Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

Sculpting Features from Noise: Reward-Guided Hierarchical Diffusion for Task-Optimal Feature Transformation

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

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

Pith's one-line read DIFFT claims automated feature transformation is best modeled as reward-guided generation in a learned latent space, reporting better accuracy than ten baselines on all 14 datasets tested.

desk verdict A plausible new generative take on feature transformation with real efficiency gains, but the empirical claims outrun the current evidence. read the letter →

arxiv 2505.15152 v1 pith:WT2KHKZC submitted 2025-05-21 cs.LG

Nanxu Gong , Zijun Li , Sixun Dong , Haoyue Bai , Wangyang Ying , Xinyuan Wang , Yanjie Fu This is my paper
classification cs.LG
keywords featuretransformationlatentdiffusionmodelrewardguidancesemi-autoregressivedecodingtabulardatavariationalautoencoderengineeringgenerative
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 proposes DIFFT, a framework that reframes automated feature transformation — creating new features from raw ones through operations like logs and products — as a reward-guided generative problem rather than a search problem. It argues that discrete search explodes combinatorially, while continuous gradient search gets stuck in local optima, and that a latent diffusion model trained on feature-set embeddings can instead sample globally while an evaluator's reward gradient steers the sample toward task-optimal features. A semi-autoregressive decoder then turns the sampled embedding into a structured feature set, predicting the number of features and generating each feature's token sequence independently in parallel. If the paper is right, feature transformation becomes faster, more robust, and better performing than the leading continuous-search baseline, with reported gains on all 14 datasets it tests.

What carries the argument

The load-bearing object is the reward-guided latent diffusion sampler. After a VAE encoder maps feature token sequences to embeddings and a GCN maps the transformed table to a condition vector, a Transformer denoiser learns the embedding distribution. At inference, the DDIM update in Equation (9) modifies the predicted noise by subtracting λ∇z(1/2)(Rψ(z)−a)2, so the reverse process follows the evaluator's reward gradient as well as the learned data distribution. The second mechanism is the semi-autoregressive decoder: it first samples the number of features T from pθ(T|z0), then generates each feature chunk's token sequence autoregressively in parallel, preserving intra-feature token dependencies while removing inter-feature order dependence.

What would settle it

On any of the 14 datasets, run DIFFT's reward guidance with the evaluator frozen, but restrict sampling to low-density regions of the latent space, decode the samples, and compare predicted evaluator reward with actual downstream F1 or RAE. If predicted and actual diverge sharply, or if the decoded features underperform the continuous-search baseline MOAT on the same data, the reward-guidance claim is falsified.

Watch

Extended reading notes

Core claim

DIFFT's central claim is that the optimal transformed feature set for a tabular dataset is best obtained by encoding candidate feature token sequences into a compact Gaussian latent space with a VAE, modeling the distribution of high-quality feature-set embeddings with a Transformer-based latent diffusion model conditioned on a graph embedding of the table, and then, during DDIM denoising, steering the trajectory with the gradient of a learned performance evaluator toward high-reward embeddings. The final embedding is decoded by a semi-autoregressive decoder that predicts the number of features and generates each feature's tokens in parallel. The paper reports that this pipeline outperforms ten baselines, including random, linear, reinforcement-learning, evolutionary, and continuous-search methods, on all 14 datasets, and that it trains roughly 25 times faster per epoch and infers more than 5 times faster than the continuous-search baseline MOAT.

Load-bearing premise

The whole pipeline rests on the evaluator network giving accurate reward predictions in latent regions where training embeddings are sparse; if it is miscalibrated there, the reward gradient steers sampling toward embeddings that look good but do not actually improve the downstream model.

Editorial extensions

If this is right

  • Feature transformation can be treated as conditional generation, so improvements in diffusion modeling and reward guidance transfer directly to automated feature engineering.
  • Because sampling is global rather than local, DIFFT-type methods should be more robust to initialization and step-size choices than continuous embedding search, a failure mode documented in earlier methods.
  • Parallel decoding across features makes generative feature transformation practical at larger feature counts, since decoding time no longer grows with the full sequence length.
  • The reward-guidance principle is downstream-model-agnostic: the same trained framework can be steered toward different learners by re-running reward-guided sampling with a different evaluator.
  • On the paper's evidence, the method gives consistent accuracy improvements across 14 diverse tabular datasets, including high-dimensional ones with over 10,000 raw features.

Reading between the lines

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

  • Editorial inference: the same reward-guided latent diffusion recipe could be lifted to other discrete structured search problems, such as neural architecture search, operator scheduling, or query rewriting, wherever a cheap evaluator can score candidate structures.
  • Editorial inference: the central failure mode is evaluator miscalibration in low-density latent regions; one cheap safeguard would be to weight reward gradients by the diffusion model's estimated likelihood or to ensemble multiple evaluators.
  • Editorial inference: because training data are collected by an RL agent, the quality ceiling of the embedding space is partly set by that collector's exploration; a testable extension is to close the loop by adding DIFFT's own generated features back into the training pool.
  • Editorial inference: a task-agnostic version could pre-train the VAE and diffusion model once and fine-tune only the evaluator per dataset, cutting the per-task cost the paper flags as a limitation.
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

5 major / 5 minor

Summary. The paper proposes DIFFT, a feature transformation method for tabular data based on reward-guided latent diffusion. A VAE maps feature token sequences into a latent space; an RL-collected dataset of transformed feature sets is used to train the VAE, a performance evaluator R_psi, and a latent diffusion model conditioned on GCN table embeddings. At inference, the reverse diffusion process is steered by evaluator gradients (Eq. 9) toward high predicted reward, and the resulting embedding is decoded by a semi-autoregressive decoder that predicts the number of features and generates each feature chunk in parallel. The paper claims that DIFFT consistently outperforms ten baselines on 14 benchmark datasets in accuracy, robustness, and training/inference efficiency.

Significance. If the empirical claims hold, DIFFT would be a meaningful step for automated feature transformation: it replaces combinatorial discrete search and initialization-sensitive continuous search with generative sampling, and the semi-autoregressive decoder is a sensible efficiency idea that avoids full autoregressive decoding while preserving intra-feature dependencies. The paper also ships public code and data, and Appendix E is candid about the per-task regeneration cost and width scaling limits. However, the central claim is entirely empirical, and the current evidence is not sufficient to support it: there are no variance estimates or significance tests, a likely train/inference mismatch in the conditioning setup, and no calibration analysis for the reward evaluator that drives the main contribution.

major comments (5)
  1. [§3.1–3.2, Table 1] Table 1 reports a single point estimate per method and dataset, and the manuscript does not describe a train/test split, cross-validation scheme, repeated-seed protocol, or significance test anywhere in Section 3 or Appendix D. Since the method has several stochastic components (VAE training, diffusion sampling, RL data collection), one run per dataset cannot support the claim in Section 3.2 that DIFFT 'consistently achieves the best results on all datasets.' Please report means and standard deviations over multiple seeds, specify how transformed features are constructed on train and test folds to rule out leakage through the evaluator or the condition, and add paired significance tests against the strongest baseline.
  2. [§2.4, Eq. (9); §2.2, Eq. (4)] The reward-guided update steers denoising with the gradient of (1/2)(R_psi(z)-a)^2, but R_psi is trained only on embeddings of feature sets collected by the RL agent in Appendix C, which may occupy a small region of the latent space. Guided diffusion can push samples into low-density regions where the evaluator's MSE fit in Eq. (4) is unconstrained, so high predicted reward need not imply high actual downstream performance. Please provide a calibration experiment on diffusion-generated latents (predicted reward versus actual downstream score after decoding), a sweep over the guidance strength lambda / reward scale mentioned in Appendix D.4, and a comparison of guided versus unguided sampling. Without such evidence, the reported gains could be an artifact of reward hacking rather than genuine feature-quality improvement.
  3. [§2.3, Appendix B, Eqs. (10)-(11)] The conditioning signal c is described as the GCN embedding of the table obtained by applying the candidate feature set to the raw data. At inference, the candidate feature set is not known before sampling, so this condition cannot be computed in the same way. If c is instead computed from the original raw table, then training and inference conditions differ and the text should say so explicitly; if c is computed from the transformed table, the model is conditioning on the very target it is supposed to generate. Please define exactly how c is obtained at inference and, if needed, retrain with a matched condition.
  4. [§3.3 Table 2, §3.5 Table 4] The ablation study covers only 2 of the 14 datasets (svmguide3 and openml_586), and the robustness check covers only SpectF, both without error bars. The statements that NoR and CS suffer 'significant performance drops' and that DIFFT 'consistently achieves the best performance across all downstream models' are therefore not supported by the presented evidence. Please run the ablations on a broader subset or all datasets, report variance, and extend the downstream-model robustness check to at least a few datasets with different sizes and feature counts.
  5. [§3.4, Table 3] Table 3 reports per-epoch training time, but total training time depends on the number of epochs and convergence criteria; Appendix D.4 states the diffusion model is trained for 800 epochs, yet no total wall-clock time is given. The claim of 'significantly less time to train' is therefore not fully supported. Please report end-to-end training time to the final reported results, including RL data collection and VAE/LDM training, and state the corresponding protocol for MOAT.
minor comments (5)
  1. [§2.4 and Appendix D.4] Eq. (9) defines lambda = 1/sigma^2, while Appendix D.4 says the 'reward scale' is set to 100; please clarify the relationship between these quantities and report the chosen value of lambda.
  2. [Figure 3] The figure appears to label its vertical axis 'Performance,' but the caption and text discuss decoding time; please relabel the axis with the time unit and ensure the data plotted are consistent with the claim about decoding speed.
  3. [Appendix D.2] The LDA baseline is cited to reference [3], which is Latent Dirichlet Allocation, not Linear Discriminant Analysis; please cite the correct LDA method.
  4. [§3.1 and Appendix D.2] The abbreviations RDG and ERG are used without expansion, and the regression metric '1-relative absolute error' is not defined with a formula; please add definitions.
  5. [§5] The conclusion refers to 'feature transformation and selection methods,' but the paper only studies feature transformation; please align the wording with the actual scope.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: final metrics are measured by downstream models on test data, not by the fitted evaluator; reward guidance is adopted from external RCGDM and no equation reduces a prediction to its inputs.

full rationale

The derivation chain is self-contained in the sense required by the circularity rubric. DIFFT first collects candidate feature sequences via an RL agent (Appendix C), encodes them with a VAE, trains a latent diffusion model on the resulting embeddings, trains a performance evaluator Rψ, and then uses Rψ's gradients in Eq. (9) only as a sampling steering signal. The headline results in Table 1 are not computed from Rψ; they are F1 and 1-RAE scores obtained by training the actual downstream random forest on the decoded feature sets. Thus the central claim is not a fitted-input-called-prediction: the evaluator is not the source of the reported performance numbers. The reward-guidance update itself is explicitly adopted from RCGDM [41], an external method, and the approximation of classifier-based guidance is attributed to Dhariwal and Nichol [5]; no load-bearing argument reduces to an unverified self-citation. The paper's self-citations to GRFG [28], ELLM [6], and MOAT [29] appear either as baselines or as inspiration for the RL data collector, but the collector is fully described in Appendix C with its own agents, state representation, reward function, and algorithm, so it is not imported as a black-box uniqueness claim. The ablation variants (NoR, CS, NAR, AR) and robustness checks provide independent evidence for the contribution of reward guidance, diffusion, and semi-autoregressive decoding. The stated limitations in Appendix E concern scalability and per-task regeneration cost, not a hidden dependence of the derivation on its own conclusion. No step in the paper's claimed derivation reduces, by construction or by self-citation, to its own inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on hand-chosen hyperparameters and domain assumptions about the RL-collected training distribution and the evaluator's reliability. No unknown entities are introduced, and the loss functions are standard, so the main burden falls on whether the training data and reward signal faithfully represent the feature transformation landscape.

free parameters (4)
  • VAE loss weights α, β, γ = not reported
    Eq. (5) combines reconstruction, feature-count, evaluator, and KL losses; the weights are chosen by hand and not reported in the text, affecting embedding quality and downstream performance.
  • Reward guidance strength λ = 1/σ² = reward scale 100 (Appendix D.4)
    Eq. (9) uses λ to control evaluator gradient influence; the paper sets it to 100 but does not report sensitivity.
  • RL data collector budget (episodes N, steps T) = not reported
    Appendix C/Algorithm 1 uses N episodes and T steps; these determine the training distribution of feature sets and are not specified.
  • Latent dimension of VAE = not reported
    Not reported in the text; determines the expressiveness of the embedding space for the diffusion model.
assumptions (5)
  • domain assumption The postfix operator set O and pairwise feature operations can express useful transformations for the target datasets.
    The whole generative space is defined by O; Sections 2.2 and Appendix A assume this representation covers the relevant formulas.
  • domain assumption The RL data collector produces a diverse and representative sample of high and low performance feature sets.
    Appendix C describes the collector; the VAE, evaluator, and diffusion model all train on this sample, so its coverage determines generalization.
  • domain assumption The evaluator Rψ provides reliable reward gradients in regions of the latent space reached by sampling.
    Eq. (9) steers denoising with ∇z Rψ; if Rψ is miscalibrated off-distribution, reward guidance can select embeddings with high predicted but low actual performance.
  • domain assumption The latent space can be treated as approximately Gaussian so that DDIM sampling with evaluator gradients is valid.
    Section 2.4 states 'we follow the gradient-steered sampler of RCGDM to consider the latent z as Gaussian'; this is an assumption about the VAE posterior.
  • domain assumption Downstream performance metrics (F1 for classification, 1-RAE for regression) with random forest are the correct optimization targets.
    Rewards for the RL collector and evaluator are defined by these metrics; Section 3.1 fixes the downstream model, so the method optimizes for that choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sculpting Features from Noise: Reward-Guided Hierarchical Diffusion for Task-Optimal Feature Transformation." pith.science (2026). https://pith.science/paper/WT2KHKZC

@misc{pith2026250515152,
  author       = {Pith},
  title        = {Pith review of: Sculpting Features from Noise: Reward-Guided Hierarchical Diffusion for Task-Optimal Feature Transformation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WT2KHKZC}},
  note         = {Machine review of arXiv:2505.15152}
}
read the original abstract

Feature Transformation (FT) crafts new features from original ones via mathematical operations to enhance dataset expressiveness for downstream models. However, existing FT methods exhibit critical limitations: discrete search struggles with enormous combinatorial spaces, impeding practical use; and continuous search, being highly sensitive to initialization and step sizes, often becomes trapped in local optima, restricting global exploration. To overcome these limitations, DIFFT redefines FT as a reward-guided generative task. It first learns a compact and expressive latent space for feature sets using a Variational Auto-Encoder (VAE). A Latent Diffusion Model (LDM) then navigates this space to generate high-quality feature embeddings, its trajectory guided by a performance evaluator towards task-specific optima. This synthesis of global distribution learning (from LDM) and targeted optimization (reward guidance) produces potent embeddings, which a novel semi-autoregressive decoder efficiently converts into structured, discrete features, preserving intra-feature dependencies while allowing parallel inter-feature generation. Extensive experiments on 14 benchmark datasets show DIFFT consistently outperforms state-of-the-art baselines in predictive accuracy and robustness, with significantly lower training and inference times.

Figures

Figures reproduced from arXiv: 2505.15152 by the authors.

Figure 1
Figure 1. Motivation example. Discrete search methods explore various feature combinations [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Framework overview. The framework consists of three key components: 1) a VAE that [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Time analysis of generating 2000 to￾kens using different methods. To investigate the contribution of each component in DIFFT, we design variant models: 1) NAR and AR, which examine the impact of the semi-autoregressive decoder by replacing it with fully non-autoregressive and fully autoregressive alternatives, respectively; 2) NoR, which removes reward guidance during the diffusion process to evaluate its effect on … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of continuous search method and optimized generation method. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Teaching Time Series to See and Speak: Forecasting with Aligned Visual and Textual Perspectives

    cs.LG 2025-06 reject novelty 5.0 of 10

    TimesCLIP aligns image-based and text-based views of the same time series via contrastive learning to improve forecasting accuracy on several benchmarks, but the full multimodal model is not used on two of the six lon...

  2. LLM-ML Teaming: Integrated Symbolic Decoding and Gradient Search for Valid and Stable Generative Feature Transformation

    cs.LG 2025-06 conditional novelty 4.0 of 10

    A product-of-experts decoder that blends a fine-tuned LLM's token probabilities with a gradient-searched sequence decoder produces more valid and stable feature transformations than either alone.

Reference graph

Works this paper leans on

42 extracted references · 19 canonical work pages · cited by 2 Pith papers

  1. [1]

    Is conditional generative modeling all you need for decision-making? arXiv preprint arXiv:2211.15657, 2022

    Anurag Ajay, Yilun Du, Abhi Gupta, Joshua Tenenbaum, Tommi Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision-making? arXiv preprint arXiv:2211.15657, 2022

    arXiv 2022

  2. [2]

    Training diffusion models with reinforcement learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023

    arXiv 2023

  3. [3]

    Latent dirichlet allocation

    David M Blei, Andrew Y Ng, and Michael I Jordan. Latent dirichlet allocation. Journal of machine Learning research, 3(Jan):993–1022, 2003

    2003

  4. [4]

    Neural feature search: A neural architecture for automated feature engineering

    Xiangning Chen, Qingwei Lin, Chuan Luo, Xudong Li, Hongyu Zhang, Yong Xu, Yingnong Dang, Kaixin Sui, Xu Zhang, Bo Qiao, et al. Neural feature search: A neural architecture for automated feature engineering. In 2019 IEEE International Conference on Data Mining (ICDM), pages 71–80. IEEE, 2019

    2019

  5. [5]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  6. [6]

    Evolutionary large language model for automated feature transformation

    Nanxu Gong, Chandan K Reddy, Wangyang Ying, Haifeng Chen, and Yanjie Fu. Evolutionary large language model for automated feature transformation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 16844–16852, 2025

    work page 2025

  7. [7]

    Unsupervised feature transformation via in-context generation, generator-critic llm agents, and duet-play teaming

    Nanxu Gong, Xinyuan Wang, Wangyang Ying, Haoyue Bai, Sixun Dong, Haifeng Chen, and Yanjie Fu. Unsupervised feature transformation via in-context generation, generator-critic llm agents, and duet-play teaming. arXiv preprint arXiv:2504.21304, 2025

    arXiv 2025

  8. [8]

    Neuro-symbolic embedding for short and effective feature selection via autoregressive generation

    Nanxu Gong, Wangyang Ying, Dongjie Wang, and Yanjie Fu. Neuro-symbolic embedding for short and effective feature selection via autoregressive generation. ACM Transactions on Intelligent Systems and Technology, 16(2):1–21, 2025

    2025

Show all 42 references
  1. [9]

    Protein design with guided discrete diffusion

    Nate Gruver, Samuel Stanton, Nathan Frey, Tim GJ Rudner, Isidro Hotzel, Julien Lafrance-Vanasse, Arvind Rajpal, Kyunghyun Cho, and Andrew G Wilson. Protein design with guided discrete diffusion. Advances in neural information processing systems, 36:12489–12517, 2023

    2023

  2. [10]

    Efficient diffusion training via min-snr weighting strategy

    Tiankai Hang, Shuyang Gu, Chen Li, Jianmin Bao, Dong Chen, Han Hu, Xin Geng, and Baining Guo. Efficient diffusion training via min-snr weighting strategy. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7441–7451, 2023

    2023

  3. [11]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  4. [12]

    The autofeat python library for automated feature engineering and selection

    Franziska Horn, Robert Pack, and Michael Rieger. The autofeat python library for automated feature engineering and selection. In Machine Learning and Knowledge Discovery in Databases: International Workshops of ECML PKDD 2019, Würzburg, Germany, September 16–20, 2019, Proceedi...

    2019

  5. [13]

    Reinforcement feature transformation for polymer property performance prediction

    Xuanming Hu, Dongjie Wang, Wangyang Ying, and Yanjie Fu. Reinforcement feature transformation for polymer property performance prediction. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 4538–4545, 2024

    2024

  6. [14]

    Planning with diffusion for flexible behavior synthesis

    Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991, 2022

    2022 arXiv

  7. [15]

    Deep feature synthesis: Towards automating data science endeavors

    James Max Kanter and Kalyan Veeramachaneni. Deep feature synthesis: Towards automating data science endeavors. In 2015 IEEE international conference on data science and advanced analytics (DSAA), pages 1–10. IEEE, 2015

    2015

  8. [16]

    Non-autoregressive machine translation with disentangled context transformer

    Jungo Kasai, James Cross, Marjan Ghazvininejad, and Jiatao Gu. Non-autoregressive machine translation with disentangled context transformer. In International conference on machine learning, pages 5144–5155. PMLR, 2020

    2020

  9. [17]

    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. 10

    2018

  10. [18]

    Semi-supervised classification with graph convolutional networks

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

    2016 arXiv

  11. [19]

    Diffusion models for black-box optimization

    Siddarth Krishnamoorthy, Satvik Mehul Mashkaria, and Aditya Grover. Diffusion models for black-box optimization. In International Conference on Machine Learning, pages 17842–17857. PMLR, 2023

    2023

  12. [20]

    Principal components analysis (pca)

    Andrzej Ma ´ckiewicz and Waldemar Ratajczak. Principal components analysis (pca). Computers & Geosciences, 19(3):303–342, 1993

    1993

  13. [21]

    Diffusion models are minimax optimal distribution estimators

    Kazusato Oko, Shunta Akiyama, and Taiji Suzuki. Diffusion models are minimax optimal distribution estimators. In International Conference on Machine Learning, pages 26517–26582. PMLR, 2023

    2023

  14. [22]

    Learning to recover from multi-modality errors for non- autoregressive neural machine translation

    Qiu Ran, Yankai Lin, Peng Li, and Jie Zhou. Learning to recover from multi-modality errors for non- autoregressive neural machine translation. arXiv preprint arXiv:2006.05165, 2020

    2006 arXiv

  15. [23]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  16. [24]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  17. [25]

    Insertion transformer: Flexible sequence generation via insertion operations

    Mitchell Stern, William Chan, Jamie Kiros, and Jakob Uszkoreit. Insertion transformer: Flexible sequence generation via insertion operations. In International Conference on Machine Learning, pages 5976–5985. PMLR, 2019

    2019

  18. [26]

    Genetic programming for feature construction and selection in classification on high-dimensional data

    Binh Tran, Bing Xue, and Mengjie Zhang. Genetic programming for feature construction and selection in classification on high-dimensional data. Memetic Computing, 8:3–15, 2016

    2016

  19. [27]

    Semi-autoregressive neural machine translation.arXiv preprint arXiv:1808.08583, 2018

    Chunqi Wang, Ji Zhang, and Haiqing Chen. Semi-autoregressive neural machine translation.arXiv preprint arXiv:1808.08583, 2018

    2018 arXiv

  20. [28]

    Group-wise reinforcement feature generation for optimal and explainable representation space reconstruction

    Dongjie Wang, Yanjie Fu, Kunpeng Liu, Xiaolin Li, and Yan Solihin. Group-wise reinforcement feature generation for optimal and explainable representation space reconstruction. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1826–1...

    2022

  21. [29]

    Reinforcement-enhanced autore- gressive feature transformation: Gradient-steered search in continuous space for postfix expressions

    Dongjie Wang, Meng Xiao, Min Wu, Yuanchun Zhou, Yanjie Fu, et al. Reinforcement-enhanced autore- gressive feature transformation: Gradient-steered search in continuous space for postfix expressions. Advances in Neural Information Processing Systems, 36:43563–43578, 2023

    2023

  22. [30]

    Towards data-centric ai: A comprehensive survey of traditional, reinforcement, and generative approaches for tabular data transformation

    Dongjie Wang, Yanyong Huang, Wangyang Ying, Haoyue Bai, Nanxu Gong, Xinyuan Wang, Sixun Dong, Tao Zhe, Kunpeng Liu, Meng Xiao, et al. Towards data-centric ai: A comprehensive survey of traditional, reinforcement, and generative approaches for tabular data transformation. arXiv...

    2025 arXiv

  23. [31]

    Knockoff-guided feature selection via a single pre-trained reinforced agent

    Xinyuan Wang, Dongjie Wang, Wangyang Ying, Rui Xie, Haifeng Chen, and Yanjie Fu. Knockoff-guided feature selection via a single pre-trained reinforced agent. arXiv preprint arXiv:2403.04015, 2024

    2024 arXiv

  24. [32]

    Mvdd: Multi-view depth diffusion models

    Zhen Wang, Qiangeng Xu, Feitong Tan, Menglei Chai, Shichen Liu, Rohit Pandey, Sean Fanello, Achuta Kadambi, and Yinda Zhang. Mvdd: Multi-view depth diffusion models. In European Conference on Computer Vision, pages 236–253. Springer, 2024

    2024

  25. [33]

    Traceable automatic feature transformation via cascading actor-critic agents

    Meng Xiao, Dongjie Wang, Min Wu, Ziyue Qiao, Pengfei Wang, Kunpeng Liu, Yuanchun Zhou, and Yanjie Fu. Traceable automatic feature transformation via cascading actor-critic agents. In Proceedings of the 2023 SIAM International Conference on Data Mining (SDM), pages 775–783. SIAM, 2023

    2023

  26. [34]

    Symmcompletion: High-fidelity and high- consistency point cloud completion with symmetry guidance

    Hongyu Yan, Zijun Li, Kunming Luo, Li Lu, and Ping Tan. Symmcompletion: High-fidelity and high- consistency point cloud completion with symmetry guidance. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 9094–9102, 2025

    2025

  27. [35]

    Self-optimizing feature generation via categorical hashing representation and hierarchical reinforcement crossing

    Wangyang Ying, Dongjie Wang, Kunpeng Liu, Leilei Sun, and Yanjie Fu. Self-optimizing feature generation via categorical hashing representation and hierarchical reinforcement crossing. In 2023 IEEE International Conference on Data Mining (ICDM), pages 748–757. IEEE, 2023

    2023

  28. [36]

    Topology-aware reinforcement feature space reconstruction for graph data

    Wangyang Ying, Haoyue Bai, Kunpeng Liu, and Yanjie Fu. Topology-aware reinforcement feature space reconstruction for graph data. arXiv preprint arXiv:2411.05742, 2024

    2024 arXiv

  29. [37]

    Feature selection as deep sequential generative learning

    Wangyang Ying, Dongjie Wang, Haifeng Chen, and Yanjie Fu. Feature selection as deep sequential generative learning. ACM Transactions on Knowledge Discovery from Data, 18(9):1–21, 2024. 11

    2024

  30. [38]

    Revolutionizing biomarker discovery: Leveraging generative ai for bio-knowledge-embedded continuous space exploration

    Wangyang Ying, Dongjie Wang, Xuanming Hu, Ji Qiu, Jin Park, and Yanjie Fu. Revolutionizing biomarker discovery: Leveraging generative ai for bio-knowledge-embedded continuous space exploration. In Proceedings of the 33rd ACM International Conference on Information and Knowledg...

    2024

  31. [39]

    Unsupervised generative feature transformation via graph contrastive pre-training and multi-objective fine- tuning

    Wangyang Ying, Dongjie Wang, Xuanming Hu, Yuanchun Zhou, Charu C Aggarwal, and Yanjie Fu. Unsupervised generative feature transformation via graph contrastive pre-training and multi-objective fine- tuning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery...

    2024

  32. [40]

    A survey on data-centric ai: Tabular learning from reinforcement learning and generative ai perspective

    Wangyang Ying, Cong Wei, Nanxu Gong, Xinyuan Wang, Haoyue Bai, Arun Vignesh Malarkkan, Sixun Dong, Dongjie Wang, Denghui Zhang, and Yanjie Fu. A survey on data-centric ai: Tabular learning from reinforcement learning and generative ai perspective. arXiv preprint arXiv:2502.08828, 2025

    2025 arXiv

  33. [41]

    Reward-directed conditional diffusion: Provable distribution estimation and reward improvement.Advances in Neural Information Processing Systems, 36:60599–60635, 2023

    Hui Yuan, Kaixuan Huang, Chengzhuo Ni, Minshuo Chen, and Mengdi Wang. Reward-directed conditional diffusion: Provable distribution estimation and reward improvement.Advances in Neural Information Processing Systems, 36:60599–60635, 2023

    2023

  34. [42]

    f1 ∗ f2, log f3, f4 + f5

    Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions On Graphics (TOG), 42(4):1–16, 2023. 12 A Preliminaries Original Sequence𝑓! 𝑓"𝑓#…Feature sequence Estimat...

    2023

Pith tools

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