Pith. sign in

REVIEW 3 major objections 5 minor 4 cited by

DrSR: LLM based Scientific Equation Discovery with Dual Reasoning from Data and Experience

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read LLM symbolic regression is improved by a closed loop of data insight and reflective idea extraction.

desk verdict A plausible incremental extension of LLM-SR with a serious internal inconsistency in the headline Oscillator 1 result. read the letter →

arxiv 2506.04282 v1 pith:FL6E2B4S submitted 2025-06-04 cs.LG

classification cs.LG
keywords symbolicregressionlargelanguagemodelsscientificequationdiscoverydata-awareinsightinductiveideaextractionclosed-loopreasoningskeletongenerationbenchmarkevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that large language models can discover scientific equations much more reliably when their generation is guided by two additional reasoning streams: explicit analysis of the raw data, and distilled lessons from previous successes and failures. It introduces DrSR, a framework in which a data-analysis module inspects a sample of input-output pairs and residual patterns, an idea-extraction module categorizes each candidate equation as positive, negative, or invalid and summarizes reusable heuristics, and both streams feed back into the equation generator. On six benchmarks spanning physics, chemistry, biology, and materials science, DrSR reports higher valid equation rates, lower normalized mean squared error, better out-of-distribution generalization, and faster convergence than LLM-SR, PySR, and other baselines. A sympathetic reading of the result is that the bottleneck in LLM-based symbolic regression is not generation ability alone but the absence of structured observation and self-correction, and that closing this loop lets the model explore the expression space far more efficiently.

What carries the argument

The load-bearing mechanism is a closed loop of three LLM-instantiations sharing one backbone: $\pi_{\text{data}}$ (data-aware insight), $\pi_{\text{idea}}$ (inductive idea extraction), and $\pi_{\text{main}}$ (equation skeleton generation). $\pi_{\text{data}}$ uniformly samples 100 points from the dataset, adds residuals $r_{t,i} = y_i - f^*(x_i)$ from the best current equation, and produces structured descriptions of monotonicity, nonlinearity, correlation, and inter-variable relations; these insights are refreshed every time a better fit appears. $\pi_{\text{idea}}$ labels each candidate as positive, negative, or invalid, extracts a one-line lesson (e.g., avoid array broadcast errors, include interaction terms), and stores it in a JSON idea library sampled from the most recent half. Both $\mathcal{D}$ and $\mathcal{I}$ are injected into the prompt of $\pi_{\text{main}}$, which generates skeletons optimized by BFGS; the loop repeats for up to 1000 iterations. The paper's Bayesian framing makes the work of this machinery precise: it converts the static prior $p_{\text{LLM}}(f)$ into a dynamic, context-aware prior $p_{\text{LLM}}(f \mid \mathcal{D}, \mathcal{I})$ that sharpens as the search proceeds.

What would settle it

Run DrSR with the data-insight output replaced by random or deliberately wrong structural hints (e.g., 'possible exponential relationship' when truth is polynomial) across the same six benchmarks; if accuracy and valid rate do not drop substantially, the insight module is not the driver of the reported gains. Alternatively, run the search on a dataset whose 100-sample subset is unrepresentative (e.g., points concentrated in one region) and check whether DrSR fails to recover the true equation while a full-data variant succeeds.

Watch

Extended reading notes

Core claim

The central claim is that coupling data-aware insight with inductive idea extraction transforms LLM-based symbolic regression from a prior-driven generator into a self-correcting scientific process, and that this yields state-of-the-art accuracy on the tested tasks. Concretely, DrSR's data module samples 100 points (augmented with residuals of the current best equation) and emits structured descriptions such as 'possible cubic relationship in the range -0.2 to 0.2'; the idea module classifies each generated skeleton as positive, negative, or invalid and stores concise lessons in a JSON library; the main generator receives both insight and ideas in its prompt. The paper reports that DrSR achieves an NMSE of $1.80\times10^{-12}$ on Oscillator 2 versus $4.59\times10^{-5}$ for LLM-SR, exceeds 90% accuracy under strict tolerances on several tasks, maintains higher valid equation rates (0.4–0.6 vs sub-0.3 on complex datasets), and converges in 1000 iterations where baselines need 2000. The authors interpret this through a Bayesian reformulation in which the generative prior becomes $p_{\text{LLM}}(f \mid \mathcal{D}, \mathcal{I})$, conditioned on evolving data insight and idea library, rather than a static language-based prior.

Load-bearing premise

The data-aware insight module, using only 100 sampled points plus residuals from the current best equation, produces structural hints (e.g., 'possible cubic relationship') that are accurate enough to steer the skeleton search toward the true expression; if the LLM's data analysis is noisy or misleading, the reflective loop could guide the search away from the correct region, and the method's advantage would vanish or reverse.

Editorial extensions

If this is right

  • DrSR raises the valid equation rate, cutting the fraction of generations wasted on syntax errors, numerical overflows, or variable mismatches, and thereby improves sample efficiency under a limited iteration budget.
  • The reported accuracy and NMSE gains—including near-exact recovery on Oscillator 2 and 90%+ tolerance accuracy on several tasks—would make DrSR a practical tool for extracting interpretable laws from experimental data in materials, biology, and chemistry.
  • Strong out-of-distribution performance (e.g., OOD NMSE $2.80\times10^{-11}$ on Oscillator 2 where uDSR degrades to $9.59\times10^{3}$) suggests the discovered expressions capture latent structure rather than memorizing the training distribution.
  • Because the loop updates insight from residuals, the framework naturally refines its structural understanding from coarse global patterns to finer local and higher-order structure over iterations, mirroring a scientist's deepening understanding.
  • The Bayesian reinterpretation implies the method can be seen as adaptive prior learning; this could transfer to other LLM-guided search tasks where the prior is initially weak and data is plentiful.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 100-point sampling plus residual refinement suggests a general recipe: a small, actively-updated data summary may be sufficient to steer an LLM through a combinatorial search space; one testable extension is to vary sample size and noise to find when the insight signal breaks.
  • The idea library mechanism is essentially a structured memory of failure modes; this pattern could generalize to other generative settings (e.g., code synthesis, PDE discovery) where reflection on invalid outputs is currently ignored.
  • The paper's framing implies that the value of an LLM in scientific discovery is not just its internal knowledge but its ability to integrate external observations; one could probe this by measuring how much performance degrades when the data insight is computed by a different, weaker model.
  • Since the ablation shows both components contribute and their removal degrades performance, a natural next step left implicit is to study how the insights and ideas themselves evolve—whether they converge to the true structure or merely to a local optimum of the prompt distribution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes DrSR, an LLM-based symbolic regression framework that extends LLM-SR with two reasoning modules: a data-aware insight module (πdata) that summarizes structural patterns from sampled data and residual-augmented data, and an inductive idea extraction module (πidea) that distills positive, negative, and error experiences into an evolving idea library. The main generation loop prompts an LLM to propose equation skeletons, fits their parameters with BFGS, and feeds back both insights and ideas to improve subsequent proposals. Experiments are reported on six tasks (two nonlinear oscillators, E. coli growth, stress-strain, two LSR-Transform tasks, and one LSR-Synth task) with Mixtral-8x7B and LLaMA3.1 backbones, against gplearn, PySR, DSR, uDSR, LaSR, and LLM-SR. The paper claims consistently higher accuracy (ACC_tau), lower NMSE, faster convergence, and higher valid solution rates, and it includes a Bayesian reinterpretation of the search objective.

Significance. The proposed architecture is a natural and plausible improvement over LLM-SR: the two modules are clearly motivated, the ablation on Oscillator 2 shows that each contributes, and the valid-rate improvement is practically important for LLM-based search. If the reported numbers are reliable, DrSR would be a strong state-of-the-art LLM-based symbolic regression method. The paper's strengths include a clean algorithmic description (Algorithms 1-3), comparisons against six diverse baselines, and a useful discussion of the insight/idea loops with prompt examples. Its main weaknesses are statistical: no repeated seeds or error bars anywhere, a small and partially selected benchmark set, and one internal contradiction in the headline Oscillator 1 result that prevents verification of the reported NMSE. No code or data release is mentioned, which further weakens reproducibility.

major comments (3)
  1. [Appendix D / Table 1] The best-equation display for Oscillator 1 is inconsistent with the reported NMSE. Appendix D states the ground truth as dv/dt = 0.8 sin(x) - 0.5 x v - 0.5 v^3 - 0.2 x^3 - x cos(x), and shows the best DrSR equation as the same expression without the -x cos(x) term. Yet Table 1 reports DrSR (Mixtral) NMSE = 3.14e-7 and DrSR (Llama) NMSE = 5.40e-7 on this task. With the stated initial conditions x0 = 0.5, v0 = 0.5 over t in [0,50], omitting -x cos(x) leaves residuals whose magnitude is not negligible relative to the target variance; an NMSE near 1e-7 would require the omitted term to be essentially zero throughout the test set, which is implausible for this simulation. The authors must reconcile the displayed equation with the reported metric: either the equation is incomplete, the data generation differs from Appendix B.1, or the metric was computed against a different target. As written, the headline accuracy numbers for Oscillator 1, and by extension the evaluation pipeline for the other five tasks, cannot be verified.
  2. [Section 4.4 / Table 1 / Figures 2-4] Every empirical result is reported as a single point estimate with no repeated seeds, confidence intervals, or error bars. The LLM backbones are stochastic and Appendix E acknowledges that outputs can be verbose and variable; BFGS can also converge to different local optima from different skeletons. Under these sources of variance, margins such as DrSR (Llama) Acc = 3.64% vs LLM-SR (Llama) Acc = 1.36% on E. coli growth, or the valid-rate differences in Figure 4, may be within run-to-run noise. Since the central claim is empirical superiority, the paper should provide distributions over at least 3-5 independent runs per configuration for Table 1 and for the ablation in Section 5.5.
  3. [Section 5.2 / Figure 2] The generalization claim is not assessable because the paper never defines the out-of-distribution (OOD) split. Figure 2 plots ID vs OOD NMSE, but the text does not say how OOD inputs are constructed (e.g., extrapolation in time, new parameter regimes, or extended variable ranges). Without an explicit definition, the claim that DrSR 'learns latent rules behind data rather than overfitting' cannot be checked, and the dramatic OOD improvements (e.g., Oscillator 2 OOD NMSE 2.80e-11 vs uDSR 9.59e3) cannot be reproduced. Please add a precise description of the OOD protocol for each dataset.
minor comments (5)
  1. [Table 1] The header 'GPlern' is a typo for 'gplearn'; also the column 'LSR-Transform-2Avg' should be named consistently with the dataset description in Section 4.1 ('LSR-Transform').
  2. [Section B.5] The reaction-kinetics equation contains '0.1899_z' with an unexplained subscript z; this appears to be a typographical or formatting artifact and should be fixed.
  3. [Section 3.4] The Bayesian formulation in Equations (1)-(2) is purely interpretive and no quantitative priors or likelihoods are specified; adding a sentence clarifying that this is a high-level reinterpretation rather than a computational model would avoid overstating its role.
  4. [Appendix A / reproducibility] No code or data availability statement is provided. Given the dependence on specific prompts (Appendix C), dataset splits, and LLM decoding parameters, releasing the implementation and data-generation scripts is necessary for reproducibility.
  5. [Appendix D / Figure 11] The text refers to terms 'highlighted in pink', but the figure in the submitted PDF appears without color highlighting; please ensure the final version preserves the highlighting or describes the marked terms in text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: DrSR's stated gains rest on held-out test evaluation, not on a fitted input renamed as a prediction or on a self-citation chain.

full rationale

DrSR is an empirical LLM-based symbolic-regression pipeline; it does not contain a mathematical derivation that reduces to its own inputs. The BFGS-fitted coefficients are optimized on training data and then evaluated on ID/OOD test metrics described in Sections 4.3 and Appendix F, so the reported NMSE/ACC values are not definitionally equal to the fitting objective. The data-aware insight module (Section 3.1, Algorithm 3) consumes training samples and residuals, which is ordinary supervised fitting rather than a prediction forced by construction; the residual loop updates the skeleton search, but this is a heuristic optimization loop, not a tautology. The Bayesian reformulation in Section 3.4 is explicitly non-objective-changing ('This formulation does not change the task objective'), so it cannot be a load-bearing circular step. The paper's citations to LLM-SR and LLM-SRBench are to other author groups and are used for benchmark construction, not to import a contested uniqueness claim or an unverified ansatz. No fitted parameter is renamed as a prediction, and no result is justified solely by a self-citation. One in-scope inconsistency exists outside circularity: Appendix D reports the best DrSR equation for Oscillator 1 as omitting the ground-truth '-x cos(x)' term while Table 1 lists NMSE 3.14e-7, which is a reproducibility and accuracy-reporting concern, but it does not constitute a reduction of the claim to its inputs. Because the central comparisons are against external baselines on external benchmark tasks and the reported test metrics are not constructed to equal the training fit, the derivation is self-contained with respect to circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method introduces no new physical entities. The free parameters are hyperparameters chosen by hand; the axioms are the correctness of the benchmark labels, the LLM's prompt-following ability, and the adequacy of BFGS for parameter fitting.

free parameters (4)
  • data sample size for insight = 100
    Number of input-output pairs sampled for data analysis; chosen by hand, affects the quality of structural hints.
  • candidate skeletons per iteration = 4
    Number of equation skeletons sampled per iteration; larger values increase compute, smaller values may reduce coverage.
  • proportion of recent ideas sampled (lambda) = 0.5
    Fraction of most recent idea entries per category used in prompts; controls recency versus diversity.
  • insight sampling temperature and top-k/top-p = temperature 0.6, top-k 30, top-p 0.3
    Decoding parameters for the data-analysis LLM; chosen ad hoc without sensitivity analysis.
assumptions (3)
  • domain assumption The benchmark ground-truth expressions are correct and representative of the scientific domains.
    Used to define ACC_tau and NMSE targets; if labels are wrong, the reported comparisons lose meaning.
  • ad hoc to paper The LLM backbones can follow the structured prompts and generate useful data insights and ideas.
    The entire method depends on this capability; Section 3.1 and Appendix C assume the model produces accurate structural descriptions rather than hallucinations.
  • ad hoc to paper BFGS converges to adequate parameter values for most skeletons.
    Parameter optimization is done with BFGS (Section 4.4); if it fails, candidate equations are discarded as invalid or scored poorly.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DrSR: LLM based Scientific Equation Discovery with Dual Reasoning from Data and Experience." pith.science (2026). https://pith.science/paper/FL6E2B4S

@misc{pith2026250604282,
  author       = {Pith},
  title        = {Pith review of: DrSR: LLM based Scientific Equation Discovery with Dual Reasoning from Data and Experience},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FL6E2B4S}},
  note         = {Machine review of arXiv:2506.04282}
}
read the original abstract

Symbolic regression is a fundamental tool for discovering interpretable mathematical expressions from data, with broad applications across scientific and engineering domains. Recently, large language models (LLMs) have demonstrated strong performance in this task, leveraging embedded scientific priors and reasoning capabilities to surpass traditional methods. However, existing LLM-based approaches, such as LLM-SR, often over-rely on internal priors, lacking explicit data understanding and systematic reflection during equation generation. To address these limitations, we propose DrSR (Dual Reasoning Symbolic Regression), a framework that combines data-driven insight with reflective learning to enhance both robustness and discovery capability. Specifically, DrSR guides LLMs to analyze structural relationships (e.g., monotonicity, nonlinearity, and correlation) within the data to generate structured descriptions. Simultaneously, it monitors equation performance and establishes a feedback loop to refine subsequent generations. By integrating data understanding and generation reflection in a closed loop, DrSR enables more efficient exploration of the symbolic expression space. Experiments across interdisciplinary datasets in physics, chemistry, biology, and materials science demonstrate that DrSR substantially improves the valid equation rate and consistently outperforms both classical and recent LLM-based methods in terms of accuracy, generalization, and search efficiency. These results underscore its potential for scientific equation discovery.

Figures

Figures reproduced from arXiv: 2506.04282 by the authors.

Figure 1
Figure 1. Overview of the DrSR framework. (a) Data-aware Insight: the LLM analyzes variable relationships to derive structured, data-informed insights; (b) Inductive Idea Extraction: based on equation fitness, the LLM summarizes successful and failed patterns into an evolving idea library; (c) Equation Generation and Selection: guided by (a) and (b), the LLM generates new equation skeletons, optimizes parameters, and caches t… view at source ↗
Figure 2
Figure 2. Generalization across scientific domains under ID and OOD settings. DrSR achieves the lowest NMSE across all tasks and settings, indicating strong generalization performance. We further assess generalization by evaluating performance under both in-distribution (ID) and out￾of-distribution (OOD) settings [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Training convergence comparison. DrSR demonstrates faster convergence and lower final error across all tasks, consistently outperforming baselines even in early-stage iterations. 5.4 Valid Solution Rate Analysis To further explain the superior efficiency of DrSR, we evaluate the valid solution rate—the proportion of generated equations that are syntactically valid, compilable, and executable [PITH_FULL_IMAGE:figure… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Valid solution rate comparison. We report the proportion of syntactically valid, compilable, and evaluable equations produced by DrSR and LLM-SR. 5.5 Ablation Study We conduct ablation experiments on the Oscillator 2 task using Mixtral backbone to evaluate the contribu…
Figure 5
Figure 5. Figure 5: Ablation results on the Oscillation 2 problem, showing the impact of Data-aware Insight and Inductive Idea Extraction on the performance of DrSR. Impact of Data-aware Insight. We vary the probability p ∈ {0, 0.5, 1.0} with which the main LLM references structured insig…
Figure 6
Figure 6. Figure 6: Prompt for Data-aware Insight. The optimized function skeleton you just answered scored higher. Please summarize useful experience. STRICTLY follow these rules: 1. Use the exact phrasing "when seeking for the mathematical function skeleton that represents {dependent_na…
Figure 7
Figure 7. Figure 7: Prompt for Inductive Idea Extraction. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Initial Data-aware Insight. "output_format": { "analysis": { "independent_to_dependent_relationships": { "x ": [ "The relationship between x and a is a non-linear one, with a possible cubic relationship in the range of -0.2 to 0.2. However, the relationship becomes mor…
Figure 9
Figure 9. Figure 9: Final Data-aware Insight. The following are previous experience summaries that can help you provide better answers. experience 1: When I optimize the equation framework, in order to avoid the output failure of the generated function framework, I need to ensure that the…
Figure 10
Figure 10. Figure 10: Final Inductive Idea Extraction. D: Results Presentation [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Performance trajectory of DrSR with the best-performing equation programs discovered [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Unified remnant models for aligned-spin, precessing, and eccentric binary black hole mergers

    gr-qc 2026-08 conditional novelty 6.0 of 10

    New analytic fits, gwModelRemS/P, predict remnant mass, spin, luminosity, and kick for black hole mergers from equal mass to q=1000, with a neural-flow model for precessing kicks.

  2. Library Reachability in LSR-Synth: How Anti-Memorization Design Changes the Measurement of Symbolic Discovery

    cs.AI 2026-07 conditional novelty 6.0 of 10

    A fixed operator library with no access to variable names or domain descriptions solves most LSR-Synth tasks, so LLM candidates rarely expand the solvable set unless the library is deliberately weakened.

  3. NSR-Boost: A Neuro-Symbolic Residual Boosting Framework for Industrial Legacy Models

    cs.AI 2026-01 conditional novelty 6.0 of 10

    NSR-Boost patches a frozen GBDT by adding LLM-generated, Bayesian-tuned symbolic experts in error regions, beating baselines on six public datasets and improving a deployed financial model.

  4. Large language models for partial differential equation workflows

    cs.AI 2026-08 conditional novelty 2.0 of 10

    A review organizing LLM-assisted PDE research into Discovery, Solving, and Optimization stages, arguing LLMs are most useful as workflow-level interfaces rather than isolated solvers.

Reference graph

Works this paper leans on

58 extracted references · 40 canonical work pages · cited by 4 Pith papers

  1. [1]

    Interpretable scientific discovery with symbolic regression: a review.Artificial Intelligence Review, 57(1):2, 2024

    Nour Makke and Sanjay Chawla. Interpretable scientific discovery with symbolic regression: a review.Artificial Intelligence Review, 57(1):2, 2024

  2. [2]

    Rediscover- ing orbital mechanics with machine learning.Machine Learning: Science and Technology, 4(4):045002, 2023

    Pablo Lemos, Niall Jeffrey, Miles Cranmer, Shirley Ho, and Peter Battaglia. Rediscover- ing orbital mechanics with machine learning.Machine Learning: Science and Technology, 4(4):045002, 2023

  3. [3]

    Discovery of a planar black hole mass scaling relation for spiral galaxies.The Astrophysical Journal Letters, 956(1):L22, 2023

    Benjamin L Davis and Zehao Jin. Discovery of a planar black hole mass scaling relation for spiral galaxies.The Astrophysical Journal Letters, 956(1):L22, 2023

  4. [4]

    Emerging materials intelligence ecosystems propelled by machine learning.Nature Reviews Materials, 6(8):655–678, 2021

    Rohit Batra, Le Song, and Rampi Ramprasad. Emerging materials intelligence ecosystems propelled by machine learning.Nature Reviews Materials, 6(8):655–678, 2021

  5. [5]

    Fast, accurate, and transferable many-body interatomic potentials by symbolic regres- sion.npj Computational Materials, 5(1):112, 2019

    Alberto Hernandez, Adarsh Balasubramanian, Fenglin Yuan, Simon AM Mason, and Tim Mueller. Fast, accurate, and transferable many-body interatomic potentials by symbolic regres- sion.npj Computational Materials, 5(1):112, 2019

  6. [6]

    Generating synergistic formulaic alpha collections via reinforcement learning

    Shuo Yu, Hongyan Xue, Xiang Ao, Feiyang Pan, Jia He, Dandan Tu, and Qing He. Generating synergistic formulaic alpha collections via reinforcement learning. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5476–5486, 2023

  7. [7]

    Hrft: Mining high- frequency risk factor collections end-to-end via transformer.arXiv preprint arXiv:2408.01271, 2024

    Wenyan Xu, Rundong Wang, Chen Li, Yonghong Hu, and Zhonghua Lu. Hrft: Mining high- frequency risk factor collections end-to-end via transformer.arXiv preprint arXiv:2408.01271, 2024

  8. [8]

    Distilling free-form natural laws from experimental data

    Michael Schmidt and Hod Lipson. Distilling free-form natural laws from experimental data. science, 324(5923):81–85, 2009

Show all 58 references
  1. [9]

    Interpretable machine learning for science with pysr and symbolicregression

    Miles Cranmer. Interpretable machine learning for science with pysr and symbolicregression. jl. arXiv preprint arXiv:2305.01582, 2023

  2. [10]

    Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients.arXiv preprint arXiv:1912.04871, 2019

    Brenden K Petersen, Mikel Landajuela, T Nathan Mundhenk, Claudio P Santiago, Soo K Kim, and Joanne T Kim. Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients.arXiv preprint arXiv:1912.04871, 2019

  3. [11]

    Neural symbolic regression that scales

    Luca Biggio, Tommaso Bendinelli, Alexander Neitz, Aurelien Lucchi, and Giambattista Paras- candolo. Neural symbolic regression that scales. InInternational Conference on Machine Learning, pages 936–945. Pmlr, 2021

  4. [12]

    End-to-end symbolic regression with transformers.Advances in Neural Information Processing Systems, 35:10269–10281, 2022

    Pierre-Alexandre Kamienny, Stéphane d’Ascoli, Guillaume Lample, and François Charton. End-to-end symbolic regression with transformers.Advances in Neural Information Processing Systems, 35:10269–10281, 2022

  5. [13]

    Llm-sr: Scientific equation discovery via programming with large language models

    Parshin Shojaee, Kazem Meidani, Shashank Gupta, Amir Barati Farimani, and Chandan K Reddy. Llm-sr: Scientific equation discovery via programming with large language models. arXiv preprint arXiv:2404.18400, 2024

  6. [14]

    Mixtral of experts.arXiv preprint arXiv:2401.04088, 2024

    Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts.arXiv preprint arXiv:2401.04088, 2024

  7. [15]

    Llama-3.1- foundationai-securityllm-base-8b technical report.arXiv preprint arXiv:2504.21039, 2025

    Paul Kassianik, Baturay Saglam, Alexander Chen, Blaine Nelson, Anu Vellore, Massimo Aufiero, Fraser Burch, Dhruv Kedia, Avi Zohary, Sajana Weerawardhena, et al. Llama-3.1- foundationai-securityllm-base-8b technical report.arXiv preprint arXiv:2504.21039, 2025

  8. [16]

    In-context symbolic regression: Leveraging large language models for function discovery.arXiv preprint arXiv:2404.19094, 2024

    Matteo Merler, Katsiaryna Haitsiukevich, Nicola Dainese, and Pekka Marttinen. In-context symbolic regression: Leveraging large language models for function discovery.arXiv preprint arXiv:2404.19094, 2024

  9. [17]

    Llm-srbench: A new benchmark for scientific equation discovery with large language models.arXiv preprint arXiv:2504.10415, 2025

    Parshin Shojaee, Ngoc-Hieu Nguyen, Kazem Meidani, Amir Barati Farimani, Khoa D Doan, and Chandan K Reddy. Llm-srbench: A new benchmark for scientific equation discovery with large language models.arXiv preprint arXiv:2504.10415, 2025. 10

  10. [18]

    The growth of bacterial cultures

    Monod and J. The growth of bacterial cultures. 3(1):371–394, 1949

  11. [19]

    Convenient model to describe the combined effects of temperature and ph on microbial growth.Applied and environmental microbiology, 61(2):610–616, 1995

    L Rosso, JR Lobry, S Bajard, and Jean-Pierre Flandrois. Convenient model to describe the combined effects of temperature and ph on microbial growth.Applied and environmental microbiology, 61(2):610–616, 1995

  12. [20]

    Ai feynman: A physics-inspired method for symbolic regression.Science advances, 6(16):eaay2631, 2020

    Silviu-Marian Udrescu and Max Tegmark. Ai feynman: A physics-inspired method for symbolic regression.Science advances, 6(16):eaay2631, 2020

  13. [21]

    A unified framework for deep symbolic regression.Advances in Neural Information Processing Systems, 35:33985– 33998, 2022

    Mikel Landajuela, Chak Shing Lee, Jiachen Yang, Ruben Glatt, Claudio P Santiago, Ignacio Aravena, Terrell Mundhenk, Garrett Mulcahy, and Brenden K Petersen. A unified framework for deep symbolic regression.Advances in Neural Information Processing Systems, 35:33985– 33998, 2022

  14. [22]

    Symbolic regression with a learned concept library.Advances in Neural Information Processing Systems, 37:44678–44709, 2024

    Arya Grayeli, Atharva Sehgal, Omar Costilla Reyes, Miles Cranmer, and Swarat Chaudhuri. Symbolic regression with a learned concept library.Advances in Neural Information Processing Systems, 37:44678–44709, 2024

  15. [23]

    Information processing, data inferences, and scientific generalization.Behav- ioral Science, 19(5):314–325, 1974

    Donald Gerwin. Information processing, data inferences, and scientific generalization.Behav- ioral Science, 19(5):314–325, 1974

  16. [24]

    Bacon: A production system that discovers empirical laws

    Pat Langley. Bacon: A production system that discovers empirical laws. InIJCAI, page 344. Citeseer, 1977

  17. [25]

    Genetic programming as a means for programming computers by natural selection

    John R Koza. Genetic programming as a means for programming computers by natural selection. Statistics and computing, 4:87–112, 1994

  18. [26]

    Symbolic physics learner: Discovering governing equations via monte carlo tree search.arXiv preprint arXiv:2205.13134, 2022

    Fangzheng Sun, Yang Liu, Jian-Xun Wang, and Hao Sun. Symbolic physics learner: Discovering governing equations via monte carlo tree search.arXiv preprint arXiv:2205.13134, 2022

  19. [27]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the national academy of sciences, 113(15):3932–3937, 2016

    Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the national academy of sciences, 113(15):3932–3937, 2016

  20. [28]

    Data-driven discovery of coordinates and governing equations.Proceedings of the National Academy of Sciences, 116(45):22445–22451, 2019

    Kathleen Champion, Bethany Lusch, J Nathan Kutz, and Steven L Brunton. Data-driven discovery of coordinates and governing equations.Proceedings of the National Academy of Sciences, 116(45):22445–22451, 2019

  21. [29]

    Extrapolation and learning equations.arXiv preprint arXiv:1610.02995, 2016

    Georg Martius and Christoph H Lampert. Extrapolation and learning equations.arXiv preprint arXiv:1610.02995, 2016

  22. [30]

    Learning equations for extrapolation and control

    Subham Sahoo, Christoph Lampert, and Georg Martius. Learning equations for extrapolation and control. InInternational Conference on Machine Learning, pages 4442–4450. Pmlr, 2018

  23. [31]

    Integration of neural network-based symbolic regression in deep learning for scientific discovery.IEEE transactions on neural networks and learning systems, 32(9):4166– 4177, 2020

    Samuel Kim, Peter Y Lu, Srijon Mukherjee, Michael Gilbert, Li Jing, Vladimir ˇCeperi´c, and Marin Soljaˇci´c. Integration of neural network-based symbolic regression in deep learning for scientific discovery.IEEE transactions on neural networks and learning systems, 32(9):4166...

  24. [32]

    Prunesymnet: A symbolic neural network and pruning algorithm for symbolic regression.arXiv preprint arXiv:2401.15103, 2024

    Min Wu, Weijun Li, Lina Yu, Wenqiang Li, Jingyi Liu, Yanjie Li, and Meilan Hao. Prunesymnet: A symbolic neural network and pruning algorithm for symbolic regression.arXiv preprint arXiv:2401.15103, 2024

  25. [33]

    Metasymnet: A tree-like symbol network with adaptive architecture and activation functions

    Yanjie Li, Weijun Li, Lina Yu, Min Wu, Jingyi Liu, Shu Wei, Yusong Deng, and Meilan Hao. Metasymnet: A tree-like symbol network with adaptive architecture and activation functions. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 27081–27089, 2025

  26. [34]

    Stanford University, Department of Computer Science Stanford, CA, 1990

    John R Koza.Genetic programming: A paradigm for genetically breeding populations of computer programs to solve problems, volume 34. Stanford University, Department of Computer Science Stanford, CA, 1990

  27. [35]

    Interaction–transformation evolutionary algorithm for symbolic regression.Evolutionary computation, 29(3):367–390, 2021

    Fabricio Olivetti de Franca and Guilherme Seidyo Imai Aldeia. Interaction–transformation evolutionary algorithm for symbolic regression.Evolutionary computation, 29(3):367–390, 2021. 11

  28. [36]

    Taylor genetic program- ming for symbolic regression

    Baihe He, Qiang Lu, Qingyun Yang, Jake Luo, and Zhiguang Wang. Taylor genetic program- ming for symbolic regression. InProceedings of the genetic and evolutionary computation conference, pages 946–954, 2022

  29. [37]

    Sr-forest: a genetic pro- gramming based heterogeneous ensemble learning method.IEEE Transactions on Evolutionary Computation, 2023

    Hengzhe Zhang, Aimin Zhou, Qi Chen, Bing Xue, and Mengjie Zhang. Sr-forest: a genetic pro- gramming based heterogeneous ensemble learning method.IEEE Transactions on Evolutionary Computation, 2023

  30. [38]

    Symbolic regression via neural-guided genetic programming population seeding.arXiv preprint arXiv:2111.00053, 2021

    T Nathan Mundhenk, Mikel Landajuela, Ruben Glatt, Claudio P Santiago, Daniel M Faissol, and Brenden K Petersen. Symbolic regression via neural-guided genetic programming population seeding.arXiv preprint arXiv:2111.00053, 2021

  31. [39]

    Improving exploration in policy gradient search: Application to symbolic optimization.arXiv preprint arXiv:2107.09158, 2021

    Mikel Landajuela, Brenden K Petersen, Soo K Kim, Claudio P Santiago, Ruben Glatt, T Nathan Mundhenk, Jacob F Pettit, and Daniel M Faissol. Improving exploration in policy gradient search: Application to symbolic optimization.arXiv preprint arXiv:2107.09158, 2021

  32. [40]

    Interactive reinforcement learning for symbolic regression from multi-format human-preference feedbacks

    Laure Crochepierre, Lydia Boudjeloud-Assala, and Vincent Barbesant. Interactive reinforcement learning for symbolic regression from multi-format human-preference feedbacks. InIJCAI, pages 5900–5903, 2022

  33. [41]

    A neural-guided dynamic symbolic network for exploring mathematical expressions from data.arXiv preprint arXiv:2309.13705, 2023

    Wenqiang Li, Weijun Li, Lina Yu, Min Wu, Linjun Sun, Jingyi Liu, Yanjie Li, Shu Wei, Yusong Deng, and Meilan Hao. A neural-guided dynamic symbolic network for exploring mathematical expressions from data.arXiv preprint arXiv:2309.13705, 2023

  34. [42]

    Discover: Deep identification of symbolically concise open-form partial differential equations via enhanced reinforcement learning.Physical Review Research, 6(1):013182, 2024

    Mengge Du, Yuntian Chen, and Dongxiao Zhang. Discover: Deep identification of symbolically concise open-form partial differential equations via enhanced reinforcement learning.Physical Review Research, 6(1):013182, 2024

  35. [43]

    Disco-dso: Coupling discrete and continuous optimization for efficient generative design in hybrid spaces

    Jacob F Pettit, Chak Shing Lee, Jiachen Yang, Alex Ho, Daniel Faissol, Brenden Petersen, and Mikel Landajuela. Disco-dso: Coupling discrete and continuous optimization for efficient generative design in hybrid spaces. InProceedings of the AAAI Conference on Artificial Intellig...

  36. [44]

    Symbolicgpt: A generative transformer model for symbolic regression (2021).arXiv preprint arXiv:2106.14131, 2021

    Mojtaba Valipour, Bowen You, Maysum Panju, and Ali Ghodsi. Symbolicgpt: A generative transformer model for symbolic regression (2021).arXiv preprint arXiv:2106.14131, 2021

  37. [45]

    Symformer: End-to-end symbolic regression using transformer-based architecture.IEEE Access, 2024

    Martin Vastl, Jonáš Kulhánek, Jiˇrí Kubalík, Erik Derner, and Robert Babuška. Symformer: End-to-end symbolic regression using transformer-based architecture.IEEE Access, 2024

  38. [46]

    Transformer-based model for symbolic regression via joint supervised learning

    Wenqiang Li, Weijun Li, Linjun Sun, Min Wu, Lina Yu, Jingyi Liu, Yanjie Li, and Songsong Tian. Transformer-based model for symbolic regression via joint supervised learning. InThe Eleventh International Conference on Learning Representations, 2022

  39. [47]

    Symbolic regression via mdlformer-guided search: from minimizing prediction error to minimizing description length.arXiv preprint arXiv:2411.03753, 2024

    Zihan Yu, Jingtao Ding, and Yong Li. Symbolic regression via mdlformer-guided search: from minimizing prediction error to minimizing description length.arXiv preprint arXiv:2411.03753, 2024

  40. [48]

    Niert: Accurate numerical interpolation through unifying scattered data representations using transformer encoder.IEEE Transactions on Knowledge and Data Engineering, 2024

    Shizhe Ding, Boyang Xia, Milong Ren, and Dongbo Bu. Niert: Accurate numerical interpolation through unifying scattered data representations using transformer encoder.IEEE Transactions on Knowledge and Data Engineering, 2024

  41. [49]

    Language model crossover: Variation through few-shot prompting.ACM Transactions on Evolutionary Learning, 4(4):1–40, 2024

    Elliot Meyerson, Mark J Nelson, Herbie Bradley, Adam Gaier, Arash Moradi, Amy K Hoover, and Joel Lehman. Language model crossover: Variation through few-shot prompting.ACM Transactions on Evolutionary Learning, 4(4):1–40, 2024

  42. [50]

    output_format

    Ping Guo, Qingfu Zhang, and Xi Lin. Coevo: Continual evolution of symbolic solutions using large language models.arXiv preprint arXiv:2412.18890, 2024. 12 Appendix A: Implementation and Experimental Details A.1 Baseline Configurations For gplearn, we use a symbolic regressor w...

  43. [52]

    Summarize ONLY the key success factors

  44. [53]

    What lessons can you draw from it? STRICTLY follow these rules:

    You need to make your answer as concise as possible The optimized function skeleton you just answered scored lower. What lessons can you draw from it? STRICTLY follow these rules:

  45. [54]

    when seeking for the mathematical function skeleton that represents {dependent_name_in_prompt} in {problem_name_in_prompt}, I can

    Use the exact phrasing "when seeking for the mathematical function skeleton that represents {dependent_name_in_prompt} in {problem_name_in_prompt}, I can ..."

  46. [56]

    You need to make your answer as concise as possible The optimized function skeleton you just answered failed with error: {error_for_sample[i]}, What lessons can you draw from it? STRICTLY follow these rules:

  47. [57]

    when seeking for the mathematical function skeleton that represents {dependent_name_in_prompt} in {problem_name_in_prompt}, I need

    Use the exact phrasing "when seeking for the mathematical function skeleton that represents {dependent_name_in_prompt} in {problem_name_in_prompt}, I need ..."

  48. [58]

    Address the SPECIFIC error: {error_for_sample[i]}

  49. [59]

    Identify ONE crucial improvement point

  50. [60]

    output_format

    You need to make your answer as concise as possible Prompt 4 - Inductive Idea Extraction (Positive case) Prompt 5 - Inductive Idea Extraction (Negative case) Prompt 6 - Inductive Idea Extraction (Invalid case) Figure 7: Prompt for Inductive Idea Extraction. 17 ### The followin...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.