Breaking the Factorization Barrier in Diffusion Language Models

Anji Liu; Benjie Wang; Guy Van den Broeck; ian Li; Rose Yu; Zilei Shao

REVIEW 4 major objections 5 minor 1 cited by

Breaking the Factorization Barrier in Diffusion Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A probabilistic-circuit prior, multiplied into the denoising distribution, removes the independence bottleneck that makes parallel decoding in diffusion language models incoherent.

desk verdict CoDD is a real, cheap, tractable trick that improves diffusion LMs on math-heavy benchmarks, but the "breaking the factorization barrier" claim outruns the evidence because the PC prior is static and the headline numbers lack error bars. read the letter →

arxiv 2603.00045 v3 pith:6J7IX7UU submitted 2026-02-09 cs.LG cs.AI

Ian Li , Zilei Shao , Benjie Wang , Rose Yu , Guy Van den Broeck , Anji Liu This is my paper

classification cs.LGcs.AI

keywords diffusionlanguagemodelsfactorizationbarrierprobabilisticcircuitstractableinferenceparalleldecodingfew-stepgenerationCoupledDiscretereasoningbenchmarks

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

Diffusion language models predict many tokens at once, but their output distribution is forced to treat those tokens as independent, which makes parallel decoding incoherent. This paper argues that the bottleneck is not the neural backbone's capacity but a structural misspecification: the output layer is restricted to a fully factorized distribution because explicit joint parameterization is too large. It proposes CoDD, which multiplies the backbone's factorized token potentials by a lightweight Probabilistic Circuit prior and normalizes the product, giving a compact yet genuinely joint distribution with exact tractable inference. On frozen LLaDA and Dream models, CoDD improves reasoning accuracy on MATH500, GSM8K, GPQA, and MBPP, prevents the accuracy collapse that normally appears at 64 diffusion steps, and costs about 3 GPU hours—under 2% of a reinforcement-learning baseline—to train. If correct, it shows the factorization barrier is removable by restructuring the output rather than enlarging the model.

What carries the argument

The central mechanism is the product composition of Eq. (5): a Probabilistic Circuit (a tractable sum/product network over variables, here an HMM with 1024 hidden states) supplies the joint structure $p_\omega(x_0)$, while the frozen Transformer supplies factorized potentials $p_\theta(x_0)$. Decomposability—the property that product nodes split variables into disjoint scopes—makes the normalization constant an instance of independent virtual evidence, computable by one feedforward pass; the same structure makes exact conditional sampling possible. Adaptive activation ($\gamma$) and dynamic windowing keep the static prior active only when it is well specified.

What would settle it

Hold out a domain whose local dependency structure is absent from the PC's training corpus—for example, code in a different language or non-English text—and run CoDD with the same frozen prior and threshold at 128 steps; if accuracy does not improve over the factorized baseline, the transfer claim is refuted. Alternatively, measure conditional log-likelihood on ground-truth continuations at mask ratios where $\gamma$ activates; if the product distribution is not better there, the threshold is masking a failure.

Watch

Extended reading notes

Core claim

The paper's central claim is that the denoising distribution in masked diffusion language models can be replaced by a normalized product of the Transformer's factorized per-token potential and a decomposable Probabilistic Circuit, $\hat{p}_{\theta,\omega}(x_0|x_t) \propto p_\omega(x_0) \cdot p_\theta(x_0)$. Because the circuit is decomposable, the partition function is a bottom-up pass in time linear in the circuit, so the joint family is far more expressive than factorized output yet remains compact. Training freezes the backbone and fits only the circuit's parameters on precomputed logits. The paper reports consistent gains in block and full diffusion settings—e.g., +5.0 accuracy on MATH500 for LLaDA at 256 steps, +10.84 on GSM8K

Load-bearing premise

The load-bearing premise is that a single static dependence model, fit once to a fixed math question–solution corpus and gated by a threshold tuned on validation data, captures the local token-dependency structure of all four evaluation benchmarks closely enough that, whenever the circuit is activated, the product distribution is closer to the true joint than the fully factorized baseline.

Editorial extensions

If this is right

CoDD removes the trade-off between parallel speed and coherence: standard decoding heuristics improve on all four benchmarks, so models need not fall back to sequential generation to avoid incoherent mixtures.
Few-step generation no longer collapses: at 64 steps CoDD lifts Dream's GSM8K accuracy from 34.0% to 56.4%, cutting inference cost substantially for the same quality.
The factorization barrier is a property of the output distribution class, not of the backbone; the same frozen model gets better without further Transformer training.
The gains are cheap: PC training uses about 3 GPU hours (under 2% of an RL baseline's budget) and inference adds only a few percent latency, so CoDD is a plug-in booster for existing diffusion LMs.
CoDD is paradigm-agnostic: it works with block diffusion, full diffusion, and multiple unmasking heuristics (random, low-confidence, margin, entropy).

Reading between the lines

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

Editorial inference: CoDD's recipe only needs frozen logits and a tractable prior, so the same product construction could improve other factorized parallel decoders—multi-token speculative decoding, non-autoregressive machine translation, or masked image/audio models—wherever a static dependency prior is available.
Editorial inference: the adaptive threshold γ is a workaround for the static prior's mismatch at high noise; a noise-conditioned circuit (which the paper sketches as latent-space or parameter-space modulation) is the natural next step and would test whether the reported gains extend beyond low-noise regimes.
Editorial inference: if the static PC transfers across domains as reported, then the marginal value of larger backbones for parallel generation may be smaller than the value of better output structure; one could test this by comparing CoDD's gains against scaling the backbone's parameters or steps.
Editorial inference: because the prior is trained on MathInstruct, the cleanest out-of-distribution test is to apply CoDD to a non-mathematical reasoning task without retraining the PC; failure there would indicate the dependency structure, not the mechanism, drives the gains.

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.

Desk Editor's Note

read the letter

Quick take: this is a legitimate and potentially useful paper, but the big claim is bigger than the evidence. The core idea — replace the fully factorized denoising distribution with the product of a decomposable probabilistic circuit and the backbone's factorized logits — is mathematically sound and clearly presented. The partition function computation via virtual evidence is correct, and the training cost of about 3 GPU hours on frozen logits is genuinely low. The empirical gains on math reasoning at few steps are large and consistent: Dream+CoDD on GSM8K at 64 steps goes from 34.0 to 56.4, and the qualitative examples show the baseline degenerating while CoDD stays coherent. That is a real effect.

What is actually new is the application of this group's PC-times-factorized-potential toolbox to discrete diffusion LMs, with dynamic windowing and adaptive activation. The novelty is incremental relative to Discrete Copula Diffusion and the tractable steering work, but the paper cites those and the engineering is not trivial. The reliance on Theorem D.2 from Liu et al. (2024) is a proper citation, not a problem.

The soft spots are in the evidence, not the math. First, every number in Tables 1 and 2 is a single run with no seeds or error bars. For accuracy differences of a few points, that is not enough to separate signal from noise. Second, the adaptive activation threshold γ and temperature τ are swept on the same benchmarks used for the main results (Appendix B), and the chosen defaults are not stated. That makes the headline gains look selected. Third, the PC is trained on MathInstruct solution segments; the four evaluation benchmarks are all reasoning/code tasks, and the biggest wins are on the math ones. The stress-test concern — that CoDD is mostly a task-specific stylistic prior that suppresses repetition rather than a general fix for the factorization barrier — is not resolved by the paper. The paper itself concedes the static PC collapses time-varying structures and needs a threshold to deactivate it at high noise. Fourth, the "negligible overhead" claim is contradicted by Table 3: CoDD+AO adds 26.9% on LLaDA at 128 steps, which is not negligible.

The math is coherent enough that I believe the method does what it says: it samples from a more expressive joint family. The open question is how much of the reported gain is a benchmark-specific artifact. The paper would be stronger with error bars, a held-out hyperparameter selection procedure, and an ablation where the PC is trained on a non-math corpus to show transfer.

Bottom line: worth serious peer review, not a desk reject. It is a solid incremental contribution from a group that knows this area. I would send it to a careful referee, with the expectation that the empirical section needs work before acceptance. For a reading group, it is a good paper to discuss because the method is simple, the claims are testable, and the gap between the math and the evidence is instructive.

Referee Report

4 major / 5 minor

Summary. The paper proposes Coupled Discrete Diffusion (CoDD), a method for discrete diffusion language models that replaces the fully factorized denoising distribution p_theta(x0|xt) with a product of the backbone's factorized potentials and a learned Probabilistic Circuit (PC) prior p_omega(x0), normalized by a partition function Z (Eq. 5). The authors show that Z is tractable when the PC is decomposable with factorized neural potentials, and they train the PC on frozen backbone logits using the diffusion objective (Eq. 7). At inference, CoDD samples from the joint product with approximate temperature scaling, uses adaptive activation controlled by a threshold gamma, and applies dynamic windowing for full-length diffusion. Experiments on LLaDA and Dream report accuracy gains on MATH500, GSM8K, GPQA, and MBPP, especially at few diffusion steps, with a claimed training cost of about 3 GPU hours. The paper argues that this breaks the 'factorization barrier' that limits parallel token prediction in diffusion LMs.

Significance. If the empirical claims hold, CoDD is a lightweight, plug-and-play enhancement to existing discrete diffusion LMs that improves coherence and few-step generation with negligible inference overhead. The theoretical observation that a decomposable PC prior can be multiplied with factorized neural potentials and normalized exactly in linear time is clean and plausible, and it gives the paper a solid conceptual core. The method is modular and does not require backpropagating through the backbone, which is a practical strength. However, the empirical evidence as presented is not yet at the standard needed to support the strong statements in the title and abstract: all accuracy tables are single-run point estimates, hyperparameters such as gamma and tau appear to be tuned on the evaluation benchmarks, and the central conceptual claim is weakened by the static, context-independent nature of the PC prior.

major comments (4)

[§6.3, Tables 1, 2, 6] All reported accuracies are single-run point estimates; no seeds, error bars, or significance tests are given. Many claimed gains are within the range of run-to-run noise in such evaluations, e.g., Table 1 LLaDA/Margin MATH500 256 steps +0.20, GSM8K 128 steps +0.30, and Table 2 Dream/Entropy MBPP 256 steps +0.20. The headline improvements (+22.44 GSM8K at 64 steps, +10.84 at 128 steps) may be real, but the text's stronger claim that CoDD yields 'robust gains across all settings' is not supported without variance estimates. Please report multiple seeds for the main configurations and provide confidence intervals or paired tests.
[§5.3 and Appendix B] The adaptive activation threshold gamma and temperature tau are ablated on MATH500 and MBPP (Tables 4–5), but the paper never states how the values used in Tables 1–2 were chosen. If gamma/tau were selected by scanning on the test benchmarks, then the 'zero-shot setting' claim in Sec. 6.1 is violated and the reported numbers are optimistically biased. Please specify a held-out validation protocol and report performance under fixed hyperparameters across all tasks.
[§4.1, Eq. (5), §5.3] The structural prior p_omega is static and unconditional; the only context-dependent factor in Eq. (5) is the fully factorized p_theta(x0). Consequently, CoDD cannot represent joint dependencies that vary with the prompt or with the noise level; it can only inject a fixed, corpus-average dependency structure. This is acknowledged in §5.3 ('a static PC... collapses these time-varying structures into a single global distribution'). Since the PC is trained on MathInstruct, which overlaps in domain with GSM8K/MATH, the reported gains may reflect a task-specific stylistic prior rather than a general resolution of the factorization barrier. Please either temper the title/abstract claim or add out-of-domain experiments (e.g., a PC trained on general text applied to math/code benchmarks) to separate these explanations.
[§5.1, Algorithm 1/2] Algorithm 1 line 10 instructs to sample from p^{1/tau}_theta(X0) * p_omega(X0), but §5.1 states that exact temperature scaling of the product is #P-hard and instead proposes latent-variable sampling or any-order autoregressive sampling. The main experiments never state which approximation is used. If latent-variable sampling is used, the actual sampled distribution is a mixture of temperature-sharpened conditionals weighted by the unsharpened latent posterior, not the sharpened product; if any-order autoregressive is used, the number of PC queries and the ordering heuristic need specification. Please clarify the exact decoding procedure for each reported row.

minor comments (5)

[Appendix B] Typos: 'adpot' -> 'adopt', 'temperatire' -> 'temperature', 'procotol' -> 'protocol', 'thoughout' -> 'throughout'.
[Eqs. (2), (4), (5)] Notation is inconsistent: p_phi and p_theta are both used for the backbone's factorized distribution. Eq. (5) reuses p_theta for the neural potentials while the network is called f_phi; please clarify the relationship among phi, f_phi, and theta.
[Algorithm 2 / Appendix D] The symbol W is used both for the set of PC windows in Algorithm 2 and for virtual evidence weights in Appendix D. Rename one of them to avoid confusion.
[Table 5] The table title says 'Dream Model with Block Diffusion,' but the main text (Sec. 6.1) describes Dream as using full diffusion and LLaDA as using block diffusion. Please check whether this is a typo or whether Table 5 reports a different setting.
[Appendix D] Theorem D.2 is stated for 'smooth and decomposable' PCs, but smoothness is not defined in the main text or Appendix C. Please state whether the HMM PC used in experiments is smooth and whether smoothness is needed for the partition-function computation.

Circularity Check

0 steps flagged · score 2.0 of 10

No central circularity: the joint product in Eq. (5) is a constructive model family and the reported benchmark gains are external; only in-house citations (Theorem D.2, anemone) are minor and non-load-bearing.

full rationale

The load-bearing claim is that replacing the fully factorized denoising distribution by the product p_omega(x0)*p_theta(x0) in Eq. (5) yields a tractable joint family. This is a constructive model choice: the PC p_omega is a genuine joint distribution, and the partition function Z is computed by the decomposability argument in Section 4.2 / Appendix D, not by fitting. Expressiveness follows by construction (factorized distributions are the special case p_omega uniform), so there is no hidden equivalence between the derivation and its inputs. The PC is trained by Eq. (7) on frozen backbone logits and MathInstruct, while the headline numbers are accuracies on MATH500/GSM8K/GPQA/MBPP, not the training objective; the gamma/tau sweeps in Appendix B are hyperparameter choices, not fitted constants renamed as predictions. The paper does cite in-house results (Theorem D.2 from Liu et al. 2024, and the anemone optimizer from Liu et al. 2025b), but these are published, checkable, standard tractability/optimization tools and do not smuggle in the factorization-barrier conclusion. Section 5.3 explicitly acknowledges the static-PC limitation, which is a generalization concern rather than circularity. No 'prediction' reduces to a fitted parameter by equation identity.

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

The central claim rests on standard tractable-probabilistic-model assumptions (decomposability and smoothness), on the unproven but plausible expressivity of a fixed HMM prior, and on the transferability of a static PC across domains. The main free parameters introduced by the method are the activation threshold, temperature, HMM size, window size, and block size; their values or selection procedures are only partially specified.

free parameters (5)

adaptive activation threshold γ = not stated (ablated: 0.3–0.8)
Controls when the PC is activated; chosen via validation ablations in Appendix B (Tables 4–5) and affects all main results.
PC temperature τ = not stated (ablated: 0.1, 0.2)
Temperature applied to latent/conditional sampling; tuned per task in Appendix B.
PC hidden state size N = 1024
HMM hidden state count for the PC prior; set by hand, controls expressivity and compute.
PC window size W = not specified
Required by dynamic windowing for full diffusion; no value or selection procedure is given in the main text.
block size Lb = 32
Fixed block length for block diffusion; hand-chosen in Section 5.2.

assumptions (4)

standard math The PC pω is smooth and decomposable, so Theorem D.2 (Liu et al. 2024) gives exact linear-time partition-function computation under independent virtual evidence.
Section 4.2 and Appendix D rely on this external theorem without proving it; the HMM instantiation is claimed to satisfy the required structural constraints.
domain assumption Multiplying a PC prior by a factorized neural potential yields a distribution family expressive enough to capture the relevant inter-token dependencies.
Section 4.1 posits a 'base-and-refine' product composition; no formal expressivity bound is given for the specific HMM prior.
domain assumption A single static PC trained on the frozen backbone's logits is a sufficient structural prior across the diffusion trajectory once the adaptive activation threshold is applied.
Section 5.3 acknowledges a static PC averages over time-varying dependency structures and introduces γ to compensate; transfer to all target tasks is assumed.
domain assumption The frozen Transformer's factorized logits pθ provide valid univariate potentials for the product distribution.
Training objective Eq. (7) and sampling in Algorithms 1–2 use pθ as fixed potentials; this is true by construction of the backbone output.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Breaking the Factorization Barrier in Diffusion Language Models." pith.science (2026). https://pith.science/paper/6J7IX7UU

@misc{pith2026260300045,
  author       = {Pith},
  title        = {Pith review of: Breaking the Factorization Barrier in Diffusion Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6J7IX7UU}},
  note         = {Machine review of arXiv:2603.00045}
}

read the original abstract

Diffusion language models theoretically allow for efficient parallel generation but are practically hindered by the ``factorization barrier'': the assumption that simultaneously predicted tokens are independent. This limitation forces a trade-off: models must either sacrifice speed by resolving dependencies sequentially or suffer from incoherence due to factorization. We argue that this barrier arises not from limited backbone expressivity, but from a structural misspecification: models are restricted to fully factorized outputs because explicitly parameterizing a joint distribution would require the Transformer to output a prohibitively large number of parameters. We propose Coupled Discrete Diffusion (CoDD), a hybrid framework that breaks this barrier by replacing the fully-factorized output distribution with a lightweight, tractable probabilistic inference layer. This formulation yields a distribution family that is significantly more expressive than standard factorized priors, enabling the modeling of complex joint dependencies, yet remains compact enough to avoid the prohibitive parameter explosion associated with full joint modeling. Empirically, CoDD seamlessly enhances diverse diffusion language model architectures with negligible overhead, matching the reasoning performance of computationally intensive Reinforcement Learning baselines at a fraction of the training cost. Furthermore, it prevents performance collapse in few-step generation, enabling high-quality outputs at significantly reduced latencies. Code available at: https://github.com/liuanji/CoDD.

Figures

Figures reproduced from arXiv: 2603.00045 by the authors.

**Figure 1.** Motivation and Intuition of CoDD. Left: Illustration of the misspecification gap. The plot reports the perplexity of LLaDA (Nie et al., 2025) on the MathInstruct validation set across varying mask ratios. Curve (a) Sequential generation represents the ideal baseline (i.e., the true joint distribution learned by the model). When restricted to (b) One-step generation, the independence assumption causes significant per… view at source ↗

**Figure 2.** Conditional Likelihood on Ground Truth. This figure illustrates the conditional log-likelihood (CLL) of the CoDD and Dream models evaluated directly on ground truth question-answer pairs from the full Math Instruct dataset. tion strategies adopted for each setting. Block Diffusion. In this regime, the sequence is generated in fixed-size segments (e.g., Lb = 32), enforcing a semi-autoregressive order where all token… view at source ↗

**Figure 3.** Left: Performance Comparison on MATH500 vs. RL Baselines for 256/128/64 diffusion steps with a fixed generation length of 512. d-GRPO denotes diffu-GRPO and is reproduced with Zhao et al. (2025)’s codebase. Methods marked with superscripts (d1† , wd1‡ , d2⋄ ) are reported from prior work (Zhao et al., 2025; Tang et al., 2025; Wang et al., 2025). Right: Training-time cost in GPU hours (↓ is better) for diffu-GRPO and… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Qualitative comparison: Exponent Equation (64 steps). At reduced sampling budgets, the baseline model suffers from severe mode collapse (repeating tokens). CoDD effectively steers the generation to recover coherent reasoning. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

**Figure 5.** Figure 5: Algebraic simplification at low compute (64 steps). The baseline model correctly derives the intermediate terms but fails at the final combination step, effectively ”forgetting” the remaining terms (3x − 6). CoDD maintains coherence through the multi-step reasoning to …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

Tensor-Train Joint Modeling for Few-Step Discrete Diffusion
cs.LG 2026-07 conditional novelty 7.0 of 10

Tensor-Train (and CPD) parameterization of the MDM clean conditional removes parallelization bias and substantially improves few-step text and molecule generation via lightweight fine-tuning.

Reference graph

Works this paper leans on

6 extracted references · 1 linked inside Pith · cited by 1 Pith paper

[1]

For leaf nodes, we evaluate the probability of the observed value (or 1 if the variable is missing)

Forward Pass (Bottom-Up):We first compute the likelihood of the evidence at every node. For leaf nodes, we evaluate the probability of the observed value (or 1 if the variable is missing). These values are propagated upward to the root, where nodencomputes the likelihood of the evidence restricted to its scope, denoted asL n(xobs)
[2]

evidence-adjusted

Backward Pass (Top-Down):We sample strictly from the missing variables xmiss using a logic similar to the unconditional case, but with “evidence-adjusted” weights: • At aSum Node n, instead of using the raw parameters ωn,c, we sample a child c using the posterior probability given the evidence: p(c|n,x obs)∝ω n,c ·L c(xobs). • At aProduct Node, we recurse...
[6]

evidence

:=pθ(xi 0 |x t), we observe that computing Z is exactly equivalent to computing the probability of soft evidence P(W) (Eq. (8)) where the “evidence” is provided by the neural network’s factorized logits. Tractability.To establish the efficiency of this computation, we reuse Theorem 1 from Liu et al. (2024). Theorem D.2.For any smooth and decomposable PC p...

2024
[2022]

content/CVPR2022/html/Chang_MaskGIT_ Masked_Generative_Image_Transformer_ CVPR_2022_paper.html

URL https://openaccess.thecvf.com/ 2GPU-hour comparison was measured on NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. content/CVPR2022/html/Chang_MaskGIT_ Masked_Generative_Image_Transformer_ CVPR_2022_paper.html. Chiu, J. and Rush, A. M. Scaling hidden markov language models. InProceedings of the 2020 Conference on Empir- ical Methods in Natural Lan...

arXiv 2020
[2024]

You are a helpful assistant

URL https://openreview.net/forum? id=NSIVHTbZBR. Liu, A., Broadrick, O., Niepert, M., and Van den Broeck, G. Discrete copula diffusion. InThe Thirteenth International Conference on Learning Representations, 2025a. Liu, A., Shao, Z., and den Broeck, G. V . Rethinking probabilistic circuit parameter learning, 2025b. URL https://arxiv.org/abs/2505.19982. Lou...

arXiv 2024
[2025]

Kisa, D., Van den Broeck, G., Choi, A., and Darwiche, A

URL https://openreview.net/forum? id=DjJmre5IkP. Kisa, D., Van den Broeck, G., Choi, A., and Darwiche, A. Probabilistic sentential decision diagrams. InKR, 2014. Lightman, H., Kosaraju, V ., Burda, Y ., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., and Cobbe, K. Let’s verify step by step, 2023. URL https: //arxiv.org/abs/2305.20...

arXiv 2014

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

For leaf nodes, we evaluate the probability of the observed value (or 1 if the variable is missing)

Forward Pass (Bottom-Up):We first compute the likelihood of the evidence at every node. For leaf nodes, we evaluate the probability of the observed value (or 1 if the variable is missing). These values are propagated upward to the root, where nodencomputes the likelihood of the evidence restricted to its scope, denoted asL n(xobs)

[2] [2]

evidence-adjusted

Backward Pass (Top-Down):We sample strictly from the missing variables xmiss using a logic similar to the unconditional case, but with “evidence-adjusted” weights: • At aSum Node n, instead of using the raw parameters ωn,c, we sample a child c using the posterior probability given the evidence: p(c|n,x obs)∝ω n,c ·L c(xobs). • At aProduct Node, we recurse...

[3] [6]

evidence

:=pθ(xi 0 |x t), we observe that computing Z is exactly equivalent to computing the probability of soft evidence P(W) (Eq. (8)) where the “evidence” is provided by the neural network’s factorized logits. Tractability.To establish the efficiency of this computation, we reuse Theorem 1 from Liu et al. (2024). Theorem D.2.For any smooth and decomposable PC p...

2024

[4] [2022]

content/CVPR2022/html/Chang_MaskGIT_ Masked_Generative_Image_Transformer_ CVPR_2022_paper.html

URL https://openaccess.thecvf.com/ 2GPU-hour comparison was measured on NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. content/CVPR2022/html/Chang_MaskGIT_ Masked_Generative_Image_Transformer_ CVPR_2022_paper.html. Chiu, J. and Rush, A. M. Scaling hidden markov language models. InProceedings of the 2020 Conference on Empir- ical Methods in Natural Lan...

arXiv 2020

[5] [2024]

You are a helpful assistant

URL https://openreview.net/forum? id=NSIVHTbZBR. Liu, A., Broadrick, O., Niepert, M., and Van den Broeck, G. Discrete copula diffusion. InThe Thirteenth International Conference on Learning Representations, 2025a. Liu, A., Shao, Z., and den Broeck, G. V . Rethinking probabilistic circuit parameter learning, 2025b. URL https://arxiv.org/abs/2505.19982. Lou...

arXiv 2024

[6] [2025]

Kisa, D., Van den Broeck, G., Choi, A., and Darwiche, A

URL https://openreview.net/forum? id=DjJmre5IkP. Kisa, D., Van den Broeck, G., Choi, A., and Darwiche, A. Probabilistic sentential decision diagrams. InKR, 2014. Lightman, H., Kosaraju, V ., Burda, Y ., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., and Cobbe, K. Let’s verify step by step, 2023. URL https: //arxiv.org/abs/2305.20...

arXiv 2014