Pith. sign in

REVIEW 4 major objections 7 minor 28 references

Adaptive Principal Components Allocation with the $\ell_{2,g}$-regularized Gaussian Graphical Model for Efficient Fine-Tuning Large Models

T0 review · 4 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A Gaussian Graphical Model can decide which SVD components stay trainable, matching LoRA on GLUE with 0.084M parameters.

desk verdict A GGM-based PEFT selection method with a plausible core and enough missing specification to make the central result unverifiable; deserves a demanding major revision, not rejection without review. read the letter →

arxiv 2412.08592 v1 pith:JRD6T7MW submitted 2024-12-11 cs.LG

Jingjing Zheng , Yankai Cao This is my paper · ORCID
classification cs.LG
keywords GaussianGraphicalModelsParameter-EfficientFine-Tuningℓ2gregularizationSVDprincipalcomponentsblockcoordinatedescentGLUEbenchmarkRoBERTastructuralsparsity
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 a parameter-efficient fine-tuning method that lets a Gaussian Graphical Model decide which principal components of each layer's weights stay trainable. It claims this is the first application of Gaussian Graphical Models to parameter-efficient fine-tuning, and reports that fine-tuning RoBERTa-base this way reaches an average score of 85.1 on four GLUE tasks with only 0.084M trainable parameters, beating LoRA's 84.9 while using less than a third of LoRA's trainable budget. The model's nodes are the SVD principal components of each layer plus its bias; their values are importance scores derived from how much each weight moves the loss. A precision matrix is then estimated under a non-convex $\ell_{2,g}$ penalty that selects whole nodes rather than individual edges, and a block coordinate descent algorithm solves the resulting problem. If the central claim holds, modeling parameter interactions before freezing weights is a viable alternative to fixed-rank or importance-score-only allocation of the fine-tuning budget.

What carries the argument

The load-bearing object is the $\ell_{2,g}$-regularized Gaussian Graphical Model precision matrix $\Omega$. Each row of $\Omega$ encodes how one trainable node interacts with the other nodes, and the regularization $\|\bar{\Omega}\|_{2,g}$ applies a non-convex concave surrogate $g$ to the $\ell_2$ norms of the rows of a restricted submatrix, so the penalty drives entire nodes out of the active set rather than merely thinning individual edges. The optimization is solved by block coordinate descent: one step updates $\Omega$ by gradient descent on a concave $\log\det$ objective with a quadratic proximal penalty, and the other step updates an auxiliary variable $\Delta$ using the non-convex proximal map $T_g(y,\tau/2\lambda)$, computed by the GAI fixed-point iteration. The selection rule that comes out of the solve — keep the nodes in $I$ plus the nodes with large row norm, freeze the rest — is what gives the parameter-efficiency result.

What would settle it

Re-run Ours-2 on the same four GLUE tasks with the same 0.084M budget, but draw the trainable SVD components uniformly at random instead of from the GGM solve, and compare the average score. If random selection matches or exceeds 85.1, the interaction model is not what carries the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that parameter interactions, not just individual importance scores, should decide which weights are trainable, and that a Gaussian Graphical Model is the right way to capture those interactions. For each layer, the method takes the top $r$ SVD principal components $(A_i,B_i)$ plus the bias as graph nodes, and assigns each node a value by averaging the importance score $s^{(k)}$ over its entries. A sample covariance matrix of these node values is formed from early training steps, and the method solves maximize over positive-semidefinite $\Omega$ of $\log\det\Omega - \langle\hat{\Sigma},\Omega\rangle - \tau\|\bar{\Omega}\|_{2,g}$, where the last term penalizes, for each node outside the set of 'important nodes' $I$, a concave surrogate $g$ of the row norm $\|\bar{\Omega}_i\|_2$. After solving, it continues training the important nodes and the nodes whose $\|\Omega_i\|_2$ is large, freezing everything else. On SST-2, MRPC, CoLA, and STS-B, the full version ('Ours-2') reports 0.084M trainable parameters and an average score of 85.1, compared with 84.9 for LoRA and 85.0 for full fine-tuning.

Load-bearing premise

The load-bearing premise is that importance scores from the first few training steps give a sample covariance matrix that accurately represents parameter interactions for the entire fine-tuning run; the paper never reports how many samples or steps were used, nor a stability check.

Editorial extensions

If this is right

  • If the GGM selection is correct, a fine-tuning budget of 0.084M parameters can match or beat LoRA's 0.3M on the four tested GLUE tasks, so interaction structure is a usable signal for allocating rank across layers.
  • The method offers a new PEFT baseline: any future low-rank adapter can be compared against a budget allocated by precision-matrix row norms instead of fixed ranks or local importance scores.
  • Fewer trainable parameters means proportionally smaller optimizer state and stored gradients during fine-tuning, which is the practical memory motivation for PEFT that the paper cites.
  • The gap between Ours-1 and Ours-2 in the ablation suggests that adding the 'important nodes' set $I$ contributes to the result, so the sample-mean node values carry information beyond the covariance structure.

Reading between the lines

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

  • The paper leaves the stability of its early-training covariance estimate untested; a practical check would be to refit the GGM after 10%, 25%, and 50% of training and compare the selected node sets.
  • A random-node control is the sharpest missing experiment: with the same 0.084M budget, random SVD components should be kept trainable to isolate whether the GGM's interaction structure, rather than the low-rank SVD parameterization, drives the 85.1 average.
  • The node definition is not BERT-specific, so the same GGM selection should apply to vision or multimodal transformers whose weight matrices also show low-rank spectra; the paper only demonstrates RoBERTa-base.
  • The method estimates the precision matrix once from early samples; a periodic update of $\Omega$ during training would test whether parameter interactions drift, at the cost of extra computation.
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

4 major / 7 minor

Summary. The paper proposes a parameter-efficient fine-tuning (PEFT) method based on a Gaussian Graphical Model (GGM). The method reparameterizes each pre-trained weight matrix via SVD, treats the top-r principal components plus the bias as nodes, computes node importance values from the PLATON-style scores during early training, fits an ℓ2,g-regularized precision matrix to the sample covariance of these values, and then fine-tunes only the nodes selected by large sample means and large ℓ2 norms of precision-matrix columns. A block coordinate descent (BCD) algorithm is presented for the non-convex optimization problem. Experiments on four GLUE tasks with RoBERTa-Base report 0.084M trainable parameters and an average score of 85.1, compared with 84.9 for LoRA.

Significance. If the results are reproducible and the comparisons are statistically meaningful, the paper would make a genuinely novel contribution: it is, to my knowledge, the first application of GGM-based structural sparsity to PEFT, and the reported parameter count (0.084M) is substantially lower than that of LoRA and AdaLoRA. The idea of using a precision matrix to capture interactions among SVD components is plausible and worth pursuing. However, the contribution is currently conditional: the surrogate function g is never fixed, the covariance estimation and node-selection thresholds are unspecified, and the experimental section does not report variability or hyperparameter settings. The paper also provides no formal convergence or consistency guarantees for the proposed BCD scheme.

major comments (4)
  1. [Section 2.3 and Table 1 (second table)] The surrogate function g is never specified for the proposed method. Section 2.3 defines the ℓ2,g regularization through an arbitrary non-negative, increasing, concave g, and Table 1 lists several examples, but the manuscript does not state which g is used in the experiments, nor the value of its internal hyperparameter (e.g., γ). Algorithm 3 explicitly depends on g'(x), and the selected node set depends on the resulting Ω*, so the reported 0.084M parameter selection is not reproducible without this information.
  2. [Section 2.2 and Algorithm 2] The sample covariance \widehatΣ is not defined in terms of the number of early training steps K or the number of nodes n, and no shrinkage or regularization is described. With n on the order of hundreds to thousands and K plausibly small, \widehatΣ is rank-deficient; then log det Ω in problem (6) is only finite for positive definite Ω, and the BCD initialization Ω^(0) = diag(\widehatΣ)^{-1} in Algorithm 2 requires strictly positive diagonal entries, which is not guaranteed because many importance scores are initially zero. The authors should specify K, n, any shrinkage (e.g., adding εI), and a well-defined initialization.
  3. [Section 3] The BCD algorithm's convergence is not analyzed. Problem (6) is non-convex, and the paper neither proves convergence of the BCD iterates to a stationary point nor states conditions under which such convergence holds. The statement that Δ converges to Ω as λ → +∞ concerns the reformulation, not the convergence of the iterates, and the claimed superlinear convergence in Algorithm 3 applies only to the scalar proximal subproblem. A convergence or stationarity guarantee is needed to support the claim that the optimization problem is 'efficiently solved.'
  4. [Section 4, Table 2] The central empirical claim—competitive performance with 0.084M trainable parameters—is not supported with statistical evidence. The results are averaged over five seeds, but no standard deviations, confidence intervals, or significance tests are reported, so the 85.1 vs. 84.9 average difference from LoRA is within the range of typical seed noise. Additionally, the experimental configuration is missing: the rank r, the regularization parameters τ and λ, the momentum parameters β1 and β2, the number of initial training steps K, the threshold for selecting the 'important nodes' I, and the chosen surrogate g are all absent. Without these, Table 2 cannot be reproduced or meaningfully compared with prior work.
minor comments (7)
  1. [Section 2.2] The definition of the 'important nodes' set I is incomplete: the manuscript says I is determined by selecting 'the nodes with the highest values in the sample mean,' but it does not specify how many nodes are selected or what threshold is used; this should be made explicit.
  2. [Throughout] There are two tables labeled 'Table 1': one is the symbol list and the other lists surrogate functions of the ℓ0 norm; renumber the tables to avoid confusion.
  3. [Section 2.1] Figure 2 is referred to as 'Figure 2.1' in the text; the citation should be to Figure 2.
  4. [Algorithm 1] Step 3 says 'Collect samples during the initial stages of the training process,' but the number of samples K is never stated; this should be a concrete hyperparameter in the experiments.
  5. [Section 4] The paper says 'GLUE benchmark' but evaluates only four of the eight GLUE tasks; the abstract and introduction should state this limitation explicitly.
  6. [Abstract] The code URL in the abstract contains a space ('Course projects.git') and appears not to be a standard repository path; a stable, accessible link should be provided.
  7. [Table 3] The ablation study in Table 3 reports results for only three tasks and a single seed; reporting the same error bars as in Table 2 would strengthen the comparison between Ours-1 and Ours-2.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the GGM selection is a data-driven pruning heuristic whose output is evaluated on held-out GLUE tasks, the 'important nodes' set is an explicit design input rather than a disguised prediction, and the only self-citation (Algorithm 3, GAI) is fully reproduced in the paper.

full rationale

The derivation chain is self-contained and does not reduce to its inputs. Node values are defined from PLATON importance scores (Eqs. 1-3); the sample mean and sample covariance are computed from those values during early training (Section 2.2, Algorithm 1 steps 3-4). The set I of 'important nodes' is explicitly determined by the highest sample-mean values, and the GGM regularization uses I only to define which off-diagonal blocks are penalized. The final selection is the union of I and nodes with large ||Omega*_i||2, where the latter is estimated from the precision matrix and therefore carries conditional-dependence information beyond the sample mean; this is not equal to the mean-based ranking by construction. The only self-citation is to Algorithm 3 / GAI [28], co-authored by one of the present authors, but the algorithm is fully reproduced in the paper and its convergence property is not load-bearing for the empirical PEFT claim. The experimental evaluation on GLUE test sets is external to the selection procedure. The conclusion's statement that the work is 'still ongoing' and the missing implementation details (e.g., number of early training steps, number of nodes, shrinkage for the sample covariance) are reproducibility and completeness concerns, not circularity. No fitted parameter is renamed as a prediction, and no uniqueness theorem or prior-work claim is used to forbid alternatives. Therefore the paper receives a circularity score of 0.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The method introduces a new node abstraction (principal components and bias per layer) but this is a modeling choice rather than a postulated physical entity. The main free parameters are numerous and mostly unspecified, which limits the paper's completeness. The axioms are standard assumptions in PEFT but are not tested.

free parameters (7)
  • r
    Number of principal components selected per layer; not specified in the paper.
  • tau
    Regularization parameter for the GGM; not specified in the paper.
  • lambda
    Penalty parameter for the BCD reformulation; not specified in the paper.
  • beta1, beta2
    Smoothing factors for the importance score; not specified in the paper.
  • number of initial training steps
    Number of samples collected for the covariance matrix; not specified in the paper.
  • threshold for 'highest values' in sample mean
    Determines the set I of important nodes; not specified in the paper.
  • surrogate function g
    The paper lists examples but does not specify which g is used in experiments.
assumptions (4)
  • domain assumption The pre-trained weight matrices have a low-rank structure that can be exploited via SVD to define a small set of nodes.
    The paper uses SVD on the query weight matrix of RoBERTa to justify low-rankness, but this is assumed for all layers and all tasks.
  • domain assumption Importance scores based on |W * gradient of loss| are informative for selecting parameters to fine-tune.
    This is the basis of the node values, borrowed from PLATON, but not re-validated in this PEFT context.
  • ad hoc to paper A Gaussian graphical model fitted to importance-score samples captures interactions among parameters that are relevant for fine-tuning.
    This is the core modeling assumption of the paper, stated without empirical validation or theoretical justification.
  • ad hoc to paper The set of important nodes I can be identified by the sample mean of importance scores before training.
    The paper selects nodes with 'highest values in the sample mean', but the size of I or the threshold is not specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Principal Components Allocation with the $\ell_{2,g}$-regularized Gaussian Graphical Model for Efficient Fine-Tuning Large Models." pith.science (2026). https://pith.science/paper/JRD6T7MW

@misc{pith2026241208592,
  author       = {Pith},
  title        = {Pith review of: Adaptive Principal Components Allocation with the $\ell_2,g$-regularized Gaussian Graphical Model for Efficient Fine-Tuning Large Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JRD6T7MW}},
  note         = {Machine review of arXiv:2412.08592}
}
abstract

In this work, we propose a novel Parameter-Efficient Fine-Tuning (PEFT) approach based on Gaussian Graphical Models (GGMs), marking the first application of GGMs to PEFT tasks, to the best of our knowledge. The proposed method utilizes the $\ell_{2,g}$-norm to effectively select critical parameters and capture global dependencies. The resulting non-convex optimization problem is efficiently solved using a Block Coordinate Descent (BCD) algorithm. Experimental results on the GLUE benchmark [24] for fine-tuning RoBERTa-Base [18] demonstrate the effectiveness of the proposed approach, achieving competitive performance with significantly fewer trainable parameters. The code for this work is available at: https://github.com/jzheng20/Course projects.git.

Figures

Figures reproduced from arXiv: 2412.08592 by the authors.

Figure 1
Figure 1. Growth of Large Model Parameters (2018–2022). [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the low-rank property in the learned over [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 19 canonical work pages

  1. [1]

    Pushing data into cp models using graphical model learn- ing and solving

    C´eline Brouard, Simon de Givry, and Thomas Schiex. Pushing data into cp models using graphical model learn- ing and solving. In Helmut Simonis, editor, Principles and Practice of Constraint Programming, pages 811–827. Springer International Publishing, 2020

    work page 2020

  2. [2]

    Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, and et al

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, and et al. Language models are few-shot learners. 2020

    work page 2020

  3. [3]

    Palm: scaling language model- ing with pathways, 2022

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, and et al. Palm: scaling language model- ing with pathways, 2022

    work page 2022

  4. [4]

    Inter-subject analysis: A partial gaussian graphical model approach

    Junwei Lu Cong Ma and Han Liu. Inter-subject analysis: A partial gaussian graphical model approach. Journal of the American Statistical Association , 116(534):746–755, 2021

    work page 2021

  5. [5]

    Qlora: efficient finetuning of quantized llms

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

  6. [6]

    Bert: pre-training of deep bidirec- tional transformers for language understanding, 2019

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: pre-training of deep bidirec- tional transformers for language understanding, 2019

    work page 2019

  7. [7]

    An image is worth 16x16 words: transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    arXiv 2010

  8. [8]

    A statistical view of some chemometrics regression tools

    LLdiko E Frank and Jerome H Friedman. A statistical view of some chemometrics regression tools. Technomet- rics, 35(2):109–135, 1993

    work page 1993

Show all 28 references
  1. [9]

    Sparse inverse covariance estimation with the graphical lasso

    Jerome Friedman, Trevor Hastie, and Robert Tibshirani. Sparse inverse covariance estimation with the graphical lasso. Biostatistics, 9(3):432–441, 2007

    2007

  2. [10]

    Fast sparse regression and classi- fication

    Jerome H Friedman. Fast sparse regression and classi- fication. International Journal of Forecasting, 28(3):722– 738, 2012

    2012

  3. [11]

    A feasible nonconvex relaxation approach to feature selec- tion

    Cuixia Gao, Naiyan Wang, Qi Yu, and Zhihua Zhang. A feasible nonconvex relaxation approach to feature selec- tion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 25, pages 356–361, 2011

    2011

  4. [12]

    Parameter-efficient fine-tuning with discrete fourier transform

    Ziqi Gao, Qichao Wang, Aochuan Chen, Zijing Liu, Bingzhe Wu, Liang Chen, and Jia Li. Parameter-efficient fine-tuning with discrete fourier transform. arXiv preprint arXiv:2405.03003, 2024

    2024 arXiv

  5. [13]

    Nonlinear image recovery with half-quadratic regularization

    Donald Geman and Chengda Yang. Nonlinear image recovery with half-quadratic regularization. IEEE trans- actions on Image Processing, 4(7):932–946, 1995

    1995

  6. [14]

    Variable selection for gaussian graphi- cal models

    Jean Honorio, Dimitris Samaras, Irina Rish, and Guillermo Cecchi. Variable selection for gaussian graphi- cal models. In Artificial Intelligence and Statistics, pages 538–546. PMLR, 2012

    2012

  7. [15]

    Parameter- efficient transfer learning for nlp, 2019

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Ges- mundo, Mona Attariyan, and Sylvain Gelly. Parameter- efficient transfer learning for nlp, 2019

    2019

  8. [16]

    Lora: low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  9. [17]

    Prefix-tuning: opti- mizing continuous prompts for generation

    Xiang Lisa Li and Percy Liang. Prefix-tuning: opti- mizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021

    2021 arXiv

  10. [18]

    Roberta: a robustly optimized bert pretraining approach, 2019

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man- dar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: a robustly optimized bert pretraining approach, 2019

    2019

  11. [19]

    P-tuning v2: prompt tuning can be comparable to fine-tuning univer- sally across scales and tasks, 2022

    Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: prompt tuning can be comparable to fine-tuning univer- sally across scales and tasks, 2022

    2022

  12. [20]

    Iterative log thresholding, 2013

    Dmitry Malioutov and Aleksandr Aravkin. Iterative log thresholding, 2013

    2013

  13. [21]

    Pissa: principal singular values and singular vectors adaptation of large language models

    Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: principal singular values and singular vectors adaptation of large language models. arXiv preprint arXiv:2404.02948, 2024

    2024 arXiv

  14. [22]

    Highly under- sampled magnetic resonance image reconstruction via ho- motopic ℓ0-minimization

    Joshua Trzasko and Armando Manduca. Highly under- sampled magnetic resonance image reconstruction via ho- motopic ℓ0-minimization. IEEE Transactions on Medical imaging, 28(1):106–121, 2008

    2008

  15. [23]

    Dylora: parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation

    Mojtaba Valipour, Mehdi Rezagholizadeh, Ivan Kobyzev, and Ali Ghodsi. Dylora: parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. arXiv preprint arXiv:2210.07558 , 2022

    2022 arXiv

  16. [24]

    Glue: a multi- task benchmark and analysis platform for natural language understanding

    Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: a multi- task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018

    2018 arXiv

  17. [25]

    Bitfit: simple parameter-efficient fine-tuning for transformer-based masked language-models, 2022

    Elad Ben Zaken, Shauli Ravfogel, and Yoav Gold- berg. Bitfit: simple parameter-efficient fine-tuning for transformer-based masked language-models, 2022

    2022

  18. [26]

    Platon: Pruning large transformer models with upper con- fidence bound of weight importance

    Qingru Zhang, Simiao Zuo, Chen Liang, Alexander Bukharin, Pengcheng He, Weizhu Chen, and Tuo Zhao. Platon: Pruning large transformer models with upper con- fidence bound of weight importance. In International con- ference on machine learning, pages 26809–26823. PMLR, 2022

    2022

  19. [27]

    Adalora: adaptive budget alloca- tion for parameter-efficient fine-tuning

    Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adalora: adaptive budget alloca- tion for parameter-efficient fine-tuning. arXiv preprint arXiv:2303.10512, 2023

    2023 arXiv

  20. [28]

    Structured sparsity optimization with non-convex surrogates of ℓ2,0- norm: A unified algorithmic framework

    Xiaoqin Zhang, Jingjing Zheng, Di Wang, Guiying Tang, Zhengyuan Zhou, and Zhouchen Lin. Structured sparsity optimization with non-convex surrogates of ℓ2,0- norm: A unified algorithmic framework. IEEE Trans- actions on Pattern Analysis and Machine Intelligence , 45(5):6386–6402, 2023

    2023

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.