Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

PiT-PO claims that fine-tuning an LLM during symbolic search — with physics constraints and token-level redundancy penalties — recovers exact governing equations and beats prompt-only methods.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 01:04 UTC pith:PZYRSRJU

load-bearing objection A genuinely useful engineering recipe for LLM-based symbolic regression, with a theory garnish that doesn't quite license the heuristic threshold it claims to justify. the 3 major comments →

arxiv 2602.10576 v2 pith:PZYRSRJU submitted 2026-02-11 cs.LG cs.AI

LLM-Based Scientific Equation Discovery via Physics-Informed Token-Regularized Policy Optimization

Boxiao Wang , Kai Li , Tianyi Liu , Chen Li , Junzhe Wang , Yifan Zhang , Jian Cheng This is my paper
classification cs.LG cs.AI
keywords large language modelssymbolic regressionreinforcement learningphysics-informed constraintstoken-level regularizationequation discoveryturbulence modelinggroup relative policy optimization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper is trying to establish that an LLM can be turned from a static equation proposer into an adaptive generator by fine-tuning its weights during symbolic search, rather than only steering it with prompts. It claims that adding two kinds of constraints as learning signals — hierarchical physical-validity penalties and token-level penalties for redundant terms — lets the model recover ground-truth equations more often and more exactly than existing prompt-based LLM symbolic regression. If true, this would change how LLM-based scientific discovery is done: search feedback would update the model itself, small open models could rival much larger closed systems, and discovered equations would come with physical and parsimony guarantees. The paper supports this with benchmark results on two suites and a turbulence-closure case study.

Core claim

The central claim is PiT-PO: use group-relative policy optimization during evolutionary search to update LLM weights, with a reward that combines fit error, complexity penalty, and gated physics constraints, and a token-aware advantage that separately penalizes tokens belonging to redundant terms. The redundancy penalty is justified by a Support Exclusion Theorem that gives a condition under which a fitted term cannot be part of the true support; in practice the method marks any term with normalized coefficient below 1e-2 as redundant. PiT-PO is reported to recover the exact equation on Oscillation 1 (NMSE ~1e-31), achieve the highest symbolic accuracy on LLM-SRBench, lower OOD error on all

What carries the argument

The load-bearing mechanism is token-aware advantage estimation inside GRPO: a group-standardized global reward (fit minus complexity minus gated physics penalty) is modified per token by subtracting a local penalty for tokens in terms flagged redundant by the Support Exclusion Theorem. This lets the policy gradient increase probability of valid tokens and decrease probability of redundant tokens, effectively pruning the search space from inside the model via LoRA weight updates.

Load-bearing premise

The method's redundancy pruning assumes that a term with normalized fitted coefficient below 1e-2 is always safe to delete, but the theorem that justifies this requires conditions — a finite dictionary containing the true expansion, bounded true coefficients, sparsity, and an overlap between candidate and true support — that the practical threshold never verifies in the open-ended LLM generation space.

What would settle it

Construct a synthetic target that is a sum of one large term and one genuine small term whose normalized fitted coefficient falls below 1e-2; run PiT-PO and check whether the final equation keeps or deletes the small term. If it deletes it, the token-level redundancy penalty is removing true structure, not just redundancy. A complementary check is to compute the theorem's internal- and external-interference bounds on this example and compare them with the implemented threshold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • LLM-based symbolic regression should be treated as an adaptive learning problem, not a fixed proposer: feedback internalized into weights yields faster convergence and fewer stalls than prompt feedback alone.
  • Small open models (1B/3B/8B) can outperform much larger or closed-source baselines on equation recovery, lowering the compute barrier to scientific discovery.
  • Adding physics constraints as gated penalties plus token-level redundancy penalties improves out-of-distribution generalization, with lower OOD NMSE reported on all four suite tasks.
  • The discovered turbulence closure, embedded into a RANS solver, matches DNS better than the k-omega SST baseline for Reynolds stress, velocity, and skin friction on the tested periodic-hill case.
  • Exact structural recovery, not just numerical fit, becomes achievable: the method is the only one in the comparison reported to recover the exact ground-truth equation for Oscillation 1.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the token-level redundancy penalty is the main driver of the gains, a natural next experiment is to ablate the RL weight update itself and keep only the pruning rule, separating the contribution of local credit assignment from global policy optimization; the paper's ablations remove one constraint at a time but not the RL update.
  • The theorem's conditions suggest a direct testable extension: construct synthetic problems with a known dictionary and known sparsity where a true term has a small coefficient, then check whether the 1e-2 threshold deletes it; if so, the pruning rule is biased against small-amplitude genuine structure.
  • The turbulence result, if validated on hill slopes other than the training geometry (e.g., different alpha or Reynolds number), would position LLM-based symbolic regression as a practical route to data-driven turbulence closures; the current CFD validation shares the same flow geometry as the training data.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. PiT-PO couples in-search GRPO-based fine-tuning of an LLM with hierarchical physics penalties and token-level redundancy penalties for symbolic regression. The paper claims state-of-the-art results on the LLM-SR Suite and LLM-SRBench, exact structural recovery on Oscillator 1, improved OOD NMSE across all suite tasks, and a turbulence closure that, when embedded in RANS, more closely matches DNS than the baseline k-ω SST model. The theoretical centerpiece is a Support Exclusion Theorem (Theorem 3.1) intended to identify provably false candidate terms, which is then operationalized as a normalized fitted-coefficient threshold and token penalty.

Significance. If the empirical findings hold, the paper makes a useful contribution: it demonstrates that in-search policy optimization with constraint-aware, token-level credit assignment can substantially improve LLM-based symbolic regression, and that small quantized open-weight models can rival larger closed models. The evaluation is comparatively thorough: multiple benchmarks, multiple backbone scales, ID/OOD splits, wall-clock comparisons, ablations of both constraint components, and an end-to-end CFD case study. The proof-of-concept turbulence closure is a valuable stress test for LLM-driven SR in a real scientific workflow. The explanatory and reproducibility material (hyperparameters, prompts, trajectory analyses, appendices) is unusually complete.

major comments (3)
  1. [§3.1.2, Eq. (5), Eq. (6), Table 3] The implemented redundancy criterion is τ_i = |b_i|/(Σ_j|b_j|+ε) ≤ ρ, a normalized coefficient-magnitude threshold, but Theorem 3.1 gives an absolute bound |b_i| < A − (internal + external interference). The algorithm never constructs a finite dictionary S containing the true support, never estimates A, B, or M, never computes the Gram-matrix interference terms, and never checks the partial-correctness condition K∩S'≠∅ used in the proof. The theorem therefore does not justify the threshold; a true term with a genuinely small coefficient (|a_i| << B) can satisfy τ_i ≤ ρ and be penalized. The ablation shows the heuristic helps on these benchmarks, but the paper's central claim of 'rigorously' distinguishing essential from redundant terms via the theorem is not established. Either implement a verifiable dictionary-based bound or explicitly reframe the token penalty as a heuristic and soften
  2. [§4.2, Table 1] The text states that PiT-PO 'is the only approach to successfully identify the exact ground-truth equation for the Oscillator 1.' Table 1 shows LLM-SR (Mixtral) achieving 100.00% Acc_avg-0.001 and NMSE 1.32e-11 on that task. If 'exact' means symbolic recovery, the paper provides no symbolic-equivalence evidence for the baselines; if 'exact' means near machine-precision numerical recovery, the Mixtral baseline is already very close. The claim should be restricted to the same-backbone comparison or supported by a formal symbolic-equality check.
  3. [§4.6, Appendix E] The turbulence case study trains and evaluates on the same periodic-hill configuration (α=0.8, Re_b=5600; Appendix E.1 and Figure 4). The post-CFD comparison against DNS is a meaningful in-sample demonstration, but the paper's abstract and Section 4.6 present this as discovering 'novel turbulence models' with practical utility. Without at least one held-out geometry or different Reynolds number, the closure's generalization is untested. The claims should be scoped to an in-sample demonstration, or additional transfer experiments should be added.
minor comments (4)
  1. [Theorem 3.1 / Appendix B] The main-text statement of Theorem 3.1 omits the partial-correctness condition K∩S'≠∅ used in the proof to bound |R| ≤ M−1. Although the contradiction argument implicitly supplies it when i∈S', the theorem statement should state all assumptions explicitly for clarity.
  2. [Figure 3] The ablation figure does not label the task or dataset. It should state which benchmark and which ID/OOD split are shown, since the text claims a general conclusion from this figure.
  3. [Table 1] The baseline is listed as 'GPlern' but the text refers to 'GPlearn'; please fix the typo. Also, the Acc_avg tolerance and NMSE are reported together but some cells have odd spacing (e.g., '0.01148.05e-10' in Table 4), which should be corrected.
  4. [Related Work] The phrase 'a unifying limitation across these methods is their reliance on the LLM as a frozen generator' is an oversimplification: LLM-SR and LaSR do update the prompt library during search, so the distinction from weight-updating in-search tuning should be stated more precisely.

Circularity Check

0 steps flagged

No significant circularity: core derivation is self-contained; the theorem-implementation gap is a rigor issue, not a circular reduction.

full rationale

The paper's derivation chain is not circular. The Support Exclusion Theorem (Thm 3.1 / Appendix B.5) is proved from first-order optimality of least squares plus triangle inequalities; it does not assume its own conclusion and is not imported from a self-citation. GRPO is a standard external algorithm; Symbolic Accuracy uses an external GPT-4o judge; OOD splits are held out. The reported SOTA is an empirical claim tied to in-search RL and constraint rewards, not a quantity equal to its own input by construction. The main weaknesses are non-circular: (i) the implemented redundancy rule tau_i <= rho (§3.1.2, Table 3) uses only fitted coefficient magnitudes and never verifies the theorem's conditions (fixed dictionary S, |S'|<=M, A<=|a_j|<=B, K∩S'≠∅), so the 'theorem-guided' pruning is actually a heuristic threshold that could delete genuine small-coefficient terms; (ii) the turbulence closure is trained and CFD-validated on the same periodic-hill case (α=0.8, Re_b=5600, Appendix E), so no out-of-geometry generalization is demonstrated. These are correctness/validity concerns, not circular reductions: no fitted parameter is renamed as a prediction, and no load-bearing step is defined in terms of the result it is supposed to establish.

Axiom & Free-Parameter Ledger

8 free parameters · 7 axioms · 0 invented entities

The method rests on standard RL machinery (GRPO, LoRA), on the Support Exclusion Theorem's unverifiable assumptions (finite dictionary, known sparsity and coefficient bounds), and on turbulence-domain assumptions (Pope tensor basis, realizability, near-wall scaling). No new physical entities are introduced. The main hand-set parameters are the redundancy threshold ρ, token penalty scale p, complexity weight λ_len, physics penalty weights, gating threshold δ_gate, and the invariant normalization scale. The largest uncharged axiom is the claim that the theorem is instantiated by the algorithm: the implemented pruning is a fitted-coefficient threshold heuristic whose regime is not shown to match the theorem's sufficient condition.

free parameters (8)
  • redundancy threshold ρ = 1e-2
    A term is 'redundant' if normalized fitted-coefficient ratio τ_i ≤ ρ (Table 3); the central pruning decision of the method; hand-set without sensitivity analysis; no link shown between this threshold and the theorem's sufficient condition.
  • token penalty scale p = 0.5
    Scales P_tok = p·max(0, −log(|b_i|+ε)) in Eq. (6); hand-set.
  • complexity penalty weight λ_len = 5e-3
    Weights the AST-node-count complexity penalty P_cplx (Eq. 8); hand-set.
  • fitting log-MSE scale α = 1.0
    Scales R_fit = −α log(MSE+ε) (Eq. 7).
  • physics penalty weights (P_dim, P_diff, w_j) = 1.0, 0.5, 0.5
    Weights for general-level and domain-specific penalties (Table 3, Eq. 20); hand-set, no sensitivity analysis.
  • gating threshold δ_gate = 1e-3 · MSE_initial
    Physics penalties activate only after MSE < δ_gate (Eq. 9); depends on the initially sampled equation.
  • invariant normalization scale = 2.0 (in tanh(I/2.0))
    Preprocessing choice in Eq. (15) mapping invariants to [−1,1]; affects fit conditioning in the turbulence task.
  • RL/search hyperparameters (KL β, LoRA rank, lr, G, N) = 0.01, 16, 1e-6, 4, 4
    GRPO KL coefficient, LoRA rank 16, learning rate 1e-6, group size 4, 4 islands → effective batch 16; standard for the framework but hand-set.
axioms (7)
  • ad hoc to paper The true equation lies in a finite dictionary span(S) with sparsity |S'| ≤ M and coefficient bounds A ≤ |a_j| ≤ B
    Theorem 3.1's guarantee requires these quantities; the algorithm searches an open-ended LLM expression space and never specifies or estimates S, M, A, or B, so the guarantee does not transfer to the implemented pruning.
  • ad hoc to paper Current support is partially correct: K ∩ S' ≠ ∅
    Used in the proof (Step 2, Appendix B.2) to bound |R| ≤ M−1; omitted from the main-text Theorem 3.1 statement.
  • standard math Empirical Gram matrix has G_ii > 0 for all dictionary functions
    Needed to define T_ij = G_ji/G_ii; standard non-degeneracy for distinct basis functions on the data.
  • standard math Least-squares residual is empirically orthogonal to the active span (first-order optimality)
    Core identity of the Support Exclusion Theorem proof (Step 1, Appendix B.2).
  • standard math GRPO provides a valid clipped policy-gradient update with KL reference regularization
    Background RL machinery (Eqs. 2-4) assumed for the in-search update, including the token-aware advantage replacing the uniform group advantage.
  • domain assumption Pope's tensor-basis expansion: Reynolds stress anisotropy b = Σ_m G_m(I1,I2) T_m with fixed finite bases and two-invariant dependence
    The turbulence SR task (Eq. 16) restricts the learned closure to three fixed tensor bases and two scalar invariants; if the true anisotropy needs more bases or variables, the symbolic form is misspecified.
  • domain assumption The four turbulence constraints (realizability, wall decay, cubic near-wall scaling, TKE production) are trustworthy physics imposed as penalties
    P(1)-P(4) in Eqs. (21-25) rest on turbulence literature (Pope; Mochizuki & Osaka; Tennekes & Lumley); their violation penalties shape the reward.

pith-pipeline@v1.3.0-alltime-deepseek · 28125 in / 32418 out tokens · 279488 ms · 2026-08-03T01:04:55.655506+00:00 · methodology

0 comments
read the original abstract

Symbolic regression aims to distill mathematical equations from observational data. Recent approaches have successfully leveraged Large Language Models (LLMs) to generate equation hypotheses, capitalizing on their vast pre-trained scientific priors. However, existing frameworks predominantly treat the LLM as a static generator, relying on prompt-level guidance to steer exploration. This paradigm fails to update the model's internal representations based on search feedback, often yielding physically inconsistent or mathematically redundant expressions. In this work, we propose PiT-PO (Physics-informed Token-regularized Policy Optimization), a unified framework that evolves the LLM into an adaptive generator via reinforcement learning. Central to PiT-PO is a dual-constraint mechanism that rigorously enforces hierarchical physical validity while simultaneously applying fine-grained, token-level penalties to suppress redundant structures. Consequently, PiT-PO aligns LLM to produce equations that are both scientifically consistent and structurally parsimonious. Empirically, PiT-PO achieves state-of-the-art performance on standard benchmarks and successfully discovers novel turbulence models for challenging fluid dynamics problems. We also demonstrate that PiT-PO empowers small-scale models to outperform closed-source giants, democratizing access to high-performance scientific discovery.

Figures

Figures reproduced from arXiv: 2602.10576 by Boxiao Wang, Chen Li, Jian Cheng, Junzhe Wang, Kai Li, Tianyi Liu, Yifan Zhang.

Figure 1
Figure 1. Figure 1: The overall framework of PiT-PO. PiT-PO trans [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation results of PiT-PO and its variants. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Schematic of the geometries for periodic hills. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Non-dimensional stream-wise velocity contours [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: NMSE versus iteration on smaller backbones, con [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Iterative trajectories of PiT-PO and LLM-SR on Os [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Evaluation and Optimization. def equation(I1: np.ndarray, I2: np.ndarray, params: np.ndarray) -> np.ndarray: """ Predict three scalar coefficients G1, G2, G3 for tensor bases. Args: I1, I2: numpy arrays of shape (N,), invariants. params: numpy array of free constants. Returns: numpy array of shape (N, 3), columns are G1, G2, G3. """ # G1 block: params[0:10] g1 = ( params[0] * I1 + params[1] * I2 + params[… view at source ↗
Figure 11
Figure 11. Figure 11: Problem Specification. E.3 Back to Turbulence Modeling From discovered expressions to a deployable closure. After obtain￾ing an explicit expression for the Reynolds-stress-related term from symbolic regression, we rewrite the discovered formula into an OpenFOAM-readable form. Embedding into OpenFOAM. We embed the resulting closure into the existing kOmegaSST implementation by replacing the routine that ev… view at source ↗
Figure 13
Figure 13. Figure 13: Equation Program Example. E.4 Domain-Specific Constraints for Turbulence Modeling To encode expert knowledge as inductive biases, we augment the reward with a domain-specific penalty term: 𝑃domain = 4 ∑ 𝑗=1 𝑤𝑗 𝑃 (𝑗) domain, (20) [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. FunctionEvolve: Structure-Guided Symbolic Regression with LLMs

    cs.LG 2026-06 unverdicted novelty 7.0

    FunctionEvolve recovers 107 exact symbolic forms out of 129 synthetic tasks (82.9% SA@50) by using expression-tree structure for evolutionary search, parent selection, mutation, and coefficient scoring with LLMs.

Reference graph

Works this paper leans on

56 extracted references · 6 canonical work pages · cited by 1 Pith paper

  1. [1]

    B. S. Aakash, JohnPatrick Connors, and Michael D Shields. 2019. Stress-strain data for aluminum 6061-T651 from 9 lots at 6 temperatures under uniaxial and plane strain tension.Data in Brief 25 (Aug 2019), 104085.doi:10.1016/j.dib.2019. 104085

    doi:10.1016/j.dib.2019 2019

  2. [2]

    Luca Biggio, Tommaso Bendinelli, Alexander Neitz, Aurelien Lucchi, and Giambattista Parascandolo. 2021. Neural Symbolic Regression that Scales. arXiv:2106.06427 [cs.LG] https://arxiv.org/abs/2106.06427

    Pith/arXiv arXiv 2021

  3. [3]

    Biggio*, T

    L. Biggio*, T. Bendinelli*, A. Neitz, A. Lucchi, and G. Parascandolo. 2021. Neural Symbolic Regression that Scales. InProceedings of 38th International Conference on Machine Learning (ICML 2021) (Proceedings of Machine Learning Research, Vol. 139). PMLR, 936–945. https://proceedings.mlr.press/v139/biggio21a.html *equal contribution

    2021

  4. [4]

    Jindou Chen, Jidong Tian, Liang Wu, ChenXinWei, Xiaokang Yang, Yaohui Jin, and Yanyan Xu. 2025. KinFormer: Generalizable Dynamical Symbolic Re- gression for Catalytic Organic Reaction Kinetics. In International Conference on Representation Learning , Y. Yue, A. Garg, N. Peng, F. Sha, and R. Yu (Eds.), Vol. 2025. 67058–67080.https://proceedings.iclr.cc/pap...

    2025

  5. [5]

    Miles Cranmer. 2023. Interpretable Machine Learning for Science with PySR and SymbolicRegression.jl. arXiv:2305.01582 [astro-ph.IM] https://arxiv.org/ abs/2305.01582

    Pith/arXiv arXiv 2023

  6. [6]

    Laure Crochepierre, Lydia Boudjeloud-Assala, and Vincent Barbesant. 2022. In- teractive Reinforcement Learning for Symbolic Regression from Multi-Format Human-Preference Feedbacks. InIJCAI 2022- 31st International Joint Conference on Artificial Intelligence. Vienne, Austria.https://hal.science/hal-03695471

    2022

  7. [7]

    Song Deng, Junjie Wang, Li Tao, Su Zhang, and Hongwei Sun. 2023. EV charg- ingloadforecastingmodelminingalgorithmbasedonhybridintelligence. Com- puters and Electrical Engineering 112 (2023), 109010.doi:10.1016/j.compeleceng. 2023.109010

    arXiv 2023

  8. [8]

    Mengge Du, Yuntian Chen, and Dongxiao Zhang. 2023. DISCOVER: Deep iden- tification of symbolically concise open-form PDEs via enhanced reinforcement- learning. arXiv:2210.02181 [cs.LG] https://arxiv.org/abs/2210.02181

    Pith/arXiv arXiv 2023

  9. [9]

    1987.Practical Methods of Optimization (2nd ed.)

    Roger Fletcher. 1987.Practical Methods of Optimization (2nd ed.). John Wiley & Sons, Chichester, New York

    1987

  10. [10]

    Arya Grayeli,AtharvaSehgal, OmarCostilla-Reyes,Miles Cranmer,and Swarat Chaudhuri. 2024. Symbolic Regression with a Learned Concept Library. arXiv:2409.09359 [cs.LG] https://arxiv.org/abs/2409.09359

    Pith/arXiv arXiv 2024

  11. [11]

    Ping Guo, Qingfu Zhang, and Xi Lin. 2025. CoEvo: Continual Evolution of Symbolic Solutions Using Large Language Models. arXiv:2412.18890 [cs.AI] https://arxiv.org/abs/2412.18890

    arXiv 2025

  12. [12]

    EdwardJ.Hu,YelongShen,PhillipWallis,ZeyuanAllen-Zhu,YuanzhiLi,Shean Wang, Lu Wang, and WeizhuChen. 2021. LoRA: Low-RankAdaptation of Large Language Models. arXiv:2106.09685 [cs.CL] https://arxiv.org/abs/2106.09685

    Pith/arXiv arXiv 2021

  13. [13]

    Ziyao Huang, Weiwei Wu, Kui Wu, Jianping Wang, and Wei-Bin Lee. 2025. CALM: Co-evolution of Algorithms and Language Model for Automatic Heuris- tic Design. arXiv:2505.12285 [cs.NE] https://arxiv.org/abs/2505.12285

    Pith/arXiv arXiv 2025

  14. [14]

    Pierre-Alexandre Kamienny, Stéphane d’Ascoli, Guillaume Lample, and François Charton. 2022. End-to-end symbolic regression with transformers. arXiv:2204.10532 [cs.LG] https://arxiv.org/abs/2204.10532

    Pith/arXiv arXiv 2022

  15. [15]

    Paul Kassianik, Baturay Saglam, Alexander Chen, Blaine Nelson, Anu Vel- lore, Massimo Aufiero, Fraser Burch, Dhruv Kedia, Avi Zohary, Sajana Weerawardhena, Aman Priyanshu, Adam Swanda, Amy Chang, Hyrum Anderson, Kojin Oshiba, Omar Santos, Yaron Singer, and Amin Kar- basi. 2025. Llama-3.1-FoundationAI-SecurityLLM-Base-8B Technical Report. arXiv:2504.21039 ...

    Pith/arXiv arXiv 2025

  16. [16]

    Geneticallybreedingpopulationsofcomputerprogramstosolve problems in artificial intelligence

    J.R.Koza.1990. Geneticallybreedingpopulationsofcomputerprogramstosolve problems in artificial intelligence. In[1990] Proceedings of the 2nd International IEEE Conference on Tools for Artificial Intelligence.819–827. doi:10.1109/TAI.1990. 130444

    work page doi:10.1109/tai.1990 1990

  17. [17]

    Mikel Landajuela, Chak Shing Lee, Jiachen Yang, Ruben Glatt, Claudio P San- tiago, Ignacio Aravena, Terrell Mundhenk, Garrett Mulcahy, and Brenden K Petersen. 2022. A Unified Framework for Deep Symbolic Regression. In Advances in Neural Information Processing Systems , S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (Eds.), Vol. 35. Curr...

    2022

  18. [18]

    Petersen, Soo K

    Mikel Landajuela, Brenden K. Petersen, Soo K. Kim, Claudio P. Santiago, Ruben Glatt, T. Nathan Mundhenk, Jacob F. Pettit, and Daniel M. Faissol. 2021. Improv- ing exploration in policy gradient search: Application to symbolic optimization. arXiv:2107.09158 [cs.LG] https://arxiv.org/abs/2107.09158

    Pith/arXiv arXiv 2021

  19. [19]

    Wenqiang Li, Weijun Li, Linjun Sun, Min Wu, Lina Yu, Jingyi Liu, Yanjie Li, and Song Tian. 2023. Transformer-based model for symbolic regression via joint supervised learning. In International Conference on Learning Representa- tions. https://api.semanticscholar.org/CorpusID:259298765

    2023

  20. [20]

    PingchuanMa,Tsun-HsuanWang,MinghaoGuo,ZhiqingSun,JoshuaB.Tenen- baum, Daniela Rus, Chuang Gan, and Wojciech Matusik. 2024. LLM and Simula- tion as Bilevel Optimizers: A New Paradigm to Advance Physical Scientific Dis- covery. InProceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235), Ruslan...

    2024

  21. [21]

    Nour Makke and Sanjay Chawla. 2024. Data-driven discovery of Tsallis-like distribution using symbolic regression in high-energy physics. PNAS Nexus 3, 11 (10 2024), pgae467. arXiv: https://academic.oup.com/pnasnexus/article- pdf/3/11/pgae467/60816181/pgae467.pdf doi:10.1093/pnasnexus/pgae467

    work page doi:10.1093/pnasnexus/pgae467 2024

  22. [22]

    Nour Makke and Sanjay Chawla. 2024. Interpretable scientific discovery with symbolic regression: a review.Artificial Intelligence Review 57 (01 2024).doi:10. 1007/s10462-023-10622-0

    2024

  23. [23]

    Florian R Menter. 1994. Two-equation eddy-viscosity turbulence models for engineering applications. AIAA journal 32, 8 (1994), 1598–1605

    1994

  24. [24]

    Florian R Menter, Martin Kuntz, Robin Langtry, et al. 2003. Ten years of in- dustrial experience with the SST turbulence model.Turbulence, heat and mass transfer 4, 1 (2003), 625–632

    2003

  25. [25]

    Matteo Merler, Katsiaryna Haitsiukevich, Nicola Dainese, and Pekka Marttinen

  26. [26]

    Shinsuke MOCHIZUKI and Hideo OSAKA. 2000. Management of a Stronger Wall Jet by a Pair of Streamwise Vortices. Reynolds Stress Tensor and Produc- tion Tenns.TRANSACTIONS OF THE JAPAN SOCIETY OF MECHANICAL ENGI- NEERS Series B 66, 646 (2000), 1309–1317.doi:10.1299/kikaib.66.646_1309

    work page doi:10.1299/kikaib.66.646_1309 2000

  27. [27]

    Monkewitz

    Peter A. Monkewitz. 2021. Asymptotics of streamwise Reynolds stress in wall turbulence. Journal of Fluid Mechanics931(Nov.2021). doi:10.1017/jfm.2021.924

    work page doi:10.1017/jfm.2021.924 2021

  28. [28]

    Jacques Monod. 1949. THE GROWTH OF BACTERIAL CULTURES. Annual Review of Microbiology 3, Volume 3, 1949 (1949), 371–394.doi:10.1146/annurev. mi.03.100149.002103

    arXiv 1949

  29. [29]

    Nathan Mundhenk, Mikel Landajuela, Ruben Glatt, Claudio P

    T. Nathan Mundhenk, Mikel Landajuela, Ruben Glatt, Claudio P. San- tiago, Daniel M. Faissol, and Brenden K. Petersen. 2021. Symbolic Regression via Neural-Guided Genetic Programming Population Seeding. arXiv:2111.00053 [cs.NE] https://arxiv.org/abs/2111.00053

    Pith/arXiv arXiv 2021

  30. [30]

    Foster, and Michael Hicks

    Iulian Neamtiu, Jeffrey S. Foster, and Michael Hicks. 2005. Understanding source code evolution using abstract syntax tree matching. InProceedings of the 2005 International Workshop on Mining Software Repositories (St. Louis, Mis- souri) (MSR ’05). Association for Computing Machinery, New York, NY, USA, 1–5. doi:10.1145/1083142.1083143

    arXiv 2005

  31. [31]

    Petersen, Mikel Landajuela, T

    Brenden K. Petersen, Mikel Landajuela, T. Nathan Mundhenk, Claudio P. San- tiago, Soo K. Kim, and Joanne T. Kim. 2021. Deep symbolic regression: Re- covering mathematical expressions from data via risk-seeking policy gradients. arXiv:1912.04871 [cs.LG] https://arxiv.org/abs/1912.04871

    Pith/arXiv arXiv 2021

  32. [32]

    Stephen B. Pope. 2000. Turbulent Flows. Cambridge University Press. doi:10. 1017/cbo9780511840531

    2000

  33. [33]

    Julien Pourcel, Cédric Colas, and Pierre-Yves Oudeyer. 2025. Self-Improving Language Models for Evolutionary Program Synthesis: A Case Study on ARC- AGI. arXiv:2507.14172 [cs.LG] https://arxiv.org/abs/2507.14172

    arXiv 2025

  34. [34]

    Julia Reuter, Hani Elmestikawy, Fabien Evrard, Sanaz Mostaghim, and Berend van Wachem. 2023. Graph Networks as Inductive Bias for Genetic Program- ming: Symbolic Models for Particle-Laden Flows. In Genetic Programming , Gisele Pappa, Mario Giacobini, and Zdenek Vasicek (Eds.). Springer Nature Switzerland, Cham, 36–51

    2023

  35. [35]

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M Pawan Kumar, Emilien Dupont, Francisco JR Ruiz, Jordan S El- lenberg, Pengming Wang, Omar Fawzi, et al. 2024. Mathematical discoveries from program search with large language models.Nature 625, 7995 (2024), 468– 475

    2024

  36. [36]

    L Rosso, J. R. Lobry, S Bajard, and J. P. Flandrois. 1995. Convenient Model To Describe the Combined Effects of Temperature and pH on Microbial Growth. Applied and Environmental Microbiology 61, 2 (Feb 1995), 610–6. doi:10.1128/ aem.61.2.610-616.1995

    1995

  37. [37]

    Michael Schmidt and Hod Lipson. 2009. Distilling Free-Form Nat- ural Laws from Experimental Data. Science 324, 5923 (2009), 81–

    2009

  38. [38]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Lan- guage Models. arXiv:2402.03300 [cs.CL] https://arxiv.org/abs/2402.03300

    Pith/arXiv arXiv 2024

  39. [39]

    AlphaForge:AFrameworktoMineandDynamically Combine Formulaic Alpha Factors

    Hao Shi, Weili Song, Xinting Zhang, Jiahe Shi, Cuicui Luo, Xiang Ao, Hamid Arian,andLuisSeco.2024. AlphaForge:AFrameworktoMineandDynamically Combine Formulaic Alpha Factors. arXiv:2406.18394 [q-fin.CP] https://arxiv. org/abs/2406.18394

    Pith/arXiv arXiv 2024

  40. [40]

    Parshin Shojaee, Kazem Meidani, Shashank Gupta, Amir Barati Farimani, and Chandan K Reddy. 2025. LLM-SR: Scientific Equation Discovery via Program- ming with Large Language Models. arXiv:2404.18400 [cs.LG] https://arxiv.org/ abs/2404.18400

    Pith/arXiv arXiv 2025

  41. [41]

    Parshin Shojaee, Ngoc-Hieu Nguyen, Kazem Meidani, Amir Barati Fari- mani, Khoa D Doan, and Chandan K Reddy. 2025. LLM-SRBench: A New Benchmark for Scientific Equation Discovery with Large Language Models. arXiv:2504.10415 [cs.CL] https://arxiv.org/abs/2504.10415

    Pith/arXiv arXiv 2025

  42. [43]

    Discovering explicit Reynolds-averaged turbulence closures for turbulent separated flows through deep learning-based symbolic regression with non-linear corrections

    HongweiTang,YanWang,TongguangWang,andLinlinTian.2023. Discovering explicit Reynolds-averaged turbulence closures for turbulent separated flows through deep learning-based symbolic regression with non-linear corrections. Physics of Fluids 35, 2 (Feb. 2023).doi:10.1063/5.0135638

    doi:10.1063/5.0135638 2023

  43. [44]

    A First Course in Turbulence

    HenkTennekesandJohnL.Lumley.1972. A First Course in Turbulence. TheMIT Press. doi:10.7551/mitpress/3014.001.0001

    work page doi:10.7551/mitpress/3014.001.0001 1972

  44. [45]

    Mojtaba Valipour, Bowen You, Maysum Panju, and Ali Ghodsi. 2021. SymbolicGPT: A Generative Transformer Model for Symbolic Regression. arXiv:2106.14131 [cs.LG] https://arxiv.org/abs/2106.14131

    Pith/arXiv arXiv 2021

  45. [46]

    Martin Vastl, Jonáš Kulhánek, Jiří Kubalík, Erik Derner, and Robert Babuška

  46. [47]

    Marco Virgolin and Solon P. Pissis. 2022. Symbolic Regression is NP-hard. arXiv:2207.01018 [cs.NE] https://arxiv.org/abs/2207.01018

    Pith/arXiv arXiv 2022

  47. [48]

    Ylva Wahlquist, Jesper Sundell, and Kristian Soltesz. 2024. Learning pharmaco- metric covariate model structures with symbolic regression networks.Journal of Pharmacokinetics and Pharmacodynamics 51, 2 (2024), 155–167.doi:10.1007/ s10928-023-09887-3

    2024

  48. [49]

    IEEE Access 12 (2024), 37840–37849

    SymFormer: End-to-End Symbolic Regression Using Transformer-Based Architecture. IEEE Access 12 (2024), 37840–37849. doi:10.1109/ACCESS.2024. 3374649

    doi:10.1109/access.2024 2024

  49. [50]

    Weller, Gavin Tabor, Hrvoje Jasak, and Christer Fureby

    H.G. Weller, Gavin Tabor, Hrvoje Jasak, and Christer Fureby. 1998. A Tenso- rialApproachtoComputationalContinuumMechanicsUsingObjectOrientated Techniques. Computers in Physics 12 (11 1998), 620–631.doi:10.1063/1.168744

    doi:10.1063/1.168744 1998

  50. [51]

    Shijie Xia, Yuhan Sun, and Pengfei Liu. 2025. SR-Scientist: Scientific Equation DiscoveryWithAgenticAI. arXiv: 2510.11661[cs.AI] https://arxiv.org/abs/2510. 11661

    arXiv 2025

  51. [52]

    Wei WANG, Yang ZHANG, and Lili CHEN. 2019. An application of the shear stresstransportlow-Reynolds-numberk-ε turbulencemodelon turbulentflows. ACTA AERODYNAMICA SINICA 37, 3 (2019), 419–425. doi:10.7638/kqdlxxb- 2016.0158

    arXiv 2019

  52. [53]

    Hengzhe Zhang, Qi Chen, Bing XUE, Wolfgang Banzhaf, and Mengjie Zhang

  53. [55]

    Heng Xiao, Jin-Long Wu, Sylvain Laizet, and Lian Duan. 2020. Flows over periodic hills of parameterized geometries: A dataset for data-driven turbu- lence modeling from direct simulations.Computers & Fluids 200 (2020), 104431. doi:10.1016/j.compfluid.2020.104431

    arXiv 2020

  54. [85]

    arXiv:https://www.science.org/doi/pdf/10.1126/science.1165893doi:10.1126/ science.1165893

    work page doi:10.1126/science.1165893doi:10.1126/

  55. [2024]

    Associa- tion for Computational Linguistics, 589–606.doi:10.18653/v1/2024.acl-srw.49

    In-Context Symbolic Regression: Leveraging Large Language Models for FunctionDiscovery.In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop) . Associa- tion for Computational Linguistics, 589–606.doi:10.18653/v1/2024.acl-srw.49

    doi:10.18653/v1/2024.acl-srw.49 2024

  56. [2025]

    error-term elimination

    RAG-SR: Retrieval-Augmented Generation for Neural Symbolic Regres- sion. In The Thirteenth International Conference on Learning Representations . https://openreview.net/forum?id=NdHka08uWn A Additional Hyperparameters of PiT-PO Table 3 lists the additional hyperparameters of PiT-PO, including (i) general-level physical penalties (dimensional homogeneity a...

    2000