REVIEW 4 major objections 3 minor 12 references
A note on simulation methods for the Dirichlet-Laplace prior
T0 review · 4 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read The MCMC algorithm published with the Dirichlet-Laplace prior draws from the wrong posterior because it samples conditional distributions in the wrong order, and this note provides two corrected samplers.
desk verdict The abstract claims a 'wrong order' Gibbs sampler, but that mechanism cannot break a Gibbs sampler; the packet also has the wrong full text, so there is nothing to referee yet. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the joint posterior distribution of the latent variables in the hierarchical Dirichlet-Laplace prior and the Gibbs-sampler update order used to explore it. The mechanism of the fix is twofold: first, rearranging the conditional draws so that the chain has the correct joint posterior as its stationary distribution; second, reparametrizing the prior through an equivalent formulation that permits a valid sampling scheme in a different order.
What would settle it
Simulate data from the generative model of the DL prior, run both the original sampler and the first corrected sampler to many iterations, and compare their empirical joint distributions for the latent variables against the closed-form full conditionals; if the original sampler's draws match the true joint posterior, the claimed ordering error cannot be real.
Extended reading notes
Core claim
The paper establishes that the Gibbs-style sampling scheme for the Dirichlet-Laplace prior, as originally published, draws from the conditional distributions in a sequence that is not a valid sampler for the joint posterior of all latent variables. The authors show where the ordering fails and supply two valid alternatives: a direct correction that restores the correct conditional-update order, and a second algorithm built on an alternative yet equivalent formulation of the prior. Both are stated to work in the conditionally Gaussian setting, and the corrigendum explicitly leaves the theoretical results of the 2015 paper unaffected.
Load-bearing premise
The second proposed sampler is only valid if the 'alternative, yet equivalent' formulation of the DL prior is equivalent in the joint latent structure the sampler exploits, not just in its marginal form.
Editorial extensions
If this is right
- Implementations of the original DL sampler should be replaced by the corrected version whenever draws from the joint posterior of the latent variables are needed.
- Applied conclusions based on the original sampler may not reflect the true DL posterior, so re-running analyses with a corrected sampler is the direct way to check sensitivity.
- The theoretical posterior-concentration and shrinkage results in the 2015 paper are not affected by the sampling-order bug.
- The alternative-formulation algorithm gives practitioners a second, equivalently valid sampler for the same conditionally Gaussian setting.
Reading between the lines
- If the original sampler has been widely used in applications, reported uncertainty intervals for shrinkage or selection quantities could be systematically off; a re-analysis of a published dataset with the corrected sampler would quantify the impact.
- The same failure mode—updating conditionals in an order that breaks the joint target—could exist in other hierarchical shrinkage samplers, so verifying that every update is a full conditional of the same joint posterior is a cheap audit for such code.
- The equivalence claim suggests a concrete diagnostic: run both corrected samplers to approximate convergence and compare their empirical joint distributions; any mismatch would reveal that the alternative formulation is not truly equivalent at the joint level.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper is advertised as a corrigendum note to Bhattacharya et al. (2015), claiming that the original MCMC algorithm for the Dirichlet-Laplace (DL) prior 'samples from conditional distributions in the wrong order' and therefore does not correctly target the joint posterior of the latent variables. It promises two fixes: a correction to the original algorithm and a new algorithm based on 'an alternative, yet equivalent, formulation of the prior.' The abstract further asserts that the theoretical results of the 2015 paper are unaffected. The submitted full text, however, is not this manuscript at all but an unrelated paper titled 'Chart-CoCa: Self-Improving Chart Understanding of Vision LMs via Code-Driven Synthesis and Candidate-Conditioned Answering' (arXiv:2508.11975). Consequently, the actual content of the note—its derivations, algorithms, proofs, or numerical checks—is absent from the submission.
Significance. If the central claim were established, the note would have substantive practical importance: the DL prior is widely used in high-dimensional Bayesian sparse regression, and a genuine failure of the published sampler would affect the validity of many posterior summaries obtained with that algorithm. A precise, verifiable correction with a correct stationary distribution would be a useful service to the community. However, the submitted text contains no derivations, no algorithm listings, no proofs, no numerical experiments, and no code. Moreover, the mechanism stated in the abstract is, as phrased, inconsistent with standard Gibbs sampling theory: for correctly specified full conditionals, update order does not change the stationary distribution. The significance assessment therefore rests entirely on an unverifiable abstract and on an internally questionable explanation.
major comments (4)
- [Abstract] The stated mechanism cannot be correct as written. In a Gibbs sampler, if every draw is from the true full conditional of the corresponding latent variable given the current values of all other variables, then the composition of such updates leaves the joint target invariant regardless of update order. Thus 'samples from conditional distributions in the wrong order' cannot, by itself, explain a failure to sample from the joint posterior. The note must identify a specific misspecified conditional, for example a step that conditions on an outdated value, omits conditioning on the response, or uses a conditional derived from a marginal instead of the joint model. The abstract provides no such identification, and no full text is available to supply it.
- [Full text] The submitted full text is not this manuscript. After the abstract of the DL-prior note, the document turns into a completely different paper, 'Chart-CoCa: Self-Improving Chart Understanding of Vision LMs via Code-Driven Synthesis and Candidate-Conditioned Answering,' with different authors, different subject, and a different arXiv identifier. None of the promised content of the note—Algorithm 1 of Bhattacharya et al., the corrected sampler, the alternative formulation, or any invariance proof—appears in the submission. This makes the central claim impossible to check and is a fundamental submission defect.
- [Abstract] The second proposed solution depends on an 'alternative, yet equivalent, formulation of the prior.' For a Gibbs sampler, the replacement must be equivalent as a joint law over all latent variables that enter the sampling scheme, not merely as the marginal law of the regression coefficient. If the equivalence holds only marginally, the induced full conditionals need not coincide and the new sampler could target a different posterior. The abstract does not even state the variables involved, let alone demonstrate joint equivalence, so this load-bearing assumption is completely unsupported.
- [Abstract] No evidence whatsoever is provided that the corrected sampler actually has the claimed stationary distribution. The abstract promises 'a correction to the original algorithm' but offers no derivation of the full conditionals, no proof of invariance, and no numerical experiment comparing the original and corrected samplers. For a corrigendum whose entire purpose is to change the target distribution of a published MCMC algorithm, the absence of any verification is a decisive gap.
minor comments (3)
- [Abstract] The phrase 'wrong order' is ambiguous and should be replaced by a precise statement of which conditional distribution is misspecified and at which step of the original Algorithm 1.
- [Abstract] The note would benefit from a brief description of the latent-variable structure of the DL sampler so that readers can understand the claimed error without consulting the 2015 paper.
- [General] If the corrected algorithm is to be usable, the submission should include pseudocode for both the corrected original sampler and the alternative-formulation sampler, as well as code or a reproducible numerical check when practical.
Circularity Check
No circularity identified from the abstract; the correction is benchmarked against the external DL posterior, not against its own outputs.
full rationale
The available material for arXiv:2508.11982 is an abstract; the full text supplied in the prompt is an unrelated paper on chart understanding (Chart-CoCa) and contains no equations from the Dirichlet-Laplace note. For the note's actual claim, the target is the joint posterior distribution induced by the Dirichlet-Laplace prior, an external probabilistic object specified independently of the MCMC scheme. Saying the original sampler 'samples from conditional distributions in the wrong order' asserts a mismatch between the sampler and that external target; even if the asserted mechanism were wrong, the mismatch would be a correctness error, not a circularity. The two proposed algorithms are described as corrections and as an alternative 'yet equivalent' prior formulation, but no derivation is available to inspect, so no equation-level reduction of a prediction to an input can be exhibited. The authors include members of the original 2015 paper, but this is a self-correction rather than load-bearing self-citation: the standard for correctness (the DL posterior) is not defined in terms of the present paper's outputs. The conceptual concern about Gibbs update order raised by the skeptic is a correctness-risk point and, per the circularity rules, does not constitute circularity without a demonstrated reduction.
Assumptions & free parameters
assumptions (3)
- domain assumption The alternative formulation of the Dirichlet-Laplace prior is exactly equivalent to the original prior
- domain assumption The corrected Gibbs sampler converges to the intended joint posterior under standard MCMC regularity conditions
- domain assumption The conditionally Gaussian setting of the original paper is preserved
Cite this review
Pith. "Pith review of A note on simulation methods for the Dirichlet-Laplace prior." pith.science (2026). https://pith.science/paper/3TAGODUN
@misc{pith2026250811982,
author = {Pith},
title = {Pith review of: A note on simulation methods for the Dirichlet-Laplace prior},
year = {2026},
howpublished = {\url{https://pith.science/paper/3TAGODUN}},
note = {Machine review of arXiv:2508.11982}
}
read the original abstract
Bhattacharya et al. (2015, Journal of the American Statistical Association 110(512): 1479-1490) introduce a novel prior, the Dirichlet-Laplace (DL) prior, and propose a Markov chain Monte Carlo (MCMC) method to simulate posterior draws under this prior in a conditionally Gaussian setting. The original algorithm samples from conditional distributions in the wrong order, i.e., it does not correctly sample from the joint posterior distribution of all latent variables. This note details the issue and provides two simple solutions: A correction to the original algorithm and a new algorithm based on an alternative, yet equivalent, formulation of the prior. This corrigendum does not affect the theoretical results in Bhattacharya et al. (2015).
Reference graph
Works this paper leans on
-
[37]
Zirui Wang, Mengzhou Xia, Luxi He, Howard Chen, Yitao Liu, Richard Zhu, Kaiqu Liang, Xindi Wu, Haotian Liu, Sadhika Malladi, Alexis Chevalier, Sanjeev Arora, and Danqi Chen. 2024. CharXiv: Charting Gaps in Realistic Chart Understanding in Multimodal LLMs.���� abs/2406.18521 (2024). doi:10.48550/ARXIV.2406.18521 arXiv:2406.18521
-
[38]
Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. �������� �� ������ ����������� ���������� �������35 (2022), 24824–24837
2022
-
[39]
Yixuan Weng, Minjun Zhu, Fei Xia, Bin Li, Shizhu He, Shengping Liu, Bin Sun, Kang Liu, and Jun Zhao. 2023. Large Language Models are Better Reasoners with Self-Verification. In�������� �� ��� ����������� ��� ������������� ������������ ����� ���� . 2550–2575
work page 2023
-
[40]
Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al
-
[41]
Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, et al . 2024. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning. ����� �������� ����������������(2024)
work page 2024
-
[42]
Guowei Xu, Peng Jin, Hao Li, Yibing Song, Lichao Sun, and Li Yuan. 2024. LLaVA- CoT: Let Vision Language Models Reason Step-by-Step. arXiv:2411.10440 [cs.CV] https://arxiv.org/abs/2411.10440
arXiv 2024
-
[43]
Zhengzhuo Xu, Sinan Du, Yiyan Qi, Chengjin Xu, Chun Yuan, and Jian Guo. 2023. Chartbench: A benchmark for complex visual reasoning in charts. ����� �������� ����������������(2023)
work page 2023
-
[44]
Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason E Weston. 2024. Self-Rewarding Language Models. In ���������� ������������� ���������� �� ������� ��������. https: //openreview.net/forum?id=0NphYCmgua
work page 2024
Show all 12 references
-
[45]
Simon Zhai, Hao Bai, Zipeng Lin, Jiayi Pan, Peter Tong, Yifei Zhou, Alane Suhr, Saining Xie, Yann LeCun, Yi Ma, and Sergey Levine. 2024. Fine- Tuning Large Vision-Language Models as Decision-Making Agents via Re- inforcement Learning. In �������� �� ������ ����������� ��������...
2024
-
[46]
Meishan Zhang, Gongyao Jiang, Shuang Liu, Jing Chen, and Min Zhang. 2024. LLM-Assisted Data Augmentation for Chinese Dialogue-Level Dependency Pars- ing. ������������� �����������50, 3 (Sept. 2024), 867–891. doi:10.1162/coli_a_ 00515
2024 doi
-
[47]
Xuanle Zhao, Xianzhen Luo, Qi Shi, Chi Chen, Shuo Wang, Wanxiang Che, Zhiyuan Liu, and Maosong Sun. 2025. ChartCoder: Advancing Multimodal Large Language Model for Chart-to-Code Generation
2025
-
[2019]
����� �������� ����������������(2019)
Huggingface’s transformers: State-of-the-art natural language processing. ����� �������� ����������������(2019)
2019
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.