Pith. sign in

REVIEW 3 major objections 4 minor 60 references

Bayesian Symbolic Regression with Entropic Reinforcement Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A neural policy can be trained to sample expressions from the Bayesian posterior

desk verdict A useful GFlowNet environment for symbolic regression, but the abstract overstates the posterior predictive performance and the coverage guarantee is under-argued. read the letter →

arxiv 2608.09617 v2 pith:BZK4TUSK submitted 2026-08-10 cs.LG

Oussama Boussif , Mohammed Mahfoud , Younesse Kaddar , Moksh Jain , Sida Li , Damiano Fornasiere , Xiaoyin Chen , Yoshua Bengio
show 1 more author
Esmeralda S. Whitammer
This is my paper · ORCID
classification cs.LG
keywords Bayesiansymbolicregressionmaximum-entropyreinforcementlearningtrajectorybalancegenerativeflownetworksposteriorsamplingexpressiontreesreversePolishnotationdimensionalanalysis
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

ERRLESS aims to solve Bayesian symbolic regression by training a neural policy to generate expression trees token by token, so that at convergence the policy's samples are draws from the posterior distribution over expressions, their constants, and the noise level, given a dataset. If true, this gives a practical way to quantify uncertainty about recovered scientific laws instead of returning a single best-fitting expression, which matters when data are scarce or noisy. The paper shows the method is competitive on the Feynman benchmark while producing short expressions, and that averaging over posterior samples predicts better than an SMC baseline in the noisiest synthetic settings. The central claim is that maximum-entropy reinforcement learning with the trajectory balance objective turns posterior sampling into a policy-learning problem.

What carries the argument

The load-bearing machinery is the trajectory balance objective, a flow-network consistency loss that minimizes the square of $\log Z_\varphi + \log \pi_\varphi(T,\theta,\sigma) - R(T,\theta,\sigma)$ for each complete tuple, with $R$ the unnormalized log posterior. A zero loss at every state forces the policy to sample exactly from the posterior and gives $\log Z_\varphi$ as the log normalizing constant, i.e., the log evidence for the data. This is supported by a bottom-up generation process: expressions are built in postorder, so partial sequences are already evaluable trees and physical-unit constraints can be enforced locally; a transformer encoder supplies next-token logits, a Gaussian-mixture head for constants, and a log-normal-mixture head for the noise scale. Off-policy training with an annealed $\epsilon$-greedy behavior policy and a prioritized replay buffer is what makes the objective tractable in practice.

What would settle it

On a tiny one-variable problem with a known posterior, enumerate every valid expression up to the node limit, compute each one's posterior probability exactly, and compare this exact posterior with the empirical distribution of ERRLESS samples; if any expression with substantial posterior mass is systematically missing or badly over-sampled, the convergence claim is false.

Watch

Extended reading notes

Core claim

The central claim is that maximizing expected log-posterior reward plus policy entropy yields a policy that at convergence samples exactly from the posterior $p(T,\theta,\sigma|D)$. The paper defines $R(T,\theta,\sigma)$ as the log prior plus log likelihood, so $\exp(R)$ is the unnormalized posterior, and trains the policy with the trajectory balance objective (5), whose zero occurs exactly when $\log \pi_\varphi + \log Z_\varphi$ equals $R$ for every complete sample. Since the objective can be driven to zero simultaneously for all states, the paper uses off-policy data and a transformer policy that emits expressions in reverse Polish notation, with separate heads for constants and noise. Empirically, ERRLESS reaches an AUC of 0.924 on the Feynman database and, on synthetic problems, its posterior predictive mean beats the SMC baseline PySIPS under the noisiest setting, while remaining finite where PySIPS's predictive mean collapses.

Load-bearing premise

The proof that the trained policy samples the posterior assumes training data cover every expression the posterior might favour; the replay buffer only keeps high-reward samples, so expressions with lower reward but real posterior probability may never be trained on, and the loss cannot reach zero there.

Editorial extensions

If this is right

  • Once trained, a policy produces posterior samples without the inner constant-optimization loop used by most competitors, which the paper identifies as the main speed bottleneck and reports as an order-of-magnitude runtime advantage.
  • Posterior predictive intervals over expressions become available, letting practitioners see where the recovered formula is uncertain and extrapolate more reliably when training data are sparse.
  • Because the sampler represents the full posterior rather than a single optimum, domain priors and unit constraints can be changed without changing the learning algorithm, only the reward.
  • The trajectory balance estimate of $\log Z$ gives a model-evidence estimate, enabling Bayesian model comparison across datasets or operator libraries.

Reading between the lines

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

  • The paper leaves unexplored a direct calibration check, but a reader could enumerate the full expression space on a tiny problem and compare the policy's samples to the exact posterior; nothing in the current experiments rules out a systematic coverage gap.
  • The same trajectory-balance reward identity applies to any countable structure with a computable prior and likelihood, so the method should transfer to symbolic integration, unit-carrying dynamical systems, or grammar-constrained design, though the paper only mentions differential equations as future work.
  • Because the operator unigram prior is fitted to a physics corpus and visibly shapes the recovered expressions, one could test how sensitive the posterior is to that corpus by running ERRLESS with a held-out prior; this would separate inductive bias from likelihood information.
  • If the empirically observed speedup holds on larger benchmarks, amortized Bayesian symbolic regression could make full posterior uncertainty practical in scientific discovery pipelines, where current MCMC and SMC samplers become too costly.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes ERRLESS, an amortized Bayesian symbolic regression method based on maximum-entropy reinforcement learning. Expressions are generated bottom-up as postorder token sequences with dimensional and structural constraints, and a transformer policy with Gaussian-mixture heads outputs expression trees, constant assignments, and noise variances. The policy is trained with the trajectory balance objective (Eq. 5), which the paper argues makes the policy sample from the Bayesian posterior over expressions, parameters, and noise at convergence. Experiments are reported on a seven-expression synthetic benchmark, the Feynman Symbolic Regression Database, and a Blackbox benchmark, with claims of competitive accuracy, short expressions, and posterior predictive performance against the SMC baseline PySIPS.

Significance. If the central posterior-sampling claim were established, ERRLESS would be a valuable contribution: it would provide an amortized alternative to MCMC/SMC Bayesian symbolic regression, avoid per-candidate constant fitting, and enable uncertainty quantification over expression structures. The paper contains several genuinely useful elements: a bottom-up generation process that enforces physical-unit constraints, a clean formulation of symbolic regression as an entropy-regularized RL problem, a public code release, and ablation studies of the prior, training objective, and replay mixing coefficient. These strengths make the approach worth developing. However, the key theoretical guarantee is not established for the actual training procedure, and the main empirical evidence in Table 1 contradicts the abstract's "high R^2" claim. The contribution is therefore promising but needs substantial revision.

major comments (3)
  1. [§3.2, §B.2, §B.3] The central claim that, at convergence, the ERRLESS policy samples from p(T,θ,σ|D) is not justified by the training procedure described. The convergence statement in §3.2 invokes Theorem 3.1 of Gritsaev et al. under an on-policy estimate of the entropy-regularized objective (4), but the actual algorithm is off-policy: §B.3 uses ε-greedy exploration, and §B.2 filters the replay buffer by discarding every newly generated trajectory whose reward is below the current buffer minimum. For any fixed buffer B, the TB loss (5) can be driven to zero by setting the policy proportional to exp(R) on B and assigning arbitrary residual mass to trajectories outside B, so zero training loss does not imply that the policy equals the full posterior. Because the Gaussian likelihood and the priors are positive on the full space of valid expressions, the true posterior has full support on that space, and the truncation is not a harmless approximation. The paper provides no diagnostic that would detect this failure, such as an importance-weighted effective sample size of policy samples relative to the posterior, or a comparison of the learned log Z with an SMC estimate of the marginal likelihood. This is a load-bearing gap: the abstract's posterior-sampling claim is the method's raison d'être.
  2. [Abstract, §6.1, Table 1] The abstract's statement that "the mean of the posterior predictive approximated by ERRLESS achieves a high coefficient of determination (R^2)" is contradicted by the paper's own Table 1. On the synthetic benchmark, ERRLESS's posterior predictive R^2_PP is negative on five of seven expressions (e.g., -0.232 for 1/sqrt(1-x^2), -5.708 for sin(y+2.5x), -11.055 for sin(x)cos(y), -9.387 for sqrt(x^2+y^2), and -2.230 for x+sin(5.5x)). A negative R^2 means the prediction is worse than predicting the test-set mean, which is not a "high" coefficient of determination. The comparison with PySIPS shows only that ERRLESS's posterior predictive mean is less catastrophically bad than PySIPS's on some problems, not that it is accurate. The empirical claims in the abstract and §6.1 need to be rewritten to match the reported numbers.
  3. [§6.1, §A.2, Table 1] The negative log-likelihood comparison in Table 1 is not apples-to-apples. As stated in §A.2, for ERRLESS the σ_j in the pointwise mixture NLL is each posterior draw's sampled noise value, while for PySIPS it is the j-th particle's root-mean-square residual on the training split. These are different quantities, and the resulting NLL values should not be compared as if they evaluated the same predictive distribution. The paper should either compute a common posterior-predictive density for both methods or explicitly discuss why the different σ definitions do not bias the comparison. As it stands, the claim in §6.1 that ERRLESS "models the posterior over expression trees more accurately" is not supported by the NLL column.
minor comments (4)
  1. [§D.3] The ablation in Table 8 shows that fixed replay-buffer mixing coefficients of 0.4, 0.5, 0.7, and 0.8 outperform the annealed schedule used in the main results, yet the main experiments use an annealing schedule from 0.9 to 0.2. The paper should explain why the annealed schedule was retained despite this ablation evidence, or update the choice.
  2. [§3.2] The sentence "Because (5) can be minimized to 0 for all samples simultaneously" is imprecise: this is only true when the behavior distribution covers the full support of the target distribution. The coverage assumption should be stated explicitly and connected to the replay-buffer construction.
  3. [§6.1] The text notes that "ERRLESS is less prone to this failure: its posterior draws yield a finite predictive mean on every problem." Finiteness of the predictive mean is a very weak guarantee and does not address the negative R^2 values reported in Table 1; the discussion should be revised to acknowledge that a finite mean can still be a poor prediction.
  4. [§5.2] The definition of R^2 in the metrics section does not explicitly state whether predictions are evaluated on the test set or the training set. In Table 1 the column "Test R^2" makes this clear, but the metric definition could be more explicit for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

Derivation is self-contained: the posterior target is fixed by Bayes rule, the reward is its log-density, and the TB loss is minimized to match it; no fitted parameter is renamed as a prediction.

full rationale

The paper's central claim, that at convergence the policy samples from the posterior p(T,θ,σ|D), is a direct consequence of the definitions, not a circular reduction. The posterior is defined independently in Eq. (2) from a prior and a Gaussian likelihood; the reward R is the log unnormalized posterior; and the trajectory-balance objective (5) is constructed so that a zero residual enforces π(T,θ,σ) ∝ exp(R). This is a training objective designed to match an independently specified target, not a fitted parameter being presented as a prediction. The convergence guarantee is attributed to an external theorem (Gritsaev et al., 2025), not to a self-citation chain, and the paper's self-citations to GFlowNet and TB work (Bengio et al., Malkin et al., Deleu et al., Mahfoud et al.) are used as standard algorithmic building blocks rather than as the source of the predictive target. The unigram operator prior is fit to an external physics corpus (Encyclopaedia Inflationaris, following Constantin et al.), not to the regression datasets, so any effect on Feynman benchmark performance is prior elicitation rather than circular fitting; no equation equates a reported result to a constant fitted from that result. The main caveat is not circularity but an unverified coverage condition: §B.2's high-reward replay filtering restricts training to trajectories with reward above the current buffer minimum, so driving the squared TB loss to zero on that buffer need not imply posterior sampling over the full support, and §3.2's 'assuming convergence' is an unresolved correctness assumption rather than a tautology. Because no prediction reduces by construction to an input or to a self-citation, the appropriate finding is no significant circularity.

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

The method rests on standard GFlowNet convergence theory, an external-corpus prior, and several hand-set priors and constraints. The main unexamined premise is the applicability of the TB convergence guarantee to the biased off-policy replay-buffer training actually used. No new entities are postulated.

free parameters (4)
  • Unigram prior frequencies = From Table 3 (e.g., × 0.1770, / 0.1328, variables 0.2877, constants 0.1892)
    Used in the reward R(T,θ,σ) (§A.2) and derived from operator frequencies in Encyclopaedia Inflationaris; fit to an external corpus, not the target data, but a hand-picked corpus that overlaps with the Feynman physics domain.
  • Constant prior standard deviation = 10 (synthetic), 20 (Feynman/Blackbox)
    Chosen per dataset (§A.2); affects the posterior over θ and thus the samples.
  • Noise prior = LogNormal(0,5) for synthetic, except N(-1,5^2) for sqrt(x^2+y^2); HalfNormal(2000) for Feynman/Blackbox
    Hand-specified per dataset and per expression (§A.2); the per-expression exception is an ad hoc choice.
  • Maximum nodes and constants = L=9, K=3 (synthetic); L=32, K=3 (Feynman)
    Hand-set search-space constraints (§5.1) that bound the support of the posterior.
assumptions (4)
  • domain assumption Convergence of trajectory balance to the posterior (Theorem 3.1 of Gritsaev et al., 2025) holds under the paper's off-policy training scheme.
    Invoked in §3.2 but conditions (on-policy estimate convergence) are not verified; training uses replay buffer and ε-greedy.
  • domain assumption Every partial sequence can be continued to a complete expression (stated in §3.1).
    Needed for the autoregressive factorization (3) to define a distribution over all trees; holds for the chosen operator library but is asserted without proof.
  • domain assumption The reward R(T,θ,σ) is finite and computable for every sampled (T,θ,σ); undefined evaluations (e.g., log of a negative number, division by zero) are handled without breaking training.
    The likelihood in (2) requires f_{T,θ}(x_i) to be defined on all training points; the paper does not specify how undefined expressions are rewarded.
  • standard math The unigram prior is a proper probability distribution over the finite set of reachable trees.
    True because maximum nodes is bounded (§D.1), so every positive weight can be normalized; explicitly stated in §D.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bayesian Symbolic Regression with Entropic Reinforcement Learning." pith.science (2026). https://pith.science/paper/BZK4TUSK

@misc{pith2026260809617,
  author       = {Pith},
  title        = {Pith review of: Bayesian Symbolic Regression with Entropic Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BZK4TUSK}},
  note         = {Machine review of arXiv:2608.09617}
}
abstract

Symbolic regression is the problem of finding an algebraic expression describing a stochastic dependence of a target variable on a set of inputs. Unlike forms of regression that fit parameters assuming a fixed model structure, symbolic regression is a search problem over the space of expressions, represented, for example, as abstract syntax trees using a library of operators. Symbolic regression is typically used in settings with limited, noisy data in the natural sciences. However, searching for a single best-fitting expression fails to capture the epistemic uncertainty about the expression, which motivates a Bayesian perspective that enables uncertainty quantification and specification of natural priors to constrain the search space. In this work, we propose ERRLESS (Entropy-Regularized Reinforcement Learning for Expression Structure Sampling), a scalable approach for sampling from the posterior distribution over expressions given data using maximum-entropy reinforcement learning. ERRLESS learns a neural policy that constructs expressions sequentially by building up their abstract syntax trees. At convergence, the policy samples expressions from the posterior. At test time, expressions can be sampled by rollouts of this policy. We demonstrate that ERRLESS achieves competitive results on the Feynman benchmark while producing short and interpretable expressions. Additionally, we demonstrate that the mean of the posterior predictive approximated by ERRLESS achieves a high coefficient of determination ($R^2$) compared to an SMC baseline, highlighting the benefits of the Bayesian perspective in symbolic regression.

Figures

Figures reproduced from arXiv: 2608.09617 by the authors.

Figure 1
Figure 1. Generation process for the expression 𝐸 = 𝑚𝑐2 . Valid states are shown in gray, while invalid states have pink fill and red outlines. Black arrows indicate valid transitions, and red arrows indicate invalid transitions. Leaf nodes are represented by double-outlined circles, unary operators by single-outlined squares, and binary operators by double-outlined squares. Wiggly arrows denote states that can terminate and … view at source ↗
Figure 2
Figure 2. Growth of search space size in maximum number [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Samples from the posterior. ERRLESS and PySIPS are trained on noised values (white circles) of the ground￾truth function 𝑥 + sin(5.5𝑥) (solid black line) at points in the training domain [−0.5, 0.5] (beige). The posterior mean (dashed blue line), that is, the mean of the individual posterior samples, fits the true function well on the beige interval in both cases, but ERRLESS extrapolates better outside the training… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Performance analysis on the Feynman benchmark. (a) Robustness profile showing the AUC of the empirical success rate P[𝑅 2 ≥ 𝑡] across varying noise levels (𝛾 ∈ {0.001, 0.01, 0.1}). Algorithms are ranked by mean AUC, with ERRLESS demonstrating superior stability. (b) Pa…
Figure 5
Figure 5. Figure 5: Exploratory analysis of the Feynman Symbolic Regression Database. Left: histogram of the number of variables per expression. Center: distribution of expression tree lengths. Right: proportion of expressions in which all variables are associated with physical units. Bla…
Figure 6
Figure 6. Figure 6: Performance analysis on the Blackbox benchmark. (a) Area Under the Curve (AUC) of the empirical success rate P[𝑅 2 ≥ 𝑡]. Algorithms are ranked by mean AUC. (b) Pareto front illustrating the trade-off between median AUC score and model parameter count. ERRLESS achieves …
Figure 7
Figure 7. Figure 7: Computational runtime. Average of the time (in seconds) taken by all methods over random seeds and expressions in the Blackbox benchmark. E COMPUTATIONAL RUNTIME [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 1 linked inside Pith

  1. [1]

    Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks , title =

    La Cava, William and Orzechowski, Patryk and Burlacu, Bogdan and de Fran. Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks , title =

  2. [2]

    , title =

    Tenachi, Wassim and Ibata, Rodrigo and Diakogiannis, Foivos I. , title =. The Astrophysical Journal , year =

  3. [3]

    Inference of compact nonlinear dynamic models by epigenetic local search , journal =

    William. Inference of compact nonlinear dynamic models by epigenetic local search , journal =. 2016 , url =

    2016

  4. [4]

    Romano and Trang T

    Joseph D. Romano and Trang T. Le and William. Bioinformatics , volume=

  5. [5]

    Science Advances , publisher =

    Udrescu, Silviu-Marian and Tegmark, Max , year =. Science Advances , publisher =

  6. [6]

    and Leighton, Robert B

    Feynman, Richard P. and Leighton, Robert B. and Sands, Matthew , year=. The Feynman lectures on physics, vol. I: The new millennium edition: Mainly mechanics, radiation, and heat , publisher=

  7. [7]

    Gomez and Lukasz Kaiser and Illia Polosukhin , title =

    Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin , title =. Neural Information Processing Systems (NeurIPS) , year =

  8. [8]

    Petersen and Soo K

    Mikel Landajuela and Brenden K. Petersen and Soo K. Kim and Cl. Improving exploration in policy gradient search: Application to symbolic optimization , journal =

Show all 60 references
  1. [9]

    and Miranda, Manuel and Pallarès, Jordi and Sales-Pardo, Marta , year =

    Guimerà, Roger and Reichardt, Ignasi and Aguilar-Mogas, Antoni and Massucci, Francesco A. and Miranda, Manuel and Pallarès, Jordi and Sales-Pardo, Marta , year =. A Bayesian machine scientist to aid in the solution of challenging scientific problems , volume =. Science Advance...

    doi

  2. [10]

    2019 , journal =

    Bayesian Symbolic Regression , author =. 2019 , journal =

    2019

  3. [11]

    International Conference on Learning Representations (ICLR) , year=

    Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients , author=. International Conference on Learning Representations (ICLR) , year=

  4. [12]

    Neural Information Processing Systems (NeurIPS) , year=

    Symbolic regression via neural-guided genetic programming population seeding , author=. Neural Information Processing Systems (NeurIPS) , year=

  5. [13]

    International Conference on Machine Learning (ICML) , year=

    Neural symbolic regression that scales , author=. International Conference on Machine Learning (ICML) , year=

  6. [14]

    Deep generative symbolic regression with

    Kamienny, Pierre-Alexandre and Lample, Guillaume and Lamprier, Sylvain and Virgolin, Marco , journal=. Deep generative symbolic regression with

  7. [15]

    International Semantic Web Conference , pages=

    SRBench: a streaming RDF/SPARQL benchmark , author=. International Semantic Web Conference , pages=. 2012 , organization=

    2012

  8. [16]

    Reversible jump

    Green, Peter J , journal=. Reversible jump. 1995 , publisher=

    1995

  9. [17]

    Philosophical Transactions of the Royal Society A , volume=

    Bayesian Symbolic Regression via Posterior Sampling , author=. Philosophical Transactions of the Royal Society A , volume=. 2026 , publisher=

    2026

  10. [18]

    Philosophical Transactions of the Royal Society A , volume=

    Bayesian symbolic regression: Automated equation discovery from a physicist's perspective , author=. Philosophical Transactions of the Royal Society A , volume=. 2026 , publisher=

    2026

  11. [19]

    Bengio, Yoshua and Lahlou, Salem and Deleu, Tristan and Hu, Edward J and Tiwari, Mo and Bengio, Emmanuel , journal=

  12. [20]

    Neural Information Processing Systems (NeurIPS) , year=

    Flow network based generative models for non-iterative diverse candidate generation , author=. Neural Information Processing Systems (NeurIPS) , year=

  13. [21]

    Trajectory balance: Improved credit assignment in

    Malkin, Nikolay and Jain, Moksh and Bengio, Emmanuel and Sun, Chen and Bengio, Yoshua , journal=. Trajectory balance: Improved credit assignment in

  14. [22]

    Li, Sida and Marinescu, Ioana and Musslick, Sebastian , journal=

  15. [23]

    Uncertainty in Artificial Intelligence (UAI) , year=

    Bayesian structure learning with generative flow networks , author=. Uncertainty in Artificial Intelligence (UAI) , year=

  16. [24]

    Proceedings of the 42nd ACM SIGPLAN International Conference on Programming Language Design and Implementation , pages=

    Ellis, Kevin and Wong, Catherine and Nye, Maxwell and Sabl. Proceedings of the 42nd ACM SIGPLAN International Conference on Programming Language Design and Implementation , pages=

  17. [25]

    Neural Information Processing Systems (NeurIPS) , year =

    Bridging the gap between value and policy based reinforcement learning , author =. Neural Information Processing Systems (NeurIPS) , year =

  18. [26]

    Genetic Programming and Evolvable Machines , volume=

    Semantically-based crossover in genetic programming: application to real-valued symbolic regression , author=. Genetic Programming and Evolvable Machines , volume=. 2011 , publisher=

    2011

  19. [27]

    Maximum Entropy

    Benjamin Eysenbach and Sergey Levine , journal=. Maximum Entropy

  20. [28]

    Philosophical Transactions of the Royal Society A , volume=

    Statistical Patterns in the Equations of Physics and the Emergence of a Meta-Law of Nature , author =. Philosophical Transactions of the Royal Society A , volume=. 2026 , publisher=

    2026

  21. [29]

    Deleu, Tristan and Nishikawa-Toomey, Mizu and Subramanian, Jithendaraa and Malkin, Nikolay and Charlin, Laurent and Bengio, Yoshua , journal =. Joint

  22. [30]

    Mingyang Zhou and Zichao Yan and Elliot Layne and Nikolay Malkin and Dinghuai Zhang and Moksh Jain and Mathieu Blanchette and Yoshua Bengio , year=

  23. [31]

    arXiv preprint arXiv:2503.06985 , year=

    Learning Decision Trees as Amortized Structure Inference , author=. arXiv preprint arXiv:2503.06985 , year=

    arXiv

  24. [32]

    Age-Fitness

    Schmidt, Michael and Lipson, Hod , year =. Age-Fitness. Genetic

  25. [33]

    2019 , month = sep, journal =

    A Probabilistic and Multi-Objective Analysis of Lexicase Selection and -Lexicase Selection , author =. 2019 , month = sep, journal =

    2019

  26. [34]

    and Suri, Srinivas and Moore, Jason H

    La Cava, William and Singh, Tilak Raj and Taggart, James P. and Suri, Srinivas and Moore, Jason H. , year =. Learning Concise Representations for Regression by Evolving Networks of Trees , journal =

  27. [35]

    McConaghy, Trent , year =. Genetic

  28. [36]

    2021 , journal =

    Improving Model-Based Genetic Programming for Symbolic Regression of Small Expressions , author =. 2021 , journal =

    2021

  29. [37]

    2021 , journal =

    Interaction-Transformation Evolutionary Algorithm for Symbolic Regression , author =. 2021 , journal =

    2021

  30. [38]

    Multiple Regression Genetic Programming , booktitle =

    Arnaldo, Ignacio and Krawiec, Krzysztof and O'Reilly, Una-May , year =. Multiple Regression Genetic Programming , booktitle =

  31. [39]

    2020 , journal =

    Parameter Identification for Symbolic Regression Using Nonlinear Least Squares , author =. 2020 , journal =

    2020

  32. [40]

    Linear Scaling with and within Semantic Backpropagation-based Genetic Programming for Symbolic Regression , booktitle =

    Virgolin, Marco and Alderliesten, Tanja and Bosman, Peter AN , year =. Linear Scaling with and within Semantic Backpropagation-based Genetic Programming for Symbolic Regression , booktitle =

  33. [41]

    2009 , publisher=

    Distilling free-form natural laws from experimental data , author=. 2009 , publisher=

    2009

  34. [42]

    Proceedings of the National Academy of Sciences , volume=

    Automated reverse engineering of nonlinear dynamical systems , author=. Proceedings of the National Academy of Sciences , volume=. 2007 , publisher=

    2007

  35. [43]

    Approximate inference in discrete distributions with

    Buesing, Lars and Heess, Nicolas and Weber, Theophane , journal=. Approximate inference in discrete distributions with

  36. [44]

    Uncertainty in Artificial Intelligence (UAI) , year=

    Discrete probabilistic inference as control in multi-path environments , author=. Uncertainty in Artificial Intelligence (UAI) , year=

  37. [45]

    Maximum entropy

    Mohammadpour, Sobhan and Bengio, Emmanuel and Frejinger, Emma and Bacon, Pierre-Luc , journal=. Maximum entropy

  38. [46]

    Generative flow networks as entropy-regularized

    Tiapkin, Daniil and Morozov, Nikita and Naumov, Alexey and Vetrov, Dmitry P , journal=. Generative flow networks as entropy-regularized

  39. [47]

    International Conference on Machine Learning (ICML) , year=

    Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor , author=. International Conference on Machine Learning (ICML) , year=

  40. [48]

    Elementy logiki matematycznej i teorii mnogo\'

    S. Elementy logiki matematycznej i teorii mnogo\'. 1963 , note =

    1963

  41. [49]

    GitHub repository , howpublished =

    Stephens, Trevor , title =. GitHub repository , howpublished =. 2015 , publisher =

    2015

  42. [50]

    ECML PKDD Workshop: Languages for Data Mining and Machine Learning , year =

    Lars Buitinck and Gilles Louppe and Mathieu Blondel and Fabian Pedregosa and Andreas Mueller and Olivier Grisel and Vlad Niculae and Peter Prettenhofer and Alexandre Gramfort and Jaques Grobler and Robert Layton and Jake VanderPlas and Arnaud Joly and Brian Holt and Ga. ECML P...

  43. [51]

    Kingma and Jimmy Ba , title =

    Diederik P. Kingma and Jimmy Ba , title =. International Conference on Learning Representations (ICLR) , year =

  44. [52]

    1929 , pages =

    Łukasiewicz, Jan , title =. 1929 , pages =

    1929

  45. [53]

    International Conference on Machine Learning (ICML) , year=

    Noisy natural gradient as variational inference , author=. International Conference on Machine Learning (ICML) , year=

  46. [54]

    How good is the

    Wenzel, Florian and Roth, Kevin and Veeling, Bastiaan S and. How good is the. International Conference on Machine Learning (ICML) , year=

  47. [55]

    Imai Aldeia and Hengzhe Zhang and Geoffrey Bomarito and Miles Cranmer and Alcides Fonseca and Bogdan Burlacu and William G

    Guilherme S. Imai Aldeia and Hengzhe Zhang and Geoffrey Bomarito and Miles Cranmer and Alcides Fonseca and Bogdan Burlacu and William G. Call for. Proceedings of the Genetic and Evolutionary Computation Conference Companion (GECCO) , pages =. 2025 , publisher =

    2025

  48. [56]

    2026 , journal =

    Complexity-Aware Deep Symbolic Regression with Robust Risk-Seeking Policy Gradients , author =. 2026 , journal =

    2026

  49. [57]

    Fangzheng Sun and Yang Liu and Jian-Xun Wang and Hao Sun , journal =

  50. [58]

    Interpretable Machine Learning for Science with

    Cranmer, Miles , journal =. Interpretable Machine Learning for Science with

  51. [59]

    Secrets of

    Tianshu Yu , year =. Secrets of

  52. [60]

    Optimizing Backward Policies in

    Gritsaev, Timofei and Morozov, Nikita and Samsonov, Sergey and Tiapkin, Daniil , journal =. Optimizing Backward Policies in

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.