acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Soft Tokens for Parallel Decoding in Diffusion Language Models

Abstract

Diffusion language models (DLMs) enable parallel generation by predicting and committing multiple tokens at each denoising step, yet they can generate individually plausible but mutually inconsistent tokens. Recent work shows that soft tokens can mitigate this issue by representing uncertain positions with continuous embeddings built from the model's predictive distribution at the previous decoding step. However, although soft tokens are commonly understood as preserving predictive uncertainty, how soft-token feedback improves parallel decoding has not been systematically examined. In this paper, we investigate this question in frozen pretrained DLMs to examine soft-token feedback without the effects of additional training. To construct soft-token inputs in a training-free setting, we identify a geometric mismatch between conventional soft-token construction and the pretrained embedding space. Based on this observation, we propose a training-free, geometry-aware construction of soft tokens. Our analysis of soft-token feedback suggests that uncertainty preservation alone does not fully explain how it reshapes subsequent predictions. To better explain how soft-token feedback improves parallel decoding, we provide empirical evidence that it favors coherent token sequences. Across four pretrained DLMs and four math and code benchmarks, our method outperforms standard parallel decoding and a training-free Euclidean soft-token baseline.

Primary area:
generative models
Keywords:
Diffusion language models, continuous feedback

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_C2iVOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper investigates soft-token feedback for parallel decoding in frozen pretrained diffusion language models (DLMs). The authors identify geometric limitations of Euclidean interpolation between the [MASK] embedding and a probability-weighted aggregate of predicted-token embeddings. They propose a training-free construction using spherical linear interpolation (SLERP), which controls angular displacement while preserving the norm of the [MASK] embedding.

The paper further studies how soft-token feedback modifies predictive distributions. It introduces an additive-mixture reference obtained by marginalizing over hypothetical discrete token realizations and compares this with a normalized multiplicative mixture. The observed predictions are closer to the multiplicative reference, motivating an agreement-seeking interpretation of soft-token feedback. Sequence-level consistency is evaluated using conditional reverse KL divergence between factorized parallel predictions and a sequential decoding reference.

Experiments on four pretrained DLMs and four mathematical reasoning and code-generation benchmarks report consistent improvements over Euclidean soft tokens and frequent improvements over vanilla parallel decoding. Additional analyses examine embedding geometry, interpolation choices, adaptive decoding, and sequence-level consistency.

The empirical findings are promising. However, the mathematical analysis does not establish that the improvements are specifically attributable to the proposed agreement-seeking mechanism rather than alternative explanations involving logit interpolation, effective interpolation strength, normalization, and prediction concentration.

Critical strengths

1. Strong empirical improvements under frozen-model decoding.

The method is evaluated across four pretrained DLMs, four benchmarks, and three parallel decoding rates. Table 2 reports improvements over Euclidean soft-token feedback in 46 of 48 settings, with ties in the remaining two. Improvements over vanilla decoding are also frequent. For example, on Dream/HumanEval at four tokens per step, accuracy increases from 17.68% (vanilla) and 23.78% (Euclidean) to 32.32%.

These results are practically interesting because the pretrained models remain frozen. They support the potential utility of continuous feedback for parallel decoding without additional training, even though they do not independently establish the proposed mechanism.

2. An interpretable construction with useful ablations.

The method separates angular interpolation from norm preservation. Let

u=eMASK∥eMASK∥2,v=m∥m∥2.u=\frac{e_{\mathrm{MASK}}}{\|e_{\mathrm{MASK}}\|_2}, \qquad v=\frac{m}{\|m\|_2}.

The proposed construction applies SLERP between uu and vv, then scales the result to norm ∥eMASK∥2\|e_{\mathrm{MASK}}\|_2. This is a simple, mathematically interpretable design motivated by the observed angular separation between the [MASK] embedding and candidate aggregates.

The paper also includes comparisons with norm-matched linear interpolation, alternative output scaling, hyperparameter sensitivity, adaptive samplers, and single-token success recovery. These evaluations strengthen the empirical contribution, although they do not completely isolate the claimed geometric and probabilistic mechanisms.

Critical weaknesses

1. The proposed probabilistic mechanism is not mathematically identified by the reference-mixture and reverse-KL analyses.

Section 3 models uncertainty preservation through a latent discrete variable ZZ and defines

padd(y)=∑zwzpθ(y∣E,Z=z).p_{\mathrm{add}}(y) =\sum_z w_z p_\theta(y\mid E,Z=z).

This is a valid marginalization under the assumed latent-variable model, but it is not a necessary consequence of uncertainty preservation through deterministic continuous embeddings. The construction assumes independent positionwise latent realizations, although the paper's motivation concerns dependencies among simultaneously predicted tokens.

Furthermore, the proposed SLERP representation is generally not the expected embedding under this model. That expectation would be

(1−λ)eMASK+λm,(1-\lambda)e_{\mathrm{MASK}}+\lambda m,

whereas SLERP interpolates normalized directions and subsequently rescales the output. No mathematical identity connects the additive reference to the proposed representation.

The multiplicative reference also admits a simpler explanation. Writing pz(y)=softmax(ℓz)(y)p_z(y)=\mathrm{softmax}(\ell_z)(y), we obtain

pmult(y)=softmaxy(∑zwzℓz(y)).p_{\mathrm{mult}}(y) =\mathrm{softmax}_y\left( \sum_z w_z\ell_z(y) \right).

Thus, the multiplicative mixture is exactly softmax over averaged logits. Closer agreement with this reference does not uniquely identify an agreement-seeking mechanism; it may also reflect ordinary logit interpolation or changes in predictive concentration.

Appendix D.2 additionally includes the queried position among the discretized inputs, introducing a self-conditioning confound. The reference predictions may be influenced by explicitly supplying a candidate token at the evaluated position rather than by resolving dependencies across positions.

Finally, the reverse-KL analysis conditions both distributions on a restricted support SS. For factorized prediction QQ and sequential reference PP,

DKL(QS∥PS)=EQS[log⁡Q(y)P(y)]+log⁡P(S)Q(S).D_{\mathrm{KL}}(Q_S\|P_S) = \mathbb E_{Q_S} \left[\log\frac{Q(y)}{P(y)}\right] +\log\frac{P(S)}{Q(S)}.

Because both distributions are renormalized, the reported divergence does not measure the unconditional probability assigned to coherent sequences. A reduction may also arise from mode concentration. Although Appendix D.3 acknowledges this limitation, the proposed mechanistic interpretation remains stronger than the evidence establishes.

These issues are consequential because the agreement-seeking explanation is presented as a central scientific contribution.

2. The geometric ablation does not isolate the claimed mechanism, and the supplementary configurations raise substantial reproducibility concerns.

Let u=eMASKu=e_{\mathrm{MASK}}, v=mv=m, and let θ∈(0,π)\theta\in(0,\pi) denote the angle between their normalized directions. SLERP produces

sλ=sin⁡((1−λ)θ)sin⁡θu∥u∥2+sin⁡(λθ)sin⁡θv∥v∥2.s_\lambda= \frac{\sin((1-\lambda)\theta)}{\sin\theta} \frac{u}{\|u\|_2} + \frac{\sin(\lambda\theta)}{\sin\theta} \frac{v}{\|v\|_2}.

However, the same direction is obtained by normalizing (1−t)u+tv(1-t)u+tv, where

t=∥u∥2sin⁡(λθ)∥u∥2sin⁡(λθ)+∥v∥2sin⁡((1−λ)θ).t= \frac{\|u\|_2\sin(\lambda\theta)} {\|u\|_2\sin(\lambda\theta)+ \|v\|_2\sin((1-\lambda)\theta)}.

Therefore, SLERP and normalized linear interpolation traverse exactly the same spherical arc for nondegenerate endpoints. Their difference is the parameterization of angular displacement, not the set of reachable directions.

Consequently, the same-λ\lambda comparison in Appendix B.2 does not isolate an intrinsically different interpolation trajectory. The performance difference in Table 5 could reflect a more favorable effective interpolation schedule rather than a distinct geometric mechanism.

More seriously, the supplied supplementary configurations appear inconsistent with Appendix A.2. The manuscript states that hyperparameters are fixed across benchmarks and parallel-decoding rates within each domain. However, inspection of dllm/metadata/*.json reveals that 26 of the 48 main-table configurations differ from these stated settings. Eight configurations use k=1k=1, although the reported search space is k∈{2,3,4}k\in\{2,3,4\}.

The supplied reproduction launcher also supports only SLERP configurations, and the included samplers reject non-SLERP interpolation modes. Consequently, the Euclidean baseline cannot be directly reproduced using the supplied implementation.

These findings do not establish that the reported results are incorrect, but they raise substantial concerns about hyperparameter selection, baseline comparability, and whether the experiments follow the stated protocol.

Finally, the main empirical comparisons lack paired uncertainty estimates. Several improvements are small relative to benchmark size, making their statistical reliability difficult to assess.

Overall, the empirical improvements are promising, but the mathematical analyses do not isolate the claimed mechanisms, and the configuration discrepancies weaken confidence in the central experimental comparisons.

Questions for authors

Q1. Mathematical justification of the probabilistic reference.

Section 3 assumes uncertainty-preserving feedback should correspond to

padd(y)=∑zwzpθ(y∣E,Z=z).p_{\mathrm{add}}(y) =\sum_z w_z p_\theta(y\mid E,Z=z).

Why should this marginalization characterize uncertainty preservation for a deterministic continuous representation? Since SLERP generally does not equal the expected embedding under wzw_z, can the authors provide a mathematical connection between the proposed representation and this reference? How would the interpretation change under correlated latent realizations?

Q2. Identifiability of agreement-seeking behavior.

The multiplicative reference is exactly equivalent to softmax over weighted-average logits. How do the authors distinguish agreement-seeking behavior from ordinary logit interpolation or predictive concentration?

Appendix D.2 also includes the queried position among the discretized inputs. Does the multiplicative preference persist when the queried position remains masked and feedback is restricted to other positions?

Q3. Exact relationship between SLERP and linear interpolation.

For nondegenerate endpoints, SLERP is exactly reproducible by normalized linear interpolation with an angle-dependent coefficient. Is the claimed advantage therefore uniform angular parameterization rather than a distinct interpolation trajectory?

Were the SLERP, norm-matched linear, and Euclidean baselines independently optimized over their effective angular displacements? What validation data and selection criterion determined the final hyperparameters?

Q4. Interpretation of restricted-support reverse KL.

Tables 4 and 8 report conditional reverse KL after renormalization on finite supports. What fractions of the original probability mass do the methods and sequential references assign to those supports?

Can the authors separate improved dependency modeling from concentration on a single preferred sequence? Reporting unconditional probability mass assigned to coherent and incoherent alternatives would be informative.

Q5. Discrepancies between the manuscript and supplementary configurations.

Appendix A.2 specifies domain-level hyperparameters fixed across benchmarks and parallel-decoding rates. However, the supplied metadata contains 26 of 48 configurations that differ from these settings, including eight instances of k=1k=1, outside the stated search space.

Were these configurations used to generate Table 2? If so, how were they selected, and on what validation data? How was test-set tuning avoided?

Furthermore, the supplied reproduction launcher and samplers support only SLERP interpolation. Can the authors provide the executable Euclidean and vanilla baseline configurations required to reproduce the main comparisons?

Q6. Statistical reliability and computational fairness.

Can the authors provide paired uncertainty estimates for the benchmark improvements, particularly where differences are small?

For adaptive decoding in Table 6, what are the measured forward-pass counts and wall-clock costs? Do the improvements persist under matched computational budgets?

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We sincerely thank Reviewer C2iV for their thoughtful and constructive review. We are especially grateful that the reviewer clearly acknowledged the contributions of our paper, including the proposed training-free soft-token method and our analysis on why soft tokens improve the parallel decoding. We believe that our paper makes the first attempt at analyzing why soft tokens improve parallel decoding in diffusion language models and can meaningfully contribute to the community. Below, we provide point-by-point responses to all questions and concerns raised. All the comments are constructive and help strengthen our work.

Q1-1. Why use an additive mixture to formalize uncertainty preservation?

A canonical way to formalize uncertainty preservation over hypothetical discrete input states is to assign a probability distribution to the possible states and propagate it through the model by marginalization. This is the standard uncertain-evidence formulation discussed in prior works [1, 2]. We apply this formulation to soft-token feedback: each soft token represents uncertainty over the unresolved [MASK] state and its top-kk candidates, and the model’s prediction is averaged over the resulting discrete input realizations.

Specifically, at position jj, we assign probability 1−λ1-\lambda to [MASK] and probability λpˉj,i\lambda \bar p_{j,i} to candidate vj,iv_{j,i}. Under the factorized joint distribution q(z∣E)=∏jqj(zj∣E)q(z\mid E)=\prod_j q_j(z_j\mid E), marginalization gives

padd(y∣E)=∑zq(z∣E) pθ(y∣E,z).p_{\mathrm{add}}(y\mid E) = \sum_z q(z\mid E)\,p_\theta(y\mid E,z).

This defines the additive reference for this particular uncertain-evidence interpretation. It is a predictive-level reference for testing that interpretation, not a necessary consequence of every deterministic continuous representation.

Q1-2. Relationship between the additive reference and SLERP

The additive reference and our construction use the same positionwise candidate weights, but operate at different levels. The additive reference averages model predictions over discrete input realizations, whereas our method maps the candidate information to a single continuous embedding using SLERP to control its direction and preserve the [MASK] norm.

We do not claim that the SLERP embedding is the expected embedding under qq. At a single position, that expectation would be

E[eZj∣E]=(1−λ)eMASK+λ∑ipˉj,iej,i.\mathbb{E}[e_{Z_j}\mid E] = (1-\lambda)e_{\mathrm{MASK}} +\lambda\sum_i \bar p_{j,i}e_{j,i}.

SLERP generally differs from this expected embedding. Moreover, even if the expected embedding were used as the model input, its prediction would not generally equal the additive reference: the model’s nonlinear transformations mean that a prediction from an expected input need not equal the average of predictions from discrete inputs. We therefore use the additive reference as a predictive-level comparison, not as a formula that determines the SLERP construction.

Q1-3. How does the interpretation change under correlated latent realizations?

Our reference uses a factorized distribution over hypothetical input states. This means the latent state assignments are independent under the reference distribution; it does not assert that actual token dependencies are absent. A correlated reference would instead require specifying a non-factorized joint distribution q(z∣E)q(z\mid E), representing uncertainty over joint configurations across positions. The marginalization formula would remain unchanged, but the weights assigned to those configurations would differ. Such correlations are not determined by the positionwise candidate probabilities or by the deterministic soft-token embedding alone, so defining a correlated reference would require an additional modeling assumption.

AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

Q2-1. What does the multiplicative-reference comparison establish?

We define the multiplicative reference as the normalized weighted geometric mean of predictions induced by hypothetical discrete input realizations:

\frac{ \exp\left(\sum_z q(z\mid E)\log p_\theta(y\mid E,z)\right) }{ \sum_{y'} \exp\left(\sum_z q(z\mid E)\log p_\theta(y'\mid E,z)\right) }.$$ The geometric mean combines predictive distributions from separate model evaluations under different discrete input realizations. Since each component distribution is a softmax over its logits, the reference can also be written algebraically as softmax over weighted-average logits. This is a property of how the reference is constructed; it does not imply that the model’s prediction from a continuous soft-token input is computed by combining predictions from discrete input realizations in this way. ### **Q2-2. Does the multiplicative preference persist when the queried position remains masked?** The multiplicative reference remains closer to our method when the queried position is kept masked and feedback is restricted to the other positions. Table 1 reports the mean JS divergence to the additive and multiplicative references of LLaDA. With the queried position masked, the divergence to the multiplicative reference is 0.0737, compared with 0.1989 for the additive reference. This preserves the same qualitative preference observed with full evidence. **Table 1. Mean JS divergence to the additive and multiplicative references under full evidence and with the queried position masked. Lower values indicate greater similarity.** | Evidence setting | Additive reference | Multiplicative reference | |---|---:|---:| | Full evidence | 0.2099 | 0.0410 | | Queried position masked | 0.1989 | 0.0737 | ### **Q3. Exact relationship between SLERP and linear interpolation** SLERP and normalized linear interpolation follow the same spherical arc for nondegenerate endpoints. At the same numerical $\lambda$, they generally reach different points on that arc: SLERP makes $\lambda$ the fraction of angular displacement, whereas the effective angular displacement under normalized Euclidean interpolation depends on $\lambda$ and the endpoint norms. The norm-matched linear baseline in Table 2 normalizes the Euclidean interpolant to the [MASK] norm. Its comparison with SLERP therefore tests the interpolation parameterization at the same numerical $\lambda$, rather than a different path or set of reachable directions. We will describe the result accordingly: the ablation supports the benefit of SLERP’s angular parameterization under the tested settings. Appendix A.2 specifies the search grid and the final settings fixed across benchmarks and parallel-decoding rates within each domain. **Table 2: Effects of interpolation and output norm on accuracy (%)** | Model | Benchmark | [MASK] norm with SLERP | Candidate norm with SLERP | [MASK] norm with Linear | |---|---|---:|---:|---:| | Dream | MATH-500 | **22.40** | 21.20 | 16.40 | | Dream | HumanEval | **50.61** | 50.00 | 50.00 | | LLaDA | MATH-500 | **23.60** | 22.60 | 21.80 | | LLaDA | HumanEval | **45.12** | 41.46 | 42.07 | | LLaDA 1.5 | MATH-500 | **24.00** | 21.80 | 21.20 | | LLaDA 1.5 | HumanEval | **43.90** | 39.63 | 37.80 |
AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

Q4-1. Probability mass retained on the restricted support

For each target, we construct the restricted support as the union of the top-10 candidate sequences from the distributions being compared. Table 3 reports the average probability mass each distribution assigns to this support before renormalization. The support captures most of the probability mass for all three distributions in both datasets.

Table 3. Mean probability mass assigned to the shared top-10 support, averaged over targets.

DatasetSoft methodAdditive referenceMultiplicative reference
GSM8K99.91%99.75%99.85%
Math50097.68%97.06%97.41%

Q4-2. How can improved dependency modeling be distinguished from concentration on a single preferred sequence?

We use reverse KL as a proxy for whether the factorized prediction concentrates probability on sequence combinations supported by the dependency-preserving sequential reference. Specifically, we measure DKL(QS∥PS)D_{\mathrm{KL}}(Q_S \| P_S), where QSQ_S is the factorized prediction and PSP_S is the sequential reference, both normalized on the shared support. This direction has a mode-seeking asymmetry: it penalizes QSQ_S for assigning probability to sequences that PSP_S considers unlikely, while not requiring QSQ_S to cover every mode of PSP_S. Thus, lower divergence indicates that the prediction’s probability mass is better aligned with sequences supported by the reference.

Table 4 shows that our method has the lowest reverse KL for all four models, indicating closer alignment with the sequential reference than either vanilla parallel decoding or the Euclidean baseline.

Table 4. Reverse KL divergence from the factorized joint prediction to the dependency-preserving sequential reference.

ModelVanillaEuclideanOurs
LLaDA2.19890.34150.2458
LLaDA-1.52.03271.92730.5624
LLaDA-2.0 mini2.73381.56101.1739
Dream4.27334.11391.7553

Q5. Hyperparameter selection and baseline reproducibility.

We use the settings specified in Appendix A.2. As in Section B.3, the method also performs consistently across a broad range of parameter values, indicating that its performance is not highly sensitive to the choice of kk and λ\lambda. We therefore selected the values of kk and λ\lambda from this stable range. The Euclidean and vanilla baselines are run by changing the mode from slerp to linear and vanilla in the configuration. This makes the baselines executable within the submitted implementation.

Q6-1. Paired uncertainty for the benchmark where the improvement differences are small?

For LLaDA-1.5 on MBPP with 8 tokens per step, the paired bootstrap 95% confidence interval for the accuracy difference (SLERP − Linear) is [−11,+5][-11, +5] percentage points. The exact McNemar test gives p=0.607p = 0.607. The interval includes zero, and the paired test finds no statistically detectable difference in this setting. The analysis compares both methods on the same benchmark examples; because decoding is greedy, there is no decoding-seed variability.

Q6-2. Computational budget for adaptive decoding methods.

For LLaDA on HumanEval, adaptive decoding with our feedback required an average of 120.5 forward passes per problem (median 118.5; range 59–191). This is approximately 76.5% fewer passes than the configured 512-step baseline. The average wall-clock time across the 164 HumanEval problems was 1,647.5 seconds (27.5 minutes), or about 10.0 seconds per problem.

References

[1] Chan and Darwiche. On the revision of probabilistic beliefs using uncertain evidence. IJCAI 2003

[2] Munk et al. Uncertain evidence in probabilistic models and stochastic simulators. ICML 2023

Reviewer_BrvxOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies soft tokens — continuous embeddings that replace a [MASK] at unresolved positions, built from the model's top-k predictive distribution — as a way to mitigate the well-known factorization problem in parallel decoding of diffusion language models (DLMs). Prior soft-token methods require training; the paper's goal is to isolate how soft-token feedback works by studying it in frozen pretrained DLMs. The method mainly contains two contributions: 1. A training-free, geometry-aware soft-token construction. The authors identify a geometric mismatch: when linearly interpolating between the [MASK] embedding and the probability-weighted candidate aggregate, the two endpoints are nearly orthogonal (empirically measured), so linear interpolation distorts angular movement and shrinks the embedding norm. They instead use spherical interpolation (SLERP) for direction and rescale to the [MASK] norm, keeping candidate aggregation in Euclidean space. 2. A mechanistic reinterpretation of soft-token feedback. Against the prevailing "uncertainty-preservation" view, the paper shows predictions induced by their construction are closer to a multiplicative mixture (Table 1), and argues this yields agreement-seeking behavior that favors coherent token combinations, measured via reverse-KL to an order-averaged sequential reference (Table 4). Experiments are done on Four frozen DLMs (LLaDA, LLaDA-1.5, LLaDA-2.0-mini, Dream) × four benchmarks (GSM8K, MATH500, HumanEval, MBPP) at 2/4/8 tokens-per-step, comparing Vanilla, a training-free Euclidean soft-token baseline.

Critical strengths
  1. A genuinely new, testable mechanistic account of soft-token feedback. Reframing the effect from "uncertainty preservation" (additive marginalization) to "agreement-seeking" (multiplicative/log-pooling) is the paper's most consequential contribution. It is specific supported by direct JS-divergence measurements against both reference mixtures, and it predicts the right qualitative behavior (suppression of incoherent cross-combinations).

  2. Training-free application to frozen models, enabling clean attribution. By constructing soft tokens that work without fine-tuning, the paper removes the confound of "learned adaptation to soft-token inputs" that clouds all prior training-based work. The geometric analysis that makes this possible (and the ablation isolating SLERP from norm preservation) is careful and well-evidenced, so the gains can be attributed to the feedback itself.

Critical weaknesses
  1. The central explanatory claim is validated on synthetic prompts, not on the tasks where accuracy is reported. The additive-vs-multiplicative and agreement-seeking evidence comes from 70 constructed toy prompts (and a finite-support proxy on completions), while the accuracy gains come from GSM8K/MATH500/HumanEval/MBPP. The paper never closes the loop by showing, on the actual benchmarks. The explanation is therefore plausible and well-motivated but not directly demonstrated on the system it is meant to explain.

  2. Causality between "lower reverse-KL" and "higher accuracy" is not established. Both quantities improve, but no controlled intervention ties them. Because gains are modest and sometimes non-monotone across baselines, a skeptic could attribute accuracy improvements partly to norm/scale effects (the method preserves [MASK] norm) rather than to agreement-seeking specifically — the Table 5 ablation hints that norm preservation alone is insufficient, but a direct dose-response experiment on real tasks is missing.

3.Missing related works baselines: 1)STRIDE: Bidirectional Soft-Token Propagation for Faster Diffusion Language Models; 2)Soft-masked diffusion language models. These works should be compared and analyzed.

Questions for authors
  1. Can you measure, on the real benchmarks (not toy prompts), a proxy for agreement-seeking (e.g., the degree of multiplicative vs. additive pooling, or reverse-KL on decoded spans) and show it correlates with per-example accuracy gains? This is the missing link between mechanism and results.
  2. Is there an intervention that modulates agreement-seeking strength (e.g., a temperature on the log-pooling, or \lambda sweep interpreted through the mixture lens) whose effect on accuracy matches the predicted direction?
  3. Given the concurrent Nigam et al. (2026) also uses spherical interpolation, can you state precisely what is novel in your construction beyond the mechanistic analysis, and whether their method would exhibit the same multiplicative behavior?
  4. In the cells where the Euclidean baseline matches or beats SLERP, is there a geometric explanation (e.g., smaller angular separation for those models/datasets)?

Generally, I think this is an interesting work, but I still have the above concerns.

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We sincerely thank Reviewer Brvx for their thoughtful and constructive review. We are especially grateful that the reviewer clearly acknowledged the contributions of our paper, including the proposed training-free soft-token method and our analysis on why soft tokens improve the parallel decoding. We believe that our paper makes the first attempt at analyzing why soft tokens improve parallel decoding in diffusion language models and can meaningfully contribute to the community. Below, we provide point-by-point responses to all questions and concerns raised. All the comments are constructive and help strengthen our work. We also appreciate that none of the points raised indicate fundamental flaws in our core ideas or methodology.

Q1. Can agreement-seeking behavior be measured on real benchmarks?

Table 1, reproduced from Table 8 in the Appendix, reports reverse-KL analysis on completions from real benchmarks. For each model-dataset pair, we sample 100 completions, mask spans of two to four tokens, and measure conditional reverse KL on a shared finite support against a sequential reference. Our method achieves the lowest mean divergence in all evaluated settings, compared with linear feedback and discrete parallel decoding. This provides evidence that the sequence-consistency effect also appears in real benchmarks.

As in Section D.4, it does not evaluate full benchmark generation: we mask only a few tokens in each completed output and compute reverse KL on those tokens. Computing reverse KL against the order-averaged reference would require an exponentially growing number of model evaluations, making it computationally infeasible for longer spans. Therefore, we use the controlled toy prompts for the analysis in the main paper.

Table 1. Mean conditional reverse KL on real benchmarks.

ModelDatasetSLERPLinearDiscrete MTP
DreamHumanEval0.0007040.0849620.001848
DreamMATH-5000.0005540.0039550.001372
LLaDA-1.5HumanEval0.0359730.0379070.038491
LLaDA-1.5MATH-5000.0159640.0169340.017006
LLaDA-8BHumanEval0.0006650.0009180.000977
LLaDA-8BMATH-5000.0532940.0538450.065190

Each value averages 100 completion-level measurements. Lower is better; the lowest mean in each row is bold.

Q2. Can an intervention modulate agreement-seeking strength?

We varied λ\lambda while keeping the other settings fixed and measured conditional reverse KL to the order-averaged sequential reference on MATH-500 completion spans (see Table 2-1). Since λ\lambda controls how much candidate information is incorporated into each soft token, this sweep tests how feedback strength affects agreement with the sequential reference.

The effect is non-monotonic: intermediate values (λ=0.3\lambda=0.3 and 0.50.5) produce the lowest divergence, while larger values substantially increase it. The accuracy sensitivity results show a compatible trend on real benchmarks: λ=0.3\lambda=0.3 gives the best accuracy across the evaluated models, and larger values reduce accuracy.

Changing λ\lambda can also change which completion the model favors. Table 2-2 shows the model's predictions given a prompt with two coherent alternatives: Harry Potter and Star Wars. The table illustrates that λ\lambda can alter the resulting predictions, including which coherent completion is favored.

Table 2-1. Conditional reverse KL on MATH-500 completion spans across λ\lambda values.

Datasetλ=0.1\lambda=0.1λ=0.3\lambda=0.3λ=0.5\lambda=0.5λ=0.7\lambda=0.7λ=0.9\lambda=0.9
MATH-5000.19040.05330.05510.54290.5688

Table 2-2. Token predictions for the Harry Potter or Star Wars prompt across λ\lambda values.

λ\lambdaFirst-token predictionSecond-token predictionResult
0.1Harry: 49%Wars: 31%No clear coherent mode
0.3Star: 58%Wars: 99%Star Wars
0.5Harry: 99%Potter: 99%Harry Potter
AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

Q3. How can lower reverse KL lead to higher accuracy?

Our hypothesis is that lower reverse KL may help parallel generation by bringing the factorized parallel prediction closer to order-preserving sequential decoding. The sequential reference conditions each prediction on tokens already filled in, thereby capturing dependencies that independent parallel predictions miss. A lower reverse KL therefore means that the parallel prediction assigns less probability to inconsistent token combinations and concentrates more on coherent sequences supported by the sequential reference.

This provides a plausible explanation for how closer agreement with order-preserving sequential decoding could benefit generation quality. The lower reverse KL of our method on both the controlled examples and benchmark-generated completion spans is consistent with this explanation. The norm-controlled ablation further shows that preserving the [MASK] norm alone does not account for the accuracy gains.

Q4. What is novel relative to Nigam et al. (2026) [1]?

Although both methods use SLERP, our novelty is not SLERP itself. Nigam et al. normalize the candidate embeddings, compute their weighted spherical Fréchet mean using iterative Karcher updates, and then apply SLERP between that mean and the [MASK] embedding [1]. Their evaluation uses continued pretraining of a 169M-parameter model.

Our construction is designed for frozen pretrained models and requires no additional training. Our geometric analysis finds that candidate embeddings are relatively aligned with one another, whereas their aggregate is nearly orthogonal to [MASK]. We therefore retain probability-weighted Euclidean aggregation for the candidates and use closed-form SLERP only for the aggregate-to-[MASK] interpolation, while preserving the [MASK] norm. This avoids iterative Fréchet-mean computation and is evaluated across four pretrained DLMs.

Using the checkpoint released by Nigam et al., we evaluated their method on our controlled toy prompts with λ=0.3\lambda=0.3, the value used for their released checkpoint. Let DJSaddD_{\mathrm{JS}}^{\mathrm{add}} and DJSmultD_{\mathrm{JS}}^{\mathrm{mult}} denote the mean Jensen–Shannon divergences to the additive and multiplicative references, respectively (see Table 3). For every tested top-kk, DJSmultD_{\mathrm{JS}}^{\mathrm{mult}} is lower, indicating that their method is consistently closer to the multiplicative reference on these prompts.

Table 3. Mean JS divergence to the additive and multiplicative references for Nigam et al.’s method.

Top-kkDJSaddD_{\mathrm{JS}}^{\mathrm{add}}DJSmultD_{\mathrm{JS}}^{\mathrm{mult}}
10.05350.0290
20.05190.0243
30.05130.0229

Q5. Do the geometric measurements explain when SLERP matches the Euclidean baseline?

Table 4 reports the accuracy gap and geometric measurements for the LLaDA-family settings.

Among these settings, only LLaDA/MATH500/8 is a tie. However, its aggregate-[MASK] and top-kk similarity angles fall within the ranges observed when SLERP improves accuracy. We therefore see no clear relationship between either geometric measurement and the accuracy gap, and these measurements alone do not explain the result in each model–benchmark setting.

A plausible explanation is that the geometric measurements characterize the input representations, while accuracy depends on how changes to those representations affect the model’s predictions for a particular task. The two constructions may lead to the same relevant token decisions in some settings and different decisions in others. We therefore view the geometry as motivation for the construction, rather than as a predictor of which benchmark will benefit.

Table 4. Accuracy gap and geometric measurements.

Setting (model / benchmark / tokens per step)SLERP - linear (pp)Aggregate-MASK angleTop-kk similarity angle
LLaDA / MATH500 / 80.0084.47° ± 2.54°7.87° ± 5.46°
LLaDA / GSM8K / 2+0.9186.88° ± 2.84°9.63° ± 6.09°
LLaDA / HumanEval / 2+3.6584.06° ± 3.03°6.78° ± 6.84°
LLaDA / MBPP / 2+1.6086.65° ± 2.98°13.14° ± 9.90°
LLaDA-1.5 / MATH500 / 2+3.0084.64° ± 2.82°8.84° ± 7.33°
LLaDA-1.5 / HumanEval / 2+4.8884.14° ± 3.13°6.79° ± 6.88°
LLaDA-1.5 / MBPP / 4+1.6086.64° ± 2.94°12.52° ± 9.93°
AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

Q6. Comparison with other baselines?

Since Soft-Masked Diffusion Language Models (SM) evaluates Dream-7B on coding benchmarks, we focus our direct comparison on HumanEval and MBPP [2]. In Table 5, we evaluate our method against SM on Dream-7B with four tokens per step.

Our training-free method outperforms the trained SM baseline on both benchmarks, by 7.52 percentage points on HumanEval and 0.50 percentage points on MBPP.

Table 5. Accuracy comparison with Soft-Masked on Dream-7B at four tokens per step.

MethodTraining-freeHumanEval (%)MBPP (%)
Soft-Masked (SM) [2]No24.8032.30
OursYes32.3232.80

We could not evaluate STRIDE because we could not find its code or checkpoint [3]. We will describe its approach in the related-work comparison but cannot provide a direct numerical comparison.

It is worth noting that our method is designed for training-free soft-token decoding with frozen pretrained DLMs. Existing soft-token methods obtain feedback through additional training or an auxiliary model [2,3]. To provide a comparison in the same training-free setting, we adapt the existing soft-token feedback mechanism into a training-free baseline, which we call Linear in the main paper. Linear constructs the probability-weighted candidate embedding directly from the frozen model’s predictions and uses linear interpolation with the [MASK] embedding. Our method uses the same candidate embedding but replaces the linear interpolation with SLERP. The two methods use the same forward-pass budget and decoding schedule, providing a matched-computation comparison that isolates the effect of the interpolation strategy.

References

[1] Lavanya Nigam, Ishaan Bansal, Aryan Sood, Vidit Aggarwal, and Gaurav Kumar Nayak. “Lost in interpolation: Why predictive feedback fails in diffusion language models.” arXiv preprint arXiv:2608.06529, 2026.

[2] M. Hersche, S. Moor-Smith, T. Hofmann, and A. Rahimi. “Soft-Masked Diffusion Language Models.” ICLR, 2026.

[3] Y. Liu and X. Ouyang. “STRIDE: Bidirectional Soft-Token Propagation for Faster Diffusion Language Models.” FMSys, 2026.

Reviewer_bb1YOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
3(+1)
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies soft-token feedback for parallel decoding in frozen masked diffusion language models. It constructs each unresolved position’s input by averaging top-k candidate embeddings, spherically interpolating between the candidate and mask directions, and preserving the mask embedding’s norm. The construction is motivated by measurements of the pretrained embedding geometry. The paper also investigates how feedback reshapes predictions, comparing against additive and multiplicative reference mixtures and measuring consistency with a sequential reference.

Critical strengths

I find the paper’s identification of the geometric effects of linear interpolation with the [MASK] embedding particularly informative and interesting. The overall motivation is clear, and the proposed normalization and interpolation procedure is straightforward to understand and implement.

Critical weaknesses
  1. The reasoning in Section 4 could be presented more clearly. I needed to consult the appendix to reconstruct the full argument, and some intermediate steps remain implicit. My understanding is that linear mixing can change vector norms and produce directions that depend on the endpoint norms. These effects apply both when aggregating candidate token embeddings and when interpolating between their aggregate and the [MASK] embedding. The authors appear to argue that the larger angular separation in the latter case makes these effects more consequential ( i.e lambda does not control how much weight over [MASK]), motivating the proposed adjustment. Although the main text states the geometric distinction, the connection between that observation and the resulting design choices needs to be explained more explicitly.

  2. The construction of esofte_{\mathrm{soft}} would benefit from a more complete geometric justification. Although the method is well motivated, I find the asymmetric treatment of the two aggregation stages insufficiently justified. The argument highlights the norm-dependent effects of linear combination, yet the method retains Euclidean averaging of the raw candidate embeddings and applies SLERP only between the normalized aggregate and the normalized [MASK] embedding. Individual candidate norms can also influence the direction of the candidate aggregate, so the same concern applies at that stage. Moreover, the reported similarities show that candidate embeddings are relatively aligned, rather than almost parallel. These averages alone do not establish that Euclidean averaging adequately addresses the geometric concerns raised by the paper.

Questions for authors
  1. Clarification of the independence assumption. The independence assumption in lines 190–192 does not hold in general. Could the authors clarify its intended interpretation and justify its use in the context of this analysis?

  2. Clarification of the reasoning in Section 4. Could the authors confirm whether the interpretation outlined in Weakness 1 is correct ?

1 comment
AuthorsOpenReview ↗

Rebuttal by Authors

We sincerely thank Reviewer bb1Y for their thoughtful and constructive review. We are especially grateful that the reviewer clearly acknowledged the contributions of our paper, including the proposed training-free soft-token method and our analysis on why soft tokens improve the parallel decoding. We believe that our paper makes the first attempt at analyzing why soft tokens improve parallel decoding in diffusion language models and can meaningfully contribute to the community. Below, we provide point-by-point responses to all questions and concerns raised. All the comments are constructive and help strengthen our work. We also appreciate that none of the points raised indicate fundamental flaws in our core ideas or methodology.

Q1. Justify the independency assumption in line 190-192

For each position j∈Sj\in S, let ZjZ_j denote its hypothetical discrete input state, distributed according to p(Zj∣E)p(Z_j\mid E). We define the joint distribution over these states using the product coupling

p(Z=z∣E)=∏j∈Sp(Zj=zj∣E),p(Z=z\mid E)=\prod_{j\in S}p(Z_j=z_j\mid E),

which corresponds to independently sampling each position’s discrete input state conditional on EE. This is a modeling choice for the reference distribution, not a claim that the positions are independent under the model’s true joint distribution.

For each joint realization zz, we replace the soft-token inputs at positions in SS with the corresponding discrete states and evaluate the frozen model, obtaining pθ(Y∣E,Z=z)p_\theta(Y\mid E,Z=z). We then average these predictions over the input realizations:

padd(y)=∑zp(Z=z∣E) pθ(Y=y∣E,Z=z).p_{\mathrm{add}}(y)=\sum_z p(Z=z\mid E)\,p_\theta(Y=y\mid E,Z=z).

Thus, this reference changes the inputs to discrete states and averages the resulting predictions; it is distinct from the model’s direct prediction given continuous soft-token embeddings.

Q2. Clarification of the reasoning in Section 4

The reviewer’s interpretation in Weakness is correct: Euclidean mixing can change both vector norms and directions, and these effects can arise in both the top-kk candidate aggregation and the interpolation between the resulting aggregate and the [MASK]\texttt{[MASK]} embedding. Our method addresses the latter but retains Euclidean averaging for the former. The reviewer’s concern about this asymmetry is well taken. We clarify the rationale here.

The choice reflects a trade-off between geometric fidelity and computational cost. As discussed in Section 4.1, angular information is important when interpolating between embeddings. Therefore, a natural geometric approach would be to perform both aggregation and interpolation on the sphere. Although SLERP provides a closed-form solution for interpolation between two points, aggregating more than two points on the sphere generally has no closed-form solution. A spherical Fréchet mean provides a geometrically principled way to aggregate them, but typically requires iterative Riemannian optimization. Because this aggregation would need to be performed for every soft-token position at every decoding step, it would add substantial computational cost.

We therefore retain the Euclidean weighted average for the top-kk candidates as a computationally efficient approximation. In our measurements, the candidate embeddings have higher pairwise cosine similarities with one another than with the [MASK]\texttt{[MASK]} embedding. This greater alignment motivates our expectation that Euclidean averaging introduces less geometric distortion when aggregating the candidates than when interpolating between their aggregate and the [MASK]\texttt{[MASK]} embedding. Euclidean averaging is also a standard and effective way to aggregate semantic embeddings in prior work [1]. We agree that relative alignment alone does not establish that the distortion is negligible. Our rationale is that the expected benefit of spherical aggregation at this stage must be weighed against its per-position iterative cost. We will make this trade-off and the distinction between the two stages more explicit in the revision.

Reference

[1] Zhen Zhang, Xuehai He, Weixiang Yan, Ao Shen, Chenyang Zhao, Xin Wang. Soft thinking: Unlocking the reasoning potential of LLMs in continuous concept space. NeurIPS 2025

Comments on the paper

1 comment

Thank you for the clarification and for updating the supplementary material and README. These changes improve the accessibility of the implementation. However, my main concerns are only partially resolved, particularly the discrepancy between the originally submitted configurations and the hyperparameters stated in the paper, whether the reported experiments consistently followed the stated protocol, and the limited uncertainty analysis across the main benchmark results. Therefore, I will maintain my score of 2 (weak rejection).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.