__ Summary and Contributions__: This paper proposes Composite Score Matching (CSM), a scalable variant of the Score Matching (SM) algorithm. CSM decomposes the original formulation into multiple univariate SM problems, resulting in a parallelizable training procedure for unnormalized autoregressive models. The paper also introduces a new model class AR-CSM, which contains autoregressive models with unnormalized conditionals and expands the usual (normalized) autoregressive models studied before. The authors then experimentally show the scalability and training efficiency (in terms of number of steps) of CSM compared to existing related methods.
-----
Post-rebuttal: The author feedback successfully addressed my concerns and also clarified some misunderstandings I had. I have thus increased my score to 6.

__ Strengths__: THEORY
- This work introduces a very clean extension of the original score matching algorithm specialized to autoregressive (AR) models. The theoretical results mirror those of the original Score Matching paper and are intuitive to understand. Also, their formulation naturally leads to a new model class that is strictly bigger than normalized AR models -- this is important because AR models currently achieve the best density estimation performance across various domains (e.g. images, audio, text).
SIGNIFICANCE & RELEVANCE
- Given the recent growing interest on unnormalized/energy-based models, combined with the impressive performance of AR models, the result of the paper is timely and relevant to the research community.

__ Weaknesses__: EXPERIMENTS
* Section 5.1, Figure 2
- Given that the vertical axes are on completely different scales, it's unclear what this comparison shows. Even though the loss curve for CSM plateaus more quickly than DSM, that doesn't necessarily imply that the trained model achieves better density estimation performance.
* Section 5.2, Figure 3
- The qualitative comparison between PixelCNN++ MLE vs. CSM doesn't seem very convincing, particularly for CelebA sample quality comparison. The paper claims "less shifted colors" when using CSM, but there doesn't seem to be a noticeable difference difference between MLE and CSM for CelebA.
- If CSM also trains a score network on top of h_d = [c_d, x_d], doesn't that have more parameters than the baseline model used to generate c_d (in which case the comparison wouldn't be fair because AR-CSM is parametrized by a strictly bigger network)?
* Section 5.3
- Image denoising can be done by a wide range of generative models. So without any comparison to other methods nor quantitative metric (such as PSNR), the denoising results only seems to serve as a quick sanity check. A more thorough experiment would be necessary to demonstrate the the models trained under CSM are "sufficiently expressive to capture complex distributions and solve difficult tasks."
* Section 6
- The NLL and FID improvements are very small. This again casts a doubt on whether the extra expressivity gained by allowing unnormalized models is actually helpful.

__ Correctness__: - One of the strengths of AR-CSM mentioned in the paper is its expressivity compared to normalized AR models. I agree that traditional AR models constitute a smaller model class than AR-CSM due to normalization. But to what extent is this beneficial? It would be informative to see a toy example that clearly shows how the expressivity gained by AR-CSM leads to better likelihood or sample quality.
- On the flip side, CSM can only be applied to a more limited function class (i.e. AR-CSM) compared to "black-box" methods such as SSM. It's unclear whether the benefits of CSM make up for this drawback (of having a more restricted model class compared to SSM) or not.

__ Clarity__: The paper is clearly written and well-organized. I especially enjoyed reading the theory section, as the presented theorems are easy to follow and very intuitive. The justification behind using CSM was also nice too, except the corresponding experimental results were rather weak. Overall, the writing quality is high.

__ Relation to Prior Work__: Yes. The paper covers several important related papers on topic -- namely Noise Conditional Score Networks (NCSN), Sliced Score Matching (SSM) and Denoising Score Matching (DSM). The presented idea is fundamentally different from the existing work as it focuses on autoregressive models in particular.

__ Reproducibility__: Yes

__ Additional Feedback__: * "CSM" seems refer to both "Composite Score Matching" and "Conditional Score Models". I found this to be slightly confusing.

__ Summary and Contributions__: The paper introduces Composite Score Matching, a new divergence between distributions based on the idea of Composite Score, Score Matching and autoregressive modelling. The authors use autoregressive decomposition of data and model distributions and apply score matching to a collection of one dimensional conditional scores thus constructing a composite score. The authors also theoretically justify this score as a proper learning objective and form a class of expressive autoregressive conditional score models, which are then implemented via autoregressive architectures like MADE and PixelCNN++. They demonstrate effectiveness of these models in a number of tasks, e.g. modelling images, density estimation, image denoising, OOD detection and increasing model flexibility in VAEs.
=== Post rebuttal and discussion update ===
I acknowledge the author rebuttal and the discussion, I maintained my score.

__ Strengths__: Exploring new divergences for generative modelling is a relevant direction for the machine learning community. To my knowledge, applying score matching in the autoregressive setting has not been examined before in the literature. The proposed autoregressive conditional score models are also theoretically justified and empirically demonstrated to improve on the previous autoregressive models.

__ Weaknesses__: I do not see major weaknesses right away, but it seems that ancestral sampling (consequently sampling each element) employed to sample from AR-CSM may take quite a time when data is very high-dimensional. It might also be the case that earlier samples x_{<d} should be drawn accurately to ensure the quality of subsequent samples x_{>d}.

__ Correctness__: It seems to be fine, although detailed proofs are delayed to the supplementary material.

__ Clarity__: Yes, the paper is well-written except for rare typos (e.g. exits — line 206, celebaA — line 267). Also some more comprehensible discussion is in the Appendix.

__ Relation to Prior Work__: Yes, to my knowledge, the prior work is well discussed.

__ Reproducibility__: Yes

__ Additional Feedback__:

__ Summary and Contributions__: This paper proposed the autoregressive score matching model, which models joint distribution in terms of unnormalized scores to improve model capacity.
In order to train the new model, the authors introduce a new divergence between distributions named Composite Score Matching (CSM) which only depends on the derivates of univariate scores, getting rid of computing the trace of the Hessian matrix. Concretely, CSM decomposes the score matching into individual conditionals, resulting a simpler optimization problem. Theoretical proof of convergence has been provided.
Experiments on three benchmarks of images demonstrate that CSM enjoys more stable and efficient training compared with previous score matching methods including denoising score matching and sliced score matching. Moreover, CSM can also be applied to natural image generation, image denoising, out-of-distribution detection and can also be applied to model the posterior distribution in VAEs.
=== Post rebuttal and discussion update ===
I acknowledge the author rebuttal and the discussion, I maintained my score.

__ Strengths__: The proposed CSM method is well-motivated, both theoretically and practically.
Theoretically, CSM improves the flexibility of traditional autoregressive model by removing the normalized score constraints, though satisfying the property of exact density estimation. Moreover, it significantly improve the efficiency of score matching by decomposing it into conditionals at each autoregressive step. Theoretical convergence guarantee has been provided, which is highly appreciated.
Practically, experimental results have illustrated the effectiveness of CSM on several tasks, and showing the efficiency of training compared with other score matching methods.

__ Weaknesses__: NA

__ Correctness__: The claims and empirical methodology are correct.

__ Clarity__: The paper is well-written and easy to follow, given multiple formulations and theorems.

__ Relation to Prior Work__: The related work is well summarized and the the relation and differences of this work compared with previous works have been discussed clearly.

__ Reproducibility__: Yes

__ Additional Feedback__: For the experiments in section 6, CSM is applied to model the posterior distribution in VAEs. Another popular approach to model the posterior distribution is using normalizing flows. Have you conducted experiments to compare CSM with normalizing flows?

__ Summary and Contributions__: This work first introduces composite score matching (CSM), a divergence between distributions, which only depends on the derivatives of univariate log-conditionals. Based the CSM, it introduces an extension of autoregressive models, which removes the normalization constraints for the univariate conditionals. The authors evaluate the proposed method on some simple density estimation and variational inference tasks. The main contribution of this paper is theoretical.

__ Strengths__: pros:
- The idea is natural but novel.
- The proposed method seems to be technically sound.

__ Weaknesses__: cons:
- The empirical results are mostly on synthetic or small-scale dataset.
- There could be some misconfigurations in experimental setup (see my detailed comments).

__ Correctness__: I am concerned about the empirical study.

__ Clarity__: Yes.

__ Relation to Prior Work__: Yes.

__ Reproducibility__: Yes

__ Additional Feedback__: Detailed comments:
- In Figure 1(a), the computational environment need to be specified.
- In Figure 2, what're sigmas for SSM and CSM?
- For image experiment, it would be much better to use the modern convolutional architecture instead of a shallow fully connected network.
- I am confused about the results in Figure 3 and their implications. For example, PixelCNN++ MLE underperforms PixelCNN++ CSM in Figure 3 (at least on MINIST). However, In Appendix D5, the architecture of AR-CSM score network is quite different from PiexelCNN++. In particular, the number of parameters and model capacity really matter, as the PiexelCNN++ only has 40 filters and 1 residual block for MNIST and CIFAR-10. In other words, the constructed PixelCNN++ MLE could be a poor baseline here.
====post rebuttal update====
Thanks for clarifications. It addresses my major concerns, so I increase my score accordingly.