.

Thank you for acknowledging the value of the additional evaluations and for sharing your remaining concerns. We will clarify the scope of our conclusions in the revised manuscript and further validate whole-equation recovery and the critic’s accuracy–cost trade-off across multiple runs. We appreciate your recognition of this promising direction.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Meta Response

Dear Area Chair and Reviewers,

Thank you for your constructive feedback. We have revised the manuscript, with changes highlighted in blue, and provided additional evidence in our individual responses. The main updates are:

  1. Methodological distinctions. We clarified STRIDE’s fitting, repair, and memory mechanisms, added DE and DoLQ to Section 5, and reported oscillator comparisons with DoLQ in our replies.STRIDE achieved lower ID and OOD NMSE and higher OOD [email protected] in these runs.

  2. Structural recovery. We clarified STC, added term precision to penalize extra terms, and reported SA judgments. STRIDE achieves the highest term precision under both backbones (Section 4.1 and Table 2). Appendix E clarifies the representative cases and Figure 15a’s structural errors.

  3. Budget and repair controls. We clarified the base-proposal budget (Section 4.1 and Appendix A.1). Our replies compare three search strategies with the fitter and memory fixed. Guided repair achieved the lowest ID and OOD NMSE in the reported Oscillator 2 control.

  4. Mechanism evidence. Additional analyses show how mixed fitting corrects skeleton rankings and how critic guidance reduces structural errors in adopted edits. We also added the equal-sample memory-coverage comparison. (Section 4.4 and Appendix B.4).

  5. Broader evaluation. Our reply to Reviewer LKFi reports predictive accuracy and SA comparisons with LLM-SR and SR-Scientist on additional LSR-Transform and Livermore tasks. STRIDE achieves leading predictive accuracy on most of these tasks, supporting its effectiveness and generalization.

  6. Reproducibility. We specified the difficulty-ranking rules (Appendix D.2), data-hint calculations and thresholds, and the interpretation of correlated candidate terms (Section 3.2 and Appendix B.1).

We welcome further questions and would appreciate your consideration of the revised manuscript and additional evidence in your final assessment.

Best regards,

The Authors

Dear Reviewers,

We are currently in the discussion phase. The authors have responded to the comments, and I would greatly appreciate it if you could take a moment to check the rebuttal and leave any follow-up comments or questions. If the response has addressed your concerns, please also consider updating your review accordingly.

Your feedback is very valuable to both the authors and the decision process. Thank you again for your help.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.