acceptodds
Under review as a conference paper at ICLR 2027

DUEL: Exact Likelihood for Masked Diffusion Reveals Improved Pre-Training Scaling Laws

Abstract

Masked diffusion models (MDMs) are trained to predict tokens in every order but generate in one. As such, MDM likelihood is an intractable sum over all orders. The evidence lower bound (ELBO) is used as a proxy, but it is a loose bound and assumes positions are unmasked uniformly at random. Frontier MDMs, however, select positions with deterministic rules such as confidence. To resolve this mismatch, we introduce DUEL, a framework that unifies common deterministic MDM samplers. DUEL allows for likelihood computation that is both exact and under the sampler that is actually used: replay generation, reveal the true tokens instead of sampled ones, and sum their log-probabilities. DUEL is a simple plug-and-play algorithm applicable to any pretrained MDM with no retraining. Evaluating with DUEL instead of the ELBO changes what we conclude about MDMs. The perplexity gap between MDMs and autoregressive (AR) models shrinks by up to 32% in-domain and 82% zero-shot. In scaling laws, the reported MDM–AR compute gap is under DUEL: between and of it was an evaluation artifact. Scoring multiple-choice answers with DUEL improves LLaDA-8B accuracy on all six tasks, by up to 9.7 points. DUEL also selects samplers correctly: it matches a gold-standard LLM judge on every decided comparison, which generative perplexity and MAUVE fail to do. DUEL reveals that MDMs are ultimately more competitive with AR models than the ELBO suggests.

Primary area:
generative models
Keywords:
masked diffusion models, discrete diffusion, diffusion language models, likelihood evaluation, exact likelihood, evidence lower bound (ELBO), deterministic unmasking, confidence-based sampling, any-order autoregressive models, perplexity, scaling laws, sampler selection, language model evaluation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JQwoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DUEL, a method that computes the exact likelihood of a masked diffusion model with deterministic unmasking. Authors observe that such model defines a proper probability distribution. Since the unmasking order is fixed, the exact likelihood can be computed just by running the same rule on the true sequence. The authors compare DUEL with the ELBO, which is only a bound and assumes a uniformly random order. With open models DUEL makes the perplexity gap between MDMs and AR models smaller. DUEL also reduces the compute and also improves LLaDA-8B's multi choice accuracy. Also the authors use DUEL perplexity to rank samplers and find the ranking align with LLM judge.

Critical strengths

The core observation is correct and practical.

The method requires no training and gives exact normalized likelihood.

And also the method reflects the actual sampler whereas the ELBO doesn't.

The ablation in Table 16 of the appendix clearly mentiond about the limits of the method.

Critical weaknesses

The framing is stronger than the evidence. DUEL and the ELBO measure different distributions. With a left-to-right order, DUEL is basically the likelihood of an AR-style decoder from MDM. Table 16 shows that the exact uniform order likelihood already beats the AR model, while DUEL is worse than both. So the reduced compute gap depends on the chosen policy

The sampler-selection evidence is weak. The perfect agreement only counts pairs with a clear judge preference. These pairs come from just a few rules, and half use small models that produces bad meaningless text. There is only one LLM judge and no human check. So more concrete evaluatino is needed. DUEL also ignores the nucleus sampling used for generation. On GSM8K, it even prefers margin over greedy, though greedy is more accurate.

Questions for authors
  1. This is my understanding of Table 16 : The ELBO seems to overstate the gap to AR mainly because of the Jensen gap, and DUEL looks like a separate decoder that is worse than the uniform-order model. Can you provide explanation on this.

  2. How much does the Table 2 compute gap change with the change in policy, for example left-to-right vs greedy vs margin.

  3. Can you add a human or other judge check on some of the pairwise labels, plus a simple baseline like entropy or gen-ppl?

2 comments

Response to Reviewer JQwo (R3)

We thank the reviewer for the careful reading. We agree with most of the points raised and answer each question with new results. If this addresses the reviewer's concerns, we would be grateful if the reviewer would consider raising their score.

W1 / Q1. DUEL and the ELBO measure different distributions.

"DUEL and the ELBO measure different distributions." We agree. Of the two, only DUEL measures the distribution we actually sample from at inference time. The ELBO bounds the likelihood of a uniform random-order sampler, which no current MDM generates with.

"With a left-to-right order, DUEL is basically the likelihood of an AR-style decoder." We agree with this too. The reason it needs an argument is that an MDM is trained on every order, so its likelihood is a sum over all of them, p(x)=∑σp(x,σ)p(x) = \sum_\sigma p(x, \sigma), which is intractable in general. Our paper shows that under a deterministic rule this sum collapses to a single term, so the likelihood becomes exact and computable. Left-to-right is the simplest such rule. The same result covers confidence and margin rules, which are not AR-style and whose order depends on xx.

Table 16. Your reading is right: the ELBO overstates the gap to AR mainly through the Jensen gap, and the uniform-order sampler is a better density than DUEL. We have now confirmed this exactly on in-domain data for two models (Table R2.1 in our response to Reviewer Lgnp). But the uniform-order sampler is not usable in practice: its exact likelihood is intractable beyond short blocks, and it generates poorly (37.8% on GSM8K, against 64.0% with greedy confidence; Table 18).

Q2. How much does the Table 2 compute gap change with the policy?

We agree that the reduction depends on the policy.

Table R3.1: Compute-optimal loss (NLL, nats/token) and, in parentheses, the MDM–AR compute gap. Table 2 protocol, same SMDM checkpoints.

Budget (FLOPs)ARELBODUEL left-to-rightDUEL greedyDUEL margin
6e182.9183.449 (×14.6)3.398 (×10.0)3.403 (×10.9)3.422 (×12.2)
1e202.4362.897 (×15.9)2.825 (×9.4)2.837 (×11.0)2.860 (×12.7)

Every rule is below the ELBO on every checkpoint. We chose left-to-right for Table 2 because it is the most comparable to the AR baseline: both models are then scored under the same generative procedure, one token at a time from left to right. Under greedy confidence the compute gap still falls from ×15–16 to about ×11 (and to ×12–13 under margin), though these rules score a different generative model than left-to-right. We will add greedy confidence to the scaling-law results.

Q3 / W2. A second judge and a simple baseline.

We re-ran the pairwise study with a judge from a different model family (OpenAI GPT-6.1 Sol): same prompts, blinding and both presentation orders, on the same 200 sample pairs for each of the 24 pairs (9,600 calls, none missing). We also added token entropy as a baseline.

Table R3.2: Agreement with each judge on the pairs that judge decides.

MetricClaude Sonnet 5GPT-6.1 Sol
DUEL14/1416/16
Gen-ppl8/149/16
MAUVE12/1413/16
Token entropy (higher = better)8/1410/16
  • All 14 pairs that Sonnet decides are decided in the same direction by the second judge, which also decides two more.
  • DUEL is the only metric that agrees with both judges at both scales. Gen-ppl fails at 110M (1/7 and 2/9) and token entropy fails at 8B (1/7 under both judges).

On the remaining points:

  • Only decided pairs are counted. Over all 24 pairs by point direction, DUEL agrees on 21/24 (Sonnet) and 22/24 (GPT-6.1 Sol), against 15/24 and 14/24 for gen-ppl.
  • Half use small models. This is where a sampler-selection metric is most needed, since small research models score near chance on benchmarks, and it is where gen-ppl fails.
  • Ignores nucleus sampling. Correct, and so does standard perplexity for AR models, which does not see temperature or nucleus sampling either (Limitations).
  • On GSM8K it prefers margin over greedy. We agree: DUEL is not a perfect predictor of accuracy, and App. E.3 presents this pair as the caveat. Likelihood and accuracy measure different things, so we expect correlation, not agreement on close pairs. This is how generative perplexity is used for sampler selection today, and unlike generative perplexity, DUEL cannot be lowered by entropy collapse.

Thank you for the detailed rebuttal and the additional experiments, which address my questions. I maintain my score of 3.

Reviewer_LgnpOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents an new framework DUEL for estimating exact likelihoods of masked diffusion language models under deterministic unmasking policies. This is inference-time method, that don't need retraining techniques. The authors demonstrate that evaluating models with DUEL instead of the standard ELBO reduces the reported perplexity (exp(log-likelihood)) gap between diffusion and autoregressive models, leads to more favorable pre-training scaling laws, and improves multiple-choice accuracy on large-scale setups. And provide interesting meaningful intuition on sampler selection for discrete diffusion language models.

Critical strengths

Important and well-motivated problem: The paper addresses the limitations of ELBO-based likelihood evaluation for masked diffusion models and introduces a practical approach to evaluating the distributions induced by their actual sampling policies. The problem is really important, because nowadays most common approaches like Gen-PPL/Entropy, MAUVE is not really good, and can't fixate generation collapse. So likelihood estimation is undeniably important task.

Extensive experimental evaluation: The proposed method is evaluated across multiple models, datasets, and scales, including small-scale (MDM, SEDD, BD3-LM) and large-scale (LLaDA-8B), with promising results on perplexity, multiple-choice benchmarks, and sampler selection.

Clear theoretical presentation: The framework is clearly formulated, with rigorous derivations and mathematically sound proofs in main and appendix paper parts.

Critical weaknesses

Limited applicability: DUEL provides exact likelihood only for deterministic unmasking policies, excluding stochastic unmasking and remasking strategies. Basically yes, I agree that confidence-based samplers often improve generation quality, stochastic alternatives remain relevant, including at larger scales. Moreover, recent large-scale diffusion models are not limited to masked diffusion. For example, DiffusionGemma [1] uses uniform-state diffusion, which falls outside the proposed framework. Thus, despite its practical value, DUEL remains restricted to a particular class of diffusion models and sampling procedures.

Comparison of different generative distributions: DUEL does not provide a tighter estimate of the original MDM likelihood. Instead, changing the unmasking policy induces a different generative distribution. I mean my main concern is that the paper compares the ELBO of the original MDM with the exact likelihood computed by DUEL, although these correspond to different generative distributions. The ELBO bounds the likelihood under the uniform unmasking policy (pθunifp_\theta^{\mathrm{unif}}), whereas DUEL evaluates (pθπFp_\theta^{\pi^F}), induced by a deterministic policy. Therefore, the reported improvement cannot be interpreted solely as a consequence of more accurate likelihood estimation. Their NLL difference decomposes as

+ \underbrace{(\mathcal L_{\mathrm{unif}}-\mathcal L_{\mathrm{DUEL}})}_{\text{Policy-induced gap}},$$ where $\mathcal L_{\mathrm{unif}}$ is exact NLL of $p_\theta^{\mathrm{unif}}$, $\mathcal L_{\mathrm{ELBO}}$ is ELBO estimation and $L_{\mathrm{DUEL}}$ is DUEL estimation. While the first term is the Jensen gap, the second term can have either sign. Thus, the main experimental comparisons conflate the looseness of the ELBO with the effect of changing the underlying generative distribution. I believe this distinction is particularly important for interpreting the reported perplexity improvements and pre-training scaling laws. Maybe results from Appendix E should be in the main part. [1] DiffusionGemma Team, “DiffusionGemma Technical Report” (2026). arXiv preprint arXiv:2608.00146
Questions for authors

See weaknesses

1 comment

Response to Reviewer Lgnp (R2)

We thank the reviewer for the careful review. We agree with both weaknesses as stated, and we have now measured the decomposition you propose exactly (Table R2.1). If this response addresses the reviewer's concerns, we would be grateful if the reviewer would consider raising their score.

W1. Limited applicability.

We agree: DUEL is exact only for deterministic unmasking rules on masked diffusion models, and we will state this more prominently in the Limitations, citing DiffusionGemma. This class is still an important one, since LLaDA, Dream, LLaDA 2.0, SDAR and Nemotron-Diffusion all report results with a deterministic rule, and their only existing likelihood is an ELBO for a random-order sampler that none of them use.

W2. Different generative distributions.

We agree. Changing the unmasking policy changes the generative distribution, and DUEL is not a tighter estimate of pθπunifp^{\pi^{\text{unif}}}_\theta. It is the exact likelihood of a different member of the family, pθπFp^{\pi^F}_\theta. The manuscript will say this explicitly in §4.1.

Your decomposition is the right way to read our results:

LELBO−LDUEL=(LELBO−Lunif)⏟Jensen gap≥0+(Lunif−LDUEL)⏟policy term.\mathcal{L}_{\text{ELBO}}-\mathcal{L}_{\text{DUEL}} = \underbrace{(\mathcal{L}_{\text{ELBO}}-\mathcal{L}_{\text{unif}})}_{\text{Jensen gap} \ge 0} + \underbrace{(\mathcal{L}_{\text{unif}}-\mathcal{L}_{\text{DUEL}})}_{\text{policy term}}.

We now compute all three quantities exactly. Lunif\mathcal{L}_{\text{unif}} is tractable for short blocks by dynamic programming over subsets of positions (2L′2^{L'} forward passes per block), which also gives the ELBO's estimand with no Monte Carlo noise. We evaluate BD3-LM on held-out OWT (256, 256 and 32 sequences at L′L' = 4, 8, 16) and LLaDA-8B on 1000 Wikitext spans of 8 tokens, all at k=1k=1.

Table R2.1. Exact perplexity (↓) of each term in the decomposition.

BD3-LM L′L'=4BD3-LM L′L'=8BD3-LM L′L'=16LLaDA-8B, 8-token spans
ELBO (LELBO\mathcal{L}_{\text{ELBO}})20.7621.7122.599.44
DUEL, greedy confidence (LDUEL\mathcal{L}_{\text{DUEL}})19.7220.3421.039.00
DUEL, left-to-right (LDUEL\mathcal{L}_{\text{DUEL}})19.0719.6820.398.92
Uniform-order sampler (Lunif\mathcal{L}_{\text{unif}})18.5818.6718.818.30
  • Ordering of samplers. In every column, perplexity is lowest for the uniform-order sampler, then left-to-right, then greedy confidence, then the ELBO. Note that this is for one token per step (k=1k=1) and may change for parallel unmasking (k>1k>1).
  • DUEL's gain comes from removing the Jensen gap, not from the choice of sampler. The ELBO sits 1.1–3.8 perplexity points above the exact likelihood of its own uniform-order sampler. Switching to a deterministic rule then costs 0.5–2.2 points (all bootstrap 95% intervals exclude 0). As you note, this sign is not guaranteed in general.
  • The uniform-order sampler scores well but generates poorly. Its likelihood averages pθ(x∣σ)p_\theta(x\mid\sigma) over all orders, and an average of probabilities is dominated by its largest terms. The few orders that explain a given xx very well therefore pull the perplexity down disproportionately, even if a typical order is poor: the best order per block reaches 14.5 perplexity at L′L'=4, against 18.6 for the mixture. A sample, however, is drawn along a single random order, so it gets a typical order, not the best one. Under it LLaDA-8B solves 37.8% of GSM8K, against 64.0% with greedy confidence (Table 18).

"Results from Appendix E should be in the main part." Agreed. We will move the decomposition into §4.1 with Table R2.1, so that every ELBO-vs-DUEL comparison is read with it in hand.

Reviewer_EPzrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposed the use of deterministic unmasking orders to exactly compute the likelihood of MDM generations. They use this method to revise the perplexity and training efficiency gaps between AR and MDM. They also use this likelihood score to multiple choice answers and select samplers.

Given that many existing works propose unmasking orders that work better than random -- using those orders to compute the exact likelihood is an immediate side effect, not a new finding -- A fixed unmasking order "will" induce a factorization of likelihood that can be computed exactly.

As for improving perplexity and scaling laws under fixed policy --- it is not guaranteed that using one fixed ordering will make likelihood higher compared to a uniform average over all orders -- the fact that it happens in the experiments is due to the particular fixed policy used -- and in that vein, paper should have tried the existing published methods for finding good orderings on a trained model and compare them [1][2][3][4] that would have added more value to the work.

The proposed practical use of the method is also not very convincing: The multiple-choice experiments compare only likelihood-based scoring methods, if the goal is to improve performance, a direct answer generation baseline should be added. The sampler-selection experiments consider only simple ordering heuristics, and in the appendix the likelihood-based does not agree with gsm8 correctness. Here also it will add value if the usefulness of sampler selection technique is shown to work in selecting the best among known sampling stratgies [1][2][3][4]

also [1] and [3] are not cited.

  1. Yang, K., Teoh, J., Yang, K., Zhang, Y., & Lamb, A. (2026). Improving Sampling for Masked Diffusion Models via Information Gain.
  2. Peng, F. Z., Bezemek, Z., Patel, S., Rector-Brooks, J., Yao, S.,Tong, A., & Chatterjee, P. (2025). Path Planning for Masked Diffusion Model Sampling (P2).
  3. Patel, D., Naseem, T., Pandey, G., Sultan, M. A., McCallum, A., & Fernandez Astudillo, R. (2025). Improved Sampling from Masked Diffusion Models with Position Contrastive Guidance.
  4. Kim, J., Shah, K., Kontonis, V., Kakade, S. M., & Chen, S. (2025). Train for the Worst, Plan for the Best: Understanding Token
Critical strengths

The paper proposed the use of deterministic unmasking orders to exactly compute the likelihood of MDM generations. They use this method to revise the perplexity and training efficiency gaps between AR and MDM. They also use this likelihood score to multiple choice answers and select samplers.

Critical weaknesses

As for improving perplexity and scaling laws under fixed policy --- it is not guaranteed that using one fixed ordering will make likelihood higher compared to a uniform average over all orders -- the fact that it happens in the experiments is due to the particular fixed policy used -- and in that vein, paper should have tried the existing published methods for finding good orderings on a trained model and compare them [1][2][3][4] that would have added more value to the work.

The proposed practical use of the method is also not very convincing: The multiple-choice experiments compare only likelihood-based scoring methods, if the goal is to improve performance, a direct answer generation baseline should be added. The sampler-selection experiments consider only simple ordering heuristics, and in the appendix the likelihood-based does not agree with gsm8 correctness. Here also it will add value if the usefulness of sampler selection technique is shown to work in selecting the best among known sampling stratgies [1][2][3][4]

also [1] and [3] are not cited.

  1. Yang, K., Teoh, J., Yang, K., Zhang, Y., & Lamb, A. (2026). Improving Sampling for Masked Diffusion Models via Information Gain.
  2. Peng, F. Z., Bezemek, Z., Patel, S., Rector-Brooks, J., Yao, S.,Tong, A., & Chatterjee, P. (2025). Path Planning for Masked Diffusion Model Sampling (P2).
  3. Patel, D., Naseem, T., Pandey, G., Sultan, M. A., McCallum, A., & Fernandez Astudillo, R. (2025). Improved Sampling from Masked Diffusion Models with Position Contrastive Guidance.
  4. Kim, J., Shah, K., Kontonis, V., Kakade, S. M., & Chen, S. (2025). Train for the Worst, Plan for the Best: Understanding Token
Questions for authors

see weaknesses section.

5 comments

Edited Nov 18: Table R1.1 updated with final numbers.

We thank the reviewer for the constructive feedback. Following the suggestions, we have:

  • evaluated DUEL under the orderings from [1–4] that admit exact likelihood, both for perplexity and for sampler selection;
  • added citations and discussion of [1] and [3].

If these additions address the reviewer's concerns, we would be grateful if the reviewer would consider raising their score.

W1. "Exact likelihood under better-than-random orders is an immediate side effect."

We agree that many works propose orders that beat random. We disagree that exact likelihood under them is immediate; it is neither trivial nor obvious.

For any policy, the MDM likelihood is a sum over all unmasking orders (Eq. 3). That is more than L!L! terms, each needing its own forward passes. A better order does not make this tractable; stochastic orders, including the uniform one, leave it intractable, which is why the ELBO is used.

What makes it tractable is determinism. Under deterministic selection, every order but one has probability zero, so the sum collapses to a single chain-rule term (Theorem 1). That order is not "fixed": under confidence- or margin-based rules it depends on xx and is unknown in advance. DUEL recovers it by replaying the rule on xx.

Exact likelihood under DUEL holds whenever:

  • (C1) position selection is a deterministic function of the revealed tokens;
  • (C2) the selected tokens are then drawn from a fixed, normalized distribution.

Some better-than-random samplers violate these conditions: P2 [2] remasks revealed tokens, and Info-Gain [1] commits the token that scores best under its lookahead. We will clarify this in the updated manuscript.

W2. "A fixed order is not guaranteed to beat the uniform average; compare published orderings [1–4]."

We agree there is no guarantee; App. E.2 already notes that DUEL need not beat the ELBO. We now test the methods from the reviewer's references that satisfy (C1)–(C2):

  • [4] Kim et al.: its top-probability-margin rule is our "probability margin" rule, already in Tables 1, 10 and 18.
  • [2] P2: without remasking, its deterministic form is greedy confidence, already in the paper. With remasking it is out of scope (Limitations).
  • [1] Info-Gain: we use its lookahead objective to choose positions only, then sample tokens from the denoiser. This variant satisfies (C2).
  • [3] PCG: it uses its guided distribution only to order positions; tokens are drawn from the unmodified denoiser, so (C1)–(C2) hold. We use its ordering with a fixed number of positions per step.

Table R1.1: Held-out perplexity under published orderings. Same checkpoints as Table 1; kk=1. LLaDA-8B: full Wikitext test set. BD3-LM: a 10M-token subset of OWT validation (every rule and the ELBO on the same tokens; the full-split ELBO is 22.27).

RuleBD3-LM L′L'=16, OWT PPL ↓LLaDA-8B, Wikitext PPL ↓
ELBO (bound)≤22.22≤15.29
Left-to-right20.0413.84
Greedy confidence (= P2 w/o remasking)20.6014.05
Probability margin [4]20.9014.45
PCG [3]20.7814.14
Info-Gain, positions only [1]20.5613.92

The LLaDA-8B greedy value supersedes the 14.50 in Table 1d, which we will update.

The gain over the ELBO does not hinge on one well-chosen rule: every ordering scores below the ELBO on both models. When unmasking one token at a time, left-to-right gives the lowest perplexity on both models, despite being the simplest strategy.

Comment 2 (edited)

Edited Nov 18: Table R1.2 and the rank correlations updated with final numbers.

W3. "The practical use is not convincing."

We see held-out perplexity and pre-training scaling laws as the primary applications. These are how MDMs are compared with AR models, and DUEL changes the conclusion: the SMDM compute gap falls from 16× to 9×. The multiple-choice experiment is not meant to maximize accuracy. It isolates the estimator, keeping the model, prompt and protocol fixed, to show that exact likelihood makes a measurable difference over the ELBO.

W4. "Sampler selection uses only simple heuristics and disagrees with GSM8K correctness."

We agree, and we will revise App. E.3 to present this as a limitation. DUEL perplexity measures distributional coverage, KL(data ‖ sampler): the forward-KL counterpart to generative perplexity, which measures KL(sampler ‖ reference LM) + H(sampler) [Pynadath et al., 2026]. Accuracy under greedy decoding depends mainly on precision. We therefore do not expect DUEL perplexity to rank close competitors, and the extended study confirms this.

Table R1.2: LLaDA-8B GSM8K. DUEL ppl of the reference solution vs pass@1 (%).

Rulekk=2 ppl ↓kk=2 acc ↑kk=4 ppl ↓kk=4 acc ↑
Left-to-right1.7961.62.3843.5
Greedy (= P2 w/o remasking)1.7364.02.1051.9
Margin [4]1.7264.12.0048.8
PCG [3]1.7365.42.0553.5
Info-Gain, positions only [1]1.7265.12.0848.2
Lowest entropy1.7460.42.1140.3
PC-Sampler (Huang et al., 2025)1.7364.42.2149.9

Over the five rules from the paper and [1–4] (the first five rows), DUEL perplexity correlates positively with accuracy: Spearman ρ = 0.60 at kk=2 and 0.50 at kk=4 (bootstrap 95% intervals [0.00, 0.97] and [0.10, 0.70]). But it does not separate close competitors. For example, at kk=4 lowest-entropy matches greedy in perplexity (2.11 vs 2.10) but is 11.6 points less accurate. For this reason, our main sampler comparison (§4.4) evaluates open-ended generation with an LLM judge, where distributional fit is the quantity of interest. GSM8K serves as a stress test.

DUEL's contribution here is held-out likelihood for deployed samplers. Recent work considers this intractable for diffusion LMs [Franca & Tong, 2026]. The current alternatives can be gamed (generative perplexity [Franca & Tong, 2026]) or are confounded by compute and temperature (MAUVE rankings [CaRE, 2026]). DUEL cannot be gamed by collapsing entropy, because its entropy term is the entropy of the data.

Citations

We will cite and discuss [1] and [3] in App. A and in the rule catalog of App. C.


What changed

Comment 1:

  1. Added the "Edited" line.
  2. Table R1.1, greedy / LLaDA: "14.05 (paper: 14.50, see notes)" → "14.05", plus the one-line note under the table.
  3. Table R1.1, Info-Gain / BD3-LM: "20.64 †" → "20.56".
  4. "every ordering is tighter than the uniform average" → "every ordering scores below the ELBO".

Comment 2:

  1. Added the "Edited" line.
  2. Table R1.2, Info-Gain row: "1.72 † / 65.4 † / 2.08 † / 48.4 †" → "1.72 / 65.1 / 2.08 / 48.2".
  3. Correlation sentence: "ρ = 0.83 [TODO] ... 0.52 [TODO]" → 5-rule values 0.60 and 0.50 with intervals, and a clause saying which five rules.

Dear reviewer, please let us know if our response answered the questions before the end of the review period. If our response addresses your concerns, we would be grateful if you would consider raising your score. Thank you.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Changes

Dear Reviewers and Area Chair,

Thank you for your time on our submission. To help wrap up the discussion, here is a short summary of what we added during the rebuttal. All results are in our individual responses below.

New experiments

  1. Exact decomposition of the ELBO–DUEL gap (Reviewers Lgnp, JQwo; Table R2.1). On two models we computed the ELBO, DUEL and the uniform-order sampler's likelihood exactly. The ordering is the same in every setting: uniform-order sampler < DUEL < ELBO. DUEL's gain comes from removing the Jensen gap, not from the choice of sampler.
  2. Published orderings (Reviewer EPzr; Table R1.1). DUEL under the orderings of Kim et al., P2, PCG and Info-Gain. Every ordering scores below the ELBO on both BD3-LM and LLaDA-8B.
  3. Scaling laws under three rules (Reviewer JQwo; Table R3.1). The 16× compute gap under the ELBO falls to about 9× (left-to-right), 11× (greedy) and 12–13× (margin). Every rule is below the ELBO on every checkpoint.
  4. Second LLM judge and entropy baseline (Reviewer JQwo; Table R3.2). A judge from a different model family decides all 14 of the original judge's pairs in the same direction. DUEL agrees with both judges on every decided pair (14/14 and 16/16), and is the only metric that does so at both model scales.
  5. Sampler selection on downstream tasks (Reviewer EPzr; Tables R1.2, R1.4, R1.5). GSM8K extended to the published orderings, plus MBPP. DUEL perplexity correlates with accuracy across samplers, but does not rank close competitors.
  6. Direct answer generation (Reviewer EPzr; Table R1.3). Free generation is far below likelihood scoring. Constrained one-token generation is above it, but with a single token there is no ordering, so it does not distinguish the ELBO from DUEL.

Points we agreed with and will reflect in the paper

  • DUEL is the exact likelihood of the deployed sampler, not a tighter estimate of the uniform-order likelihood. We will state this in §4.1 and move the decomposition there.
  • The compute-gap reduction depends on the rule. We will add greedy confidence to the scaling-law results.
  • DUEL is not a perfect predictor of benchmark accuracy. App. E.3 will present this as a limitation.
  • Scope: deterministic unmasking on masked diffusion models. We will state this more prominently and cite DiffusionGemma.
  • We will cite and discuss Info-Gain and PCG, and make explicit in §4.3 that the multiple-choice experiment isolates the estimator and is not meant to maximize accuracy.

We are happy to answer any further questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.