Pith. sign in

REVIEW 3 major objections 5 minor 33 references

SoftPipe: A Soft-Guided Reinforcement Learning Framework for Automated Data Preparation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read SoftPipe replaces hard constraints in data-preparation reinforcement learning with a Bayesian soft-guidance policy and reports the best average accuracy, 0.835, and best average rank, 1.78, across 18 datasets.

desk verdict The central optimality claim does not survive contact with the paper's own equations: Eq. (6) and Eq. (10) are algebraically inconsistent unless β=1, and Table 6 uses β=2, so the main theorems are unproven for the reported configuration. read the letter →

arxiv 2507.13710 v2 pith:JUNTFYSW submitted 2025-07-18 cs.DB cs.LG

Jing Chang , Chang Liu , Jinbin Huang , Shuyuan Zheng , Rui Mao , Jianbin Qin This is my paper · ORCID
classification cs.DBcs.LG
keywords automateddatapreparationsoftguidancereinforcementlearninglargelanguagemodelpriortorankKL-regularizedpolicypipelinesearchAutoML
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

The paper sets out to show that the hard constraints used by current reinforcement-learning systems for automated data preparation—particularly the rule that an operator type cannot appear twice in one pipeline—are not harmless search shortcuts but structural barriers to optimal pipelines. SoftPipe replaces these constraints with a soft guidance policy that combines three signals: an LLM-generated strategic prior over operator types, a supervised Learning-to-Rank score, and the RL agent's Q-value. The combination is derived as Bayesian inference over action optimality, and the resulting policy is shown to be the closed-form optimum of a KL-regularized RL objective. Across 18 public benchmark datasets, SoftPipe reports the best average accuracy (0.835) and best average rank (1.78), with up to 13.9% higher pipeline quality and 2.8x faster convergence than the baselines. If the claims hold, soft guidance can reach pipelines that hard-constraint methods cannot even represent.

What carries the argument

The load-bearing object is Eq. (6), the SoftPipe policy. It is built from a Bayesian decomposition: the LLM supplies a prior over operator types, marginalized to operators; the exponential of αQ + γ r_LTR supplies the likelihood; applying softmax to the sum of the log-prior and log-likelihood gives the posterior policy. The same expression is recognized as the closed-form solution of the KL-regularized objective in Eq. (7), which is what lets the paper claim optimality and convergence (Theorem 1). Theorem 2's bound uses the LLM prior's positivity on the optimal operator to show SoftPipe keeps a positive selection probability where hard constraints have zero.

What would settle it

Take a dataset whose optimal pipeline repeats an operator type, such as wall-robot-nav, and re-run SoftPipe with the LTR score re-scaled or replaced by noise so it is no longer comparable to Q-values in scale. The bound in Theorem 2 is supposed to guarantee a positive advantage over hard constraints; if the measured advantage vanishes or reverses under this perturbation, the unbiasedness and scale-alignment assumption is carrying the proof.

Watch

Extended reading notes

Core claim

SoftPipe's central claim is that the policy π(a_t|s_t) = softmax(αQ(s_t,a_t) + β log P_LLM(a_t|s_t) + γ r_LTR(a_t|s_t)) is the theoretically grounded answer to pipeline search: it is the posterior of a Bayesian action-optimality model, and equivalently the optimal solution of maximizing expected αQ + γ r_LTR while keeping the policy within KL distance β of the LLM prior. Theorem 1 asserts convergence of this policy under standard RL assumptions, and Theorem 2 gives a lower bound on how much better SoftPipe can do than a hard-constrained policy when the optimal operator is excluded. The empirical claim is that this soft-guidance design consistently outperforms hard-constraint, hierarchical, pure-LLM, and classical AutoML baselines, with the wall-robot-nav case showing a pipeline of [QuantileTransformer, StandardScaler, PolynomialFeatures] that a hard-constraint method cannot access.

Load-bearing premise

The load-bearing premise is that the LTR model's scores are unbiased estimates of action optimality and are on a scale comparable to Q-values; the paper invokes this as 'For simplicity, assume rLTR provides unbiased estimates' in the proof of Theorem 2, and if it fails the claimed lower bound does not follow.

Editorial extensions

If this is right

  • Pipelines that repeat an operator type become reachable; the wall-robot-nav case reaches 0.962 accuracy via the combination [QuantileTransformer, StandardScaler, PolynomialFeatures], above the hard-constrained baseline's 0.946.
  • Warm-starting with the LLM prior and dense LTR scores makes early exploration efficient, yielding 2.8x faster convergence in the reported learning curves.
  • Average accuracy rises to 0.835 and average rank to 1.78 across 18 datasets, outperforming all listed baselines including classical AutoML systems.
  • Generated pipelines average 2.72 operators, shorter than the hard-constraint baseline's 6.00, making them cheaper and more interpretable.
  • Removing any one of the three signals degrades accuracy in the ablation, supporting the three-signal fusion as the source of the gain.

Reading between the lines

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

  • The β parameter in the KL-regularized objective is a natural exploration-commitment dial: larger β pins the policy closer to the LLM prior, while smaller β lets Q and LTR dominate; an empirical study of how β should scale with search-space size is a direct test of the framework's practical reach.
  • Because the optimality-gap bound only requires the LLM prior to be positive on the optimal action, the same soft-guidance recipe should transfer to other AutoML search problems, such as feature selection or hyperparameter schedules, where hard pruning rules are common.
  • An online-trained ranker that replaces the unreleased offline LTR corpus would test whether the warm-start gains come from the scoring function itself or from its pretrained quality; this variant is not in the paper but follows directly from its ablation logic.
  • The theory predicts that improvement should concentrate on datasets whose best pipelines need repeated operator types; sorting the 18 results by whether the winning pipeline repeats a type would be a cheap check of the core mechanism.
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 SoftPipe, an RL framework for automated data preparation in which action selection is a softmax over a weighted combination of a Q-value, a Learning-to-Rank score, and the log of an LLM prior. It claims this policy is the optimal solution of a KL-regularized RL objective, provides convergence and optimality-gap theorems, and reports experiments on 18 datasets showing improvements in average accuracy and rank over hard-constraint baselines. The paper's central narrative is that soft guidance from an LLM prior, combined with empirical estimators, replaces rigid hard constraints and leads to better pipelines.

Significance. The soft-guidance idea is timely and the empirical study is broad: 18 datasets, multiple baselines, ablations, and a hyperparameter analysis. The inclusion of a code link and ablation variants is a positive. However, the theoretical justification is the paper's primary claimed novelty, and the derivations contain a clear algebraic inconsistency that invalidates the claim that Eq. (6) is the optimal solution of Eq. (7). If the empirical results hold up, the approach has practical value, but the theoretical gaps and the unverified LTR assumption currently prevent the central 'theoretically grounded' claim from being accepted.

major comments (3)
  1. [5.2, Eq. (6), (7), (10); Table 6] Equation (6) defines π(a|s) ∝ P_LLM(a|s)^β exp(αQ + γr_LTR), while solving the KL-regularized objective in Eq. (7) for a fixed state gives π*(a|s) ∝ P_LLM(a|s) exp((αQ + γr_LTR)/β), as stated in Eq. (10). These two expressions coincide only when β = 1. Since Table 6 sets β = 2 for SoftPipe, the policy actually used in the experiments is not the optimal solution of the objective the paper claims to solve. The statement in the proof that Eq. (10) 'matches our policy formulation' is therefore incorrect. This is a load-bearing defect because the paper's theoretical grounding rests on this equivalence.
  2. [5.4, Eq. (11)-(17)] The lower bound on π(o*|s) in Eq. (15) does not follow from the logit in Eq. (5)/(6): the logit contains β log P_LLM(o*) and γ r_LTR(o*), but the bound uses P_LLM(o*) exp(αQ*/β) with no exponent β on the prior and no LTR term. Even under the stated 'r_LTR unbiased' simplification, the algebra is inconsistent. Moreover, the unbiasedness and scale-comparability of r_LTR with Q are assumed 'for simplicity' but no evidence is given; Section 5.1.2 says the LTR model is trained on a large offline dataset of past experiences, which is not described or released, so the assumption is unverifiable. Finally, the move from c/(1+e^{-αΔ/β}) in Eq. (16) to c/(1+e^{-Δ/β}) in Eq. (11) by 'absorbing α into β' is not justified, since the constant c defined in Eq. (16) contains P_LLM(o*) and the max over P_LLM(o') and does not have a simple dependence on α. The theorem as stated is not proven.
  3. [5.2, Theorem 1] Theorem 1 is titled 'Convergence of Collaborative Policy', but its proof only observes that Eq. (6) is a softmax and states that as the Q-function converges through RL updates, the policy converges to this optimal form. No convergence analysis of the RL algorithm (Algorithm 2's Monte Carlo Q update) under the stochastic softmax policy is provided. Standard Q-learning convergence results require conditions such as greedy-in-the-limit exploration, which are not verified here. In addition, the trajectory-level objective in Eq. (8) places the KL term inside an expectation over π, which is not the same as the per-state objective in Eq. (7); this change is not discussed. Thus the claimed convergence guarantee is not established.
minor comments (5)
  1. [Throughout] There are numerous typographical errors, such as 'exploreReinforcement learning' (§2), 'constriants' (§1), 'trnsitions' and 'colloaboation' (§4), and 'demostrate' (§2). These should be corrected.
  2. [Figures 2 and 3] Figures 2 and 3 are rendered as long strings of '/uni...' codes in the manuscript, making the learning curves and inference-time plots illegible. The figures must be replaced with actual plots.
  3. [Table 3, Section 6.2] SoftPipe does not beat the best baseline on several datasets (e.g., abalone, connect-4, page-blocks, pol, run_or_walk, uscensus). The claim of 'consistent superiority' is too strong as stated and should be moderated or supported by pairwise significance tests or effect sizes.
  4. [Section 5.1.2] The offline dataset used to train the LTR model is not described: its size, sources, collection procedure, and release status are all absent. Since the LTR model is a key component of the method, this is an important reproducibility gap.
  5. [Appendix A.4] The hyperparameter analysis reports multiple local maxima for β and γ, but the selection criterion for the final values (β=2, γ=2) is not stated explicitly. Please specify the exact rule used to pick these values.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the policy is an explicit construction and the empirical comparison is against external benchmarks; the proof of Theorem 1 is algebraically inconsistent for β=2, but that is a correctness flaw, not an input-output equivalence.

full rationale

No load-bearing step in SoftPipe reduces to its own inputs by construction. Eq. (6) is a stated softmax definition, and Theorem 1 attempts to show it solves Eq. (7); the proof's own Eq. (10), π*(a|s) ∝ P_LLM(a|s) exp((αQ(s,a)+γr_LTR(s,a))/β), coincides with Eq. (6) only at β=1, whereas Table 6 sets β=2. That is an internal derivation error, not circularity: the paper neither fits Eq. (6) to Eq. (10) nor defines one in terms of the other. Theorem 2's lower bound likewise rests on an unverified 'assume rLTR provides unbiased estimates' and Eq. (15) drops both the γr_LTR term and the β exponent on P_LLM, so the stated bound does not follow as written; this is a missing-support or proof-gap concern, not a circular reduction. The empirical claim is tested against external baselines on published OpenML datasets, and the core comparison is not forced by the method's definition. One mild leakage concern is that Appendix A.4 tunes α, β, γ on ada_prior, uscensus, and wall-robot-nav, all of which also appear in the main Table 3; this makes those three rows optimistic but does not reduce the central average-accuracy claim by construction. The KL-regularized-RL citations ([25]-[27]) are external, with no author overlap and no self-citation chain used to forbid alternatives. Under the stated standard, the honest finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on three tuned weights, a fixed pipeline length, an offline-trained ranker, and two strong assumptions: an always-positive LLM prior and unbiased LTR estimates. These are the main costs the paper pulls from prior work or postulates without independent evidence.

free parameters (5)
  • α (Q-value weight) = 1.0
    Weight on αQ in Eqs. (5)-(6); tuned one-variable-at-a-time in Appendix A.4; the policy depends on it.
  • β (LLM prior weight) = 2.0
    Controls strength of the LLM prior; Appendix A.4 shows multiple local maxima for β, so the chosen value is dataset-sensitive.
  • γ (LTR weight) = 2.0
    Weight on the LTR score; has the same tuning caveat as β.
  • Fixed pipeline length T = 8
    Maximum pipeline length in Algorithm 2 and Table 6; chosen by hand and applied to every dataset, which bounds the search space.
  • LTR ranker parameters = unreleased
    LightGBM ranker trained on an offline dataset of (state, action, performance) tuples; the model weights and training tuples are not provided.
assumptions (5)
  • standard math Bayes rule and control-as-inference formulation with binary optimality variable O apply to action selection.
    Used to motivate Eq. (2), following Levine [25]; standard, but the exponential likelihood form is chosen for convenience.
  • domain assumption Dataset state is adequately represented by meta-features such as row counts, missing values, skew, outliers, and the partial pipeline.
    Section 3 and Section 5.1; if meta-features are insufficient, Q, LTR, and LLM all operate on incomplete state.
  • ad hoc to paper The LLM prior is positive for every operator type that contains an optimal action.
    Theorem 1 explicitly assumes PLLM(a|s)>0; if the LLM assigns near-zero probability to a useful type, soft guidance behaves like a hard constraint.
  • ad hoc to paper LTR scores are unbiased estimates of action optimality and are scale-comparable with Q-values.
    Used in the Theorem 2 proof; no evidence is provided, and LTR is trained offline while Q is learned online.
  • ad hoc to paper Hyperparameters tuned on a small validation set generalize to all 18 datasets.
    Appendix A.4 reports tuning on ada_prior, uscensus, and wall-robot-nav, which also appear in Table 3; if these are evaluation datasets, the tuned weights leak test information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SoftPipe: A Soft-Guided Reinforcement Learning Framework for Automated Data Preparation." pith.science (2026). https://pith.science/paper/JUNTFYSW

@misc{pith2026250713710,
  author       = {Pith},
  title        = {Pith review of: SoftPipe: A Soft-Guided Reinforcement Learning Framework for Automated Data Preparation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JUNTFYSW}},
  note         = {Machine review of arXiv:2507.13710}
}
abstract

Data preparation is a foundational yet notoriously challenging component of the machine learning lifecycle, characterized by a vast combinatorial search space. While reinforcement learning (RL) offers a promising direction, state-of-the-art methods suffer from a critical limitation: to manage the search space, they rely on rigid ``hard constraints'' that prematurely prune the search space and often preclude optimal solutions. To address this, we introduce SoftPipe, a novel RL framework that replaces these constraints with a flexible ``soft guidance'' paradigm. SoftPipe formulates action selection as a Bayesian inference problem. A high-level strategic prior, generated by a Large Language Model (LLM), probabilistically guides exploration. This prior is combined with empirical estimators from two sources through a collaborative process: a fine-grained quality score from a supervised Learning-to-Rank (LTR) model and a long-term value estimate from the agent's Q-function. Through extensive experiments on 18 diverse datasets, we demonstrate that SoftPipe achieves up to a 13.9\% improvement in pipeline quality and 2.8$\times$ faster convergence compared to existing methods.

Figures

Figures reproduced from arXiv: 2507.13710 by the authors.

Figure 1
Figure 1. The SoftPipe Framework. optimal, p(at|st, O = 1). Using Bayes’ rule, this posterior can be decomposed: p(at|st, O = 1) = p(O = 1|st, at)p(at|st) p(O = 1|st) ∝ p(O = 1|st, at) | {z } Likelihood · p(at|st) | {z } Prior (2) This formulation provides the theoretical scaffold for SoftPipe. The framework’s components are designed to instantiate the prior and the likelihood terms as described next. 5.1.1 The LLM as a Strat… view at source ↗
Figure 2
Figure 2. Learning curves on six representative datasets (abalone, ada_prior, avila, house-prices, pbcseq, wall-robot-nav) [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Comparison of inference time 18. The learning curve for SoftPipe consistently shows a much steeper initial ascent, which can be attributed to two core components of our framework: the LTR model provides a dense, immediate reward signal, giving the agent a “warm start,” while the LLM prior guides exploration towards promising regions from the very beginning. In contrast, both the standard Q-Learning and the HRL basel… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Hyperparameter analysis on the weight of Synergistic Policy [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 12 canonical work pages

  1. [1]

    Data preparation for machine learning

    Pecan. Data preparation for machine learning. https://www.pecan.ai/blog/ data-preparation-for-machine-learning/ , 2024. Accessed: 2025-06-25

    2024

  2. [2]

    What are the challenges of implementing automl? https://milvus.io/ai-quick-reference/ what-are-the-challenges-of-implementing-automl// , 2025

    Zilliz. What are the challenges of implementing automl? https://milvus.io/ai-quick-reference/ what-are-the-challenges-of-implementing-automl// , 2025. Accessed: 2025-06-10

    work page 2025

  3. [3]

    Agile data preparation & exploration for cloud machine learning

    John Morrell. Agile data preparation & exploration for cloud machine learning. https://www.datameer.com/ blog/agile-data-preparation-exploration-for-cloud-machine-learning// , 2024. Accessed: 2025-06-25

    2024

  4. [4]

    Learn2clean: Optimizing the sequence of tasks for web data preparation

    Laure Berti-Equille. Learn2clean: Optimizing the sequence of tasks for web data preparation. In The world wide web conference, pages 2580–2586, 2019

    2019

  5. [5]

    Deepline: Automl tool for pipelines generation using deep reinforcement learning and hierarchical actions filtering

    Yuval Heffetz, Roman Vainshtein, Gilad Katz, and Lior Rokach. Deepline: Automl tool for pipelines generation using deep reinforcement learning and hierarchical actions filtering. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2103–2113, 2020

    2020

  6. [6]

    Democratizing data science through interactive curation of ml pipelines

    Zeyuan Shang, Emanuel Zgraggen, Benedetto Buratti, Ferdinand Kossmann, Philipp Eichmann, Yeounoh Chung, Carsten Binnig, Eli Upfal, and Tim Kraska. Democratizing data science through interactive curation of ml pipelines. In Proceedings of the 2019 international conference on management of data, pages 1171–1188, 2019

    2019

  7. [7]

    Auto-pipeline: synthesizing complex data pipelines by-target using reinforcement learning and search

    Junwen Yang, Yeye He, and Surajit Chaudhuri. Auto-pipeline: synthesizing complex data pipelines by-target using reinforcement learning and search. arXiv preprint arXiv:2106.13861, 2021

    arXiv 2021

  8. [8]

    Cleansurvival: Automated data preprocessing for time-to-event models using reinforcement learning

    Yousef Koka, David Selby, Gerrit Großmann, and Sebastian V ollmer. Cleansurvival: Automated data preprocessing for time-to-event models using reinforcement learning. arXiv preprint arXiv:2502.03946, 2025

    arXiv 2025

Show all 33 references
  1. [9]

    Advancing multimodal reasoning: From optimized cold start to staged reinforcement learning

    Shuang Chen, Yue Guo, Zhaochen Su, Yafu Li, Yulun Wu, Jiacheng Chen, Jiayu Chen, Weijie Wang, Xiaoye Qu, and Yu Cheng. Advancing multimodal reasoning: From optimized cold start to staged reinforcement learning. arXiv preprint arXiv:2506.04207, 2025. 11 A PREPRINT - S EPTEMBER 21, 2025

    2025

  2. [10]

    Ctxpipe: Context-aware data preparation pipeline construction for machine learning

    Haotian Gao, Shaofeng Cai, Tien Tuan Anh Dinh, Zhiyong Huang, and Beng Chin Ooi. Ctxpipe: Context-aware data preparation pipeline construction for machine learning. Proc. ACM Manag. Data, 2(6), December 2024

    2024

  3. [11]

    Haipipe: Combining human-generated and machine-generated pipelines for data preparation

    Sibei Chen, Nan Tang, Ju Fan, Xuemi Yan, Chengliang Chai, Guoliang Li, and Xiaoyong Du. Haipipe: Combining human-generated and machine-generated pipelines for data preparation. Proceedings of the ACM on Management of Data, 1(1):1–26, 2023

    2023

  4. [12]

    Evaluation of a tree-based pipeline optimization tool for automating data science

    Randal S Olson, Nathan Bartley, Ryan J Urbanowicz, and Jason H Moore. Evaluation of a tree-based pipeline optimization tool for automating data science. In Proceedings of the genetic and evolutionary computation conference 2016, pages 485–492, 2016

    2016

  5. [13]

    Efficient and robust automated machine learning

    Matthias Feurer, Aaron Klein, Katharina Eggensperger, Jost Tobias Springenberg, Manuel Blum, and Frank Hutter. Efficient and robust automated machine learning. In Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume 2, NIPS’15, pag...

    2015

  6. [14]

    Data civilizer 2.0: A holistic framework for data preparation and analytics

    El Kindi Rezig, Lei Cao, Michael Stonebraker, Giovanni Simonini, Wenbo Tao, Samuel Madden, Mourad Ouzzani, Nan Tang, and Ahmed K Elmagarmid. Data civilizer 2.0: A holistic framework for data preparation and analytics. Proceedings of the VLDB Endowment, 12(12):1954–1957, 2019

    1954

  7. [15]

    Dynaml: a scala & jvm machine learning toolbox for research, education & industry., 2021

    Amit Kumar Jaiswal Mandar Chandorkar. Dynaml: a scala & jvm machine learning toolbox for research, education & industry., 2021

    2021

  8. [16]

    Deep Learning with H2O, 3 2025

    Candel Arno and LeDell Erin. Deep Learning with H2O, 3 2025

    2025

  9. [17]

    Causalcomrl: Context-based offline meta- reinforcement learning with causal representation

    Zhengzhe Zhang, Wenjia Meng, Haoliang Sun, and Gang Pan. Causalcomrl: Context-based offline meta- reinforcement learning with causal representation. arXiv preprint arXiv:2502.00983, 2025

    2025 arXiv

  10. [18]

    Towards sample efficient reinforcement learning

    Yang Yu. Towards sample efficient reinforcement learning. In IJCAI, pages 5739–5743, 2018

    2018

  11. [19]

    Feature engineering for predictive modeling using reinforcement learning

    Udayan Khurana, Horst Samulowitz, and Deepak Turaga. Feature engineering for predictive modeling using reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018

    2018

  12. [20]

    Metaprep: Data preparation pipelines recommen- dation via meta-learning

    Fernando Rezende Zagatti, Lucas Cardoso Silva, Lucas Nildaimon Dos Santos Silva, Bruno Silva Sette, Helena de Medeiros Caseli, Daniel Lucrédio, and Diego Furtado Silva. Metaprep: Data preparation pipelines recommen- dation via meta-learning. In 2021 20th IEEE International Con...

    2021

  13. [21]

    Evaluating large language models trained on code

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021

    2021 arXiv

  14. [22]

    Can large language models be data-driven reasoners for answering tabular questions? arXiv preprint arXiv:2402.18086, 2024

    Yong-qiang Chen, Xin-yu Zhang, Lei-lei Li, Jia-xin Bai, Wang-chun Shu, Ai-ti Awuti, Bo-wen Zhang, Yuan-lin Zhang, and Ai-xin Feng. Can large language models be data-driven reasoners for answering tabular questions? arXiv preprint arXiv:2402.18086, 2024

    2024 arXiv

  15. [23]

    Automl-gpt: A gpt-based tool for automated machine learning

    Zhipeng Gao, Yuqi Zhang, Jiali Li, Yifan Zhang, Zhongyi Liu, Zhineng Chen, Jianjun Chen, Shaohua Yan, Sicheng Zhao, and Wei Huang. Automl-gpt: A gpt-based tool for automated machine learning. arXiv preprint arXiv:2311.08772, 2023

    2023 arXiv

  16. [24]

    Jupyter-eval: A large-scale evaluation of large language models for data science code

    Bhaskar Ghosh, Anijus Khot, Debanjali Mitra, Pin-Yu Chen, and Danish Pruthi. Jupyter-eval: A large-scale evaluation of large language models for data science code. arXiv preprint arXiv:2402.16315, 2024

    2024 arXiv

  17. [25]

    Reinforcement learning and control as probabilistic inference: Tutorial and review

    Sergey Levine. Reinforcement learning and control as probabilistic inference: Tutorial and review. arXiv preprint arXiv:1805.00909, 2018

    2018 arXiv

  18. [26]

    Efficient reinforcement learning with large language model priors

    Xue Yan, Yan Song, Xidong Feng, Mengyue Yang, Haifeng Zhang, Haitham Bou Ammar, and Jun Wang. Efficient reinforcement learning with large language model priors. In Submitted to International Conference on Learning Representations, 2025. Anonymous submission

    2025

  19. [27]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290, 2023

    2023 arXiv

  20. [28]

    Openml: networked science in machine learning

    Joaquin Vanschoren, Jan N Van Rijn, Bernd Bischl, and Luis Torgo. Openml: networked science in machine learning. ACM SIGKDD Explorations Newsletter, 15(2):49–60, 2014

    2014

  21. [29]

    Diffprep: Differentiable data preprocessing pipeline search for learning over tabular data

    Peng Li, Zhiyi Chen, Xu Chu, and Kexin Rong. Diffprep: Differentiable data preprocessing pipeline search for learning over tabular data. Proceedings of the ACM on Management of Data, 1(2):1–26, 2023

    2023

  22. [30]

    Cleanml: A study for evaluating the impact of data cleaning on ml classification tasks

    Peng Li, Xi Rao, Jennifer Blase, Yue Zhang, Xu Chu, and Ce Zhang. Cleanml: A study for evaluating the impact of data cleaning on ml classification tasks. In 2021 IEEE 37th International Conference on Data Engineering (ICDE), pages 13–24. IEEE, 2021. 12 A PREPRINT - S EPTEMBER 21, 2025

    2021

  23. [31]

    H2o automl: Scalable automatic machine learning

    Erin LeDell and Sebastien Poirier. H2o automl: Scalable automatic machine learning. In Proceedings of the AutoML Workshop at ICML, volume 2020, page 24, 2020

    2020

  24. [32]

    Qwen3 technical report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

  25. [33]

    Llamafactory: Unified efficient fine-tuning of 100+ language models

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models. arXiv preprint arXiv:2403.13372, 2024. A Appendix A A.1 A.1 Algorithm Pseudocode Algorithm 2 outlines the cor...

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.