acceptodds
Under review as a conference paper at ICLR 2027

Anchor and Follower: Narrowing Valid Answers for Parallel Decoding in Diffusion Language Models

Abstract

Diffusion large language models (dLLMs) can accelerate decoding by committing tokens at multiple masked positions in parallel. However, when multiple valid answers remain compatible with the current context, parallel commitment can mix tokens from different answers and produce an output consistent with none of them. Sequential decoding reduces this risk by refining between commitments, suggesting a view of denoising as progressively narrowing the set of compatible valid answers. To study how predictions over valid answers change during denoising, we introduce *valid answer uncertainty* . Across a controlled graph task and three ParallelBench tasks, we find that commitments differ sharply in their effect on . Commitments that distinguish among the remaining valid answers can substantially reduce it, while shared commitments change it little. Intermediate refinement tends to reduce it further. Building on these findings, we propose **Anchor and Follower (AnF)**, a training-free decoding method that selects an anchor whose commitment can guide subsequent refinement toward lower answer uncertainty. AnF then commits the anchor, refines the remaining predictions, and commits high-confidence followers in parallel, combining sequential refinement with parallel efficiency. With LLaDA-8B-Instruct, AnF achieves 16.9% higher accuracy with 31.0% fewer NFEs than DAPD on HumanEval and the highest accuracy among compared methods on HumanEval, MBPP, and IFEval.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Diffusion Language Models, Parallel Decoding, Token Dependencies

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HKPNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Based on my understanding, this paper studies potential conflicts between tokens during parallel decoding in dLLMs. In the analysis part, it introduces valid answer uncertainty as a metric to quantify these conflicts. In the method part, it proposes AnF decoding, which works in three steps: selecting an anchor, refining the remaining predictions conditioned on that anchor, and then committing follower tokens in parallel. I find the research topic interesting, but I have concerns about the writing, the experimental setup, the method design and the evaluation results.

Critical strengths

S1. Personally, I find the topic of this paper interesting. However, I am not sure whether the broader community would see it as a key scientific problem, or whether the paper's novelty meets the expected bar. I will therefore take the other reviewers' opinions into account before making an overall assessment of novelty.

S2. The paper's entry point, token conflicts during parallel decoding, is interesting. To be honest, though, similar ideas exist in other works, albeit with a different framing.

S3. The method is relatively flexible.

Critical weaknesses

W1. The writing needs substantial improvement. A paper should convey its core motivation, key scientific question and research approach concisely, clearly and in an easy-to-follow way. At present, many statements in the paper, including in the abstract, introduction and method sections, are not sufficiently clear or rigorous.

W2. The reported results raise concerns about cherry-picking. I would like the authors to provide a more extensive and detailed hyperparameter sensitivity analysis, including a comparison of sensitivity across different models, to establish the robustness of the method.

W3. The anchor and follower concepts in this paper emphasize first committing an anchor during decoding so that it shapes the remaining context. These concepts are related to ADLM (https://openreview.net/pdf?id=E8adS5srds) and CoTA (https://openreview.net/pdf?id=mOz9jVYxsD), and the authors should discuss these works in the paper.

W4. The reliability of the valid answer uncertainty metric N_t still lacks supporting evidence. There is no rigorous mathematical proof of it, and the experiments lack statistical evidence. It would help to add a post-hoc statistical analysis of the actual conflicts that occur under parallel decoding, as corroborating evidence for N_t. Showing the correlation between these two statistics would make the metric considerably more convincing.

W5. The paper does not make clear the logical connection between its analysis and its proposed method. AnF also appears to perform poorly in Table 5. Moreover, the results in App. G.2 seem to contradict the paper's motivation. I therefore have doubts about the method's actual effectiveness, and I wonder whether the paper's narrative overstates what the method achieves.

W6. How common is cross-token inconsistency, and what drives it? Beyond the controlled tasks and illustrative examples such as Fig. 8, the paper does not show how common this phenomenon is on real benchmarks, so it is unclear whether the examples are representative. The authors should analyze two dimensions:

  1. Decoding steps. The effect may partly reflect model uncertainty in early decoding steps rather than inconsistency between tokens. The authors should break down conflicts by step, and separate tokens that are individually low-confidence from tokens that are each confident but jointly inconsistent.
  2. Position distance. The authors should report how far apart conflicting tokens typically are, and clarify how a "conflict" is defined for distant tokens.
Questions for authors

The evaluation should report the approximate caching and block settings in more detail.

5 comments

We appreciate the detailed and thoughtful feedback. We included additional comments to further improve our work.


[W1] Clarity of the Writing

Thank you for the feedback on the clarity of the writing. We revised the paper to clarify the motivation, central research question, and proposed approach. The revisions are highlighted in blue.

[A for W1] We revised the paper to clearly connect answer mixing, progressive answer narrowing, and Anchor and Follower. We also clarified the interpretation of NtN_t, strengthened the theoretical motivation for entropy-based anchor selection, and added direct evidence of answer mixing on generation benchmarks.


Details

I. Abstract and Introduction (Section 1)

We rewrote the abstract and introduction to clarify the core motivation, central research question, and proposed approach. We made explicit how answer mixing arises during parallel decoding and how this motivates using anchor-conditioned refinement to reduce mixing while retaining parallel efficiency.

II. Preliminaries (Section 3)

We clarified the meaning and interpretation of valid answer uncertainty NtN_t in Section 3.3.

III. Analysis (Section 4)

We added Section 4.3 to clarify the relationship between valid answer uncertainty NtN_t and conflict risk. We established a theoretical bound on conflict probability and empirically showed that low NtN_t is associated with low conflict risk across four analysis tasks.

IV. Method (Section 5)

We clarified the motivation for the two-stage anchor selection and added a theoretical connection between predictive entropy and NtN_t in Section 5.1.

V. Experiments (Section 6)

We added a direct evaluation of answer mixing across four generation benchmarks, reporting mixing rates alongside accuracy, GPU-forward throughput, and memory in Table 1.

We further analyze the distance and confidence of mixed tokens in Figure 8 to clarify how answer mixing occurs during parallel decoding.

[W2] Robustness Across Settings and Models

Thank you for the detailed comment regarding robustness.

All methods use greedy decoding, and we evaluate each baseline over parameter ranges based on its original paper. We report accuracy–NFE trade-offs across the evaluated settings and the highest-accuracy operating point for each method, rather than selecting a single favorable configuration.

Beyond LLaDA-8B-Instruct, we evaluated two additional models, LLaDA-1.5 and Dream-Instruct-7B, and further added LLaDA2.0-mini, using the same hyperparameter settings as LLaDA-8B-Instruct without model-specific tuning.

[A for W2] Across different models and parameter settings, AnF generally achieves the highest or competitive accuracy among parallel decoders while maintaining competitive efficiency. These results demonstrate the robustness of AnF across models and decoding configurations.


Details

I. Evaluation Protocol and Parameter Sensitivity

All methods use greedy decoding and are evaluated using the same prompts and evaluation procedures. For each baseline, we follow the configurations proposed in its original paper and sweep the method-specific parameters over the ranges specified in Appendix E.

Rather than reporting results from a single configuration, we present accuracy–NFE curves across multiple parameter settings for each method, allowing a more comprehensive comparison of accuracy–efficiency trade-offs.

We analyze AnF's sensitivity to the follower threshold τ\tau, target confidence γ\gamma, and candidate budget KK on LLaDA-8B-Instruct (Figure 21). Additional γ\gamma and τ\tau evaluations on LLaDA-1.5 and Dream-Instruct-7B are presented in Figures 25 and 26. The generation-length comparison is provided in Table 8.

II. Robustness Across Models

Table A. LLaDA-1.5 (Table 9)

MethodHumanEvalMBPPGSM8KIFEval
Acc.NFEAcc.NFEAcc.NFEAcc.NFE
Sequential (top-1)45.7512.048.2512.083.7512.064.91024.0
Threshold26.832.443.221.649.325.047.0120.8
PUNT25.0170.941.2112.169.7502.954.5540.0
DAPD40.255.147.5106.878.279.563.072.0
ETE34.8121.840.129.981.9121.059.3259.7
AnF46.379.249.829.178.578.263.6211.9

Table B. Dream-Instruct-7B (Table 10)

MethodHumanEvalMBPPGSM8KIFEval
Acc.NFEAcc.NFEAcc.NFEAcc.NFE
Sequential (top-1)49.4512.062.6512.083.5512.052.91024.0
Threshold28.0150.752.153.044.9165.229.820.2
PUNT23.2203.944.499.057.2151.248.4162.9
DAPD26.8243.127.2298.564.4108.531.1424.1
ETE29.964.349.0120.073.7120.136.4257.1
AnF50.031.665.837.684.865.960.8206.8

Table C. LLaDA2.0-mini (Table 11)

MethodHumanEvalMBPPGSM8KIFEval
Acc.NFEAcc.NFEAcc.NFEAcc.NFE
Sequential (baseline)73.2228.777.070.590.1276.082.6360.0
Threshold71.331.875.517.290.1116.381.9292.5
PUNT60.4161.466.578.388.9399.380.6703.0
ETE57.925.560.310.488.679.876.2176.7
AnF72.639.075.117.991.0108.981.5282.6

Across the evaluated models, AnF achieves the highest accuracy among parallel decoders on most benchmarks and remains competitive on the others, while maintaining favorable accuracy–NFE trade-offs.

These results show the robustness of AnF across different diffusion language models.


[W3] Discussion of ADLM and CoTA

Thank you for pointing out these related works.

[A for W3] We added ADLM and CoTA to Related Work (Section 2), discussing training-based anchor prediction in ADLM and how context tokens guide subsequent predictions in CoTA.

[W4] Reliability of Valid Answer Uncertainty NtN_t

Thank you for the detailed comment regarding the reliability of NtN_t. In this revision, We strengthened its justification through both theoretical and empirical analyses of its relationship with answer-mixing risk.

We added an analysis of the conflict probability CtC_t under independent parallel completion, deriving a theoretical bound relating $C_t$ to NtN_t and examining their empirical relationship across four analysis tasks.

The analysis is presented in Section 4.3, with the definition and proof in Appendix C.

[A for W4] We establish Ct≤1−1/NtC_t \le 1-1/\sqrt{N_t}, showing that low NtN_t guarantees low conflict probability under independent parallel completion. Empirically, states with low NtN_t exhibit low conflict probability CtC_t across all four tasks.


Details

I. Conflict

To examine how NtN_t relates to answer mixing, we compare it with the conflict probability CtC_t. A completion oo is conflicting if it contains at least one individually valid unit, but no single valid answer contains all of those units. We define

Conflict⁡(o)=1[∥x(o)∥1>0 ∧ ∥b(o)∥1=0],Ct=Eo∼pθ(⋅∣zt)[Conflict⁡(o)].\operatorname{Conflict}(o)=\mathbf 1\big[\|\mathbf x(o)\|_1>0\ \wedge\ \|\mathbf b(o)\|_1=0\big],\qquad C_t=\mathbb E_{o\sim p_\theta(\cdot\mid z_t)}\big[\operatorname{Conflict}(o)\big].

CtC_t measures the probability of answer mixing when all remaining masked positions are completed independently using the current predictions.

II. Relation between NtN_t and Conflict Probability

Theoretical relation. Conflicting completions contribute no answer contribution mass, so St=∑ist,i≤1−CtS_t=\sum_i s_{t,i}\le 1-C_t. Since ∑ist,i2≤St2\sum_i s_{t,i}^2\le S_t^2, we have St≥1/NtS_t\ge 1/\sqrt{N_t}. Combining these inequalities yields

Ct≤1−St≤1−1Nt.C_t\le 1-S_t\le 1-\frac{1}{\sqrt{N_t}}.

Thus, a low NtN_t guarantees a low conflict probability under independent parallel completion.

Empirical relation. We compare NtN_t with CtC_t, estimated from 512 independent completions at pre-commitment states of top-1, top-2, and top-4 decoding across four analysis tasks.

Table D. Correlation between NtN_t and CtC_t by NtN_t range

TaskNt≤5N_t\le5Nt≤10N_t\le10Nt≤100N_t\le100
Graph0.310.480.55
Replace Random0.810.590.09
Shuffle0.790.840.57
Latin Square0.960.960.98

Table D shows positive correlations between NtN_t and CtC_t across all four tasks for Nt≤10N_t\le10, ranging from 0.48 to 0.96.

These results show that low NtN_t is associated with low conflict probability, indicating that NtN_t reflects answer-mixing risk in practice.


[W6] Frequency and Characteristics of Answer Mixing on Generation Benchmarks

Thank you for your comment, which helps improve the readability of our paper.

We added a direct analysis of answer mixing across four generation benchmarks. We then measure the frequency of answer mixing and analyze the confidence and distance of mixed tokens. The analysis is presented in Section 6.2 and Appendix F.

[A for W6] Answer mixing increases from 3.7% under top-2 to 5.3% under top-4, while AnF achieves 1.6% despite committing more tokens per step. Mixed tokens are typically close together, and mixing can occur even at high confidence, showing that token confidence alone does not prevent answer mixing.


Details

I. Measuring Answer Mixing

We use GPT-6.1 to identify mixed parts in generated outputs, distinguishing answer mixing from ordinary prediction errors. We then match these parts to decoding traces and count mixing only when the identified parts are committed in the same decoding step.

The judging protocol and full prompt are provided in Appendix F.

II. Frequency of Answer Mixing

Table E. Mixing rate and tokens committed per step

MethodMixing (%)Tokens/step
Top-23.71.99
Top-45.33.95
Threshold2.62.51
Fast-dLLM1.42.33
PUNT6.06.88
DAPD8.53.64
ETE5.56.49
AnF1.65.11

Mixing increases with parallelism from top-2 to top-4. In contrast, AnF maintains a low mixing rate while committing more tokens per step than top-4.

These results are reported in Table 1, with additional results in Appendix F.

III. Characteristics of Answer Mixing

We analyze the distance between mixed parts and their confidence at commitment across decoding methods.

Table F. Distance and confidence of mixed tokens

MethodMedian distanceMean confidence
Top-220.63
Top-420.57
Threshold30.98
Fast-dLLM20.97
PUNT70.89
DAPD200.92
ETE30.80
AnF30.98

These results show that answer mixing occurs even at high token confidence, indicating that confident individual predictions can still produce cross-token inconsistency when committed in parallel.

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

The paper studies how parallel decoding in diffusion language models can mix incompatible valid answers. It introduces an answer-uncertainty measure and uses controlled tasks to study how token commitments affect it. The proposed AnF decoder selects a confident anchor using the entropy reduction after refinement, then commits high-confidence followers in parallel; experiments on three models and four generation benchmarks show promising accuracy–efficiency trade-offs.

Critical strengths
  • The HumanEval result is compelling: with LLaDA-8B-Instruct, AnF reaches 47.0% accuracy at 51.7 NFE, versus DAPD's 40.2% at 74.9 NFE. Table 1 also shows higher GPU-forward throughput, including candidate evaluation costs. The additional model results in Appendix F support the method's practical value.

  • The controlled comparisons help explain why intermediate refinement matters. In particular, the sequential and parallel arms commit the same set of positions, so the difference is not simply due to selecting different positions. This is a useful complement to the benchmark results.

Critical weaknesses

My main concern is how much the entropy score adds beyond choosing a confident anchor and refining once. The random-anchor comparison in Appendix C changes both the candidate pool and the selection rule, while the K ablation only tests 2–4 candidates. Neither isolates the benefit of Eq. (8). This matters because closely related lookahead strategies already exist in ETE [1] and LoPA [2].

[1] From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language Models

[2] LoPA: Scaling dLLM Inference via Lookahead Parallel Decoding

Questions for authors
  • How does AnF compare with a highest-confidence anchor (K = 1) and random selection within the same top-K candidate pool? Please keep the window, padding treatment, cache and follower threshold fixed, and report accuracy and runtime.

  • Among states that still admit a valid answer, how often does committing the selected anchor eliminate all valid answers? Appendix D excludes such states from the follower-role proportions; this rate would help interpret Figure 7.

3 comments

We appreciate the thoughtful feedback on our paper, and we included additional comments to further improve our work.


[W1, Q1] Contribution of the Entropy-Based Anchor Score

Thank you for the comment regarding the contribution of the entropy score.

To isolate the effect of the anchor score, we additionally conducted an anchor-score comparison using the same K=3K=3 candidate pool and decoder settings to evaluate entropy reduction against alternative selection rules.

Furthermore, we have now established a theoretical connection between entropy reduction and NtN_t through an upper bound on log⁡Nt\log N_t under an explicit assumption.

Entropy reduction achieves the highest average accuracy and lowest mean NFE.

[A for W1] With all decoder settings fixed and only the anchor score varied, entropy reduction achieves the highest average accuracy and lowest mean NFE. It outperforms ETE and performs comparably to LoPA. We also establish a theoretical connection to NtN_t, showing that entropy reduction tightens an upper bound on log⁡Nt\log N_t under an explicit assumption.


Details

I. Anchor Selection within the Same Candidate Pool

We compare anchor selection rules using LLaDA-8B-Instruct with K=3K=3 and follower threshold τ=0.9\tau=0.9, keeping all other decoder settings fixed. Decoding time is measured on an NVIDIA H200.

Table A. Accuracy (%), mean NFE, and decoding time (s) across anchor selection rules

Anchor ruleGSM8KHumanEvalMBPPIFEval
Acc.NFETimeAcc.NFETimeAcc.NFETimeAcc.NFETime
Top-1 confidence (K=1K=1)75.3107.12.9739.656.21.6649.033.80.8760.8195.68.18
Random in pool (3 seeds)74.4104.82.6741.953.01.5347.132.20.8359.7221.37.83
ETE score77.486.52.3441.544.41.3949.030.30.7960.4193.36.57
LoPA score79.583.72.1447.042.71.2647.929.00.8057.3183.76.53
Entropy (AnF)79.680.92.1546.342.71.2547.128.60.7960.1166.36.52

Compared with ETE, AnF achieves higher average accuracy with fewer forward passes. LoPA uses a similar scoring criterion based on post-refinement confidence, which may explain its comparable accuracy and runtime.

II. Relation between Entropy Reduction and NtN_t

We establish a theoretical connection between predictive entropy and NtN_t by deriving an upper bound on log⁡Nt\log N_t.

We assume that anchor-conditioned refinement does not decrease Λt\Lambda_t, which measures answer-contribution concentration relative to token-prediction concentration, i.e., Λt(a)≥Λt>0\Lambda_t^{(a)}\ge\Lambda_t>0.

Under this assumption, we obtain

log⁡Nt(a)≤∑j∈Mt∖{a}H(pt,j(a))−log⁡Λt.\log N_t^{(a)}\le\sum_{j\in\mathcal M_t\setminus\{a\}}H(p_{t,j}^{(a)})-\log\Lambda_t.

Thus, lower predictive entropy after refinement yields a lower upper bound on log⁡Nt\log N_t, providing a theoretical motivation for entropy-based anchor selection.

The theoretical relation is outlined in Section 5.1, with the full derivation in Appendix D.


[Q2] Elimination of Remaining Valid Answers

[A for Q2] Most AnF anchor commitments preserve at least one valid answer, particularly on ParallelBench. This clarifies the scope of the follower-role analysis in Figure 9 (previously Figure 7).

Thank you for the comment. We measured how often the selected anchor eliminates all valid answers among states with at least one valid answer remaining.

Table B. Rate of anchor commitments eliminating all remaining valid answers

TaskElimination rate
Graph13.4%
Replace Random0.3%
Shuffle0%
Latin Square1.6%

Table B shows that anchor elimination is rare on ParallelBench but more frequent on Graph. This indicates that Figure 9 (previously Figure 7) excludes few anchor events due to answer elimination on ParallelBench, but a larger fraction on Graph.

I appreciate the author's response. I am keeping my score.

Thank you for your response. We’re pleased that our explanation helped clarify your concerns. If you have any additional questions, we’d be happy to discuss them.

Reviewer_y4PfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper targets diffusion language models (DLMs) where parallel commitment can mix tokens from different answers and produce an output consistent with none of them. The authors find that commitments differ sharply in the effect on valid answer uncertainty. Commitments that distinguish among the remaining valid answers can substantially reduce it, while shared commitments change it little. Building on these findings, we propose Anchor and Follower (AnF), a training-free decoding method that selects an anchor whose commitment can guide subsequent refinement toward lower answer uncertainty.

Critical strengths
  • The topic is interesting where enhancing the parallel ability and remaining the accuacry is a central problem of existing DLMs.

  • The method is training-free and can be applied into different models.

  • The authors provide many detailed discussion and analyses.

Critical weaknesses
  • I think the effectiveness of the valid answer uncertainty is not discussed well. For example, the reducation of the valid answer uncertainty does not necessarily indicate fewer competing valid answers.

  • AnF proposes to make the token entropy reduction, which does not ensure that valid answer uncertainty and final correctness can be established.

  • Efficiency evaluation need to be enhanced. HumanEval GPU-forward throughput is encouraging, but broader end-to-end latency and memory measurements are needed.

  • The scale of the used model is about 7-8B, the author should test the proposed method on a lager-scale model such as llada 2.0 mini and more.

Questions for authors

please see the weakness

4 comments

We appreciate the thoughtful feedback on our paper, and we included additional comments to further improve our work.


[W1] Effectiveness of valid answer uncertainty NtN_t

Thank you for the comment regarding valid answer uncertainty. NtN_t does not count the number of remaining valid answers. Instead, it captures how much predictive mass can be coherently assigned to valid answers and how broadly that mass is distributed across them.

To further demonstrate the effectiveness of NtN_t, we conducted an additional analysis comparing it with the conflict probability CtC_t, which measures answer mixing under parallel completion. We proved that Ct≤1−1/NtC_t \le 1-1/\sqrt{N_t} and examined their correlation across four analysis tasks. We found that states with low NtN_t tend to exhibit low CtC_t across all four tasks.

The analysis is presented in Section 4.3, with further theoretical and experimental details in Appendix C. We also clarified this interpretation in the revised manuscript.

[A for W1] NtN_t captures the spread and total mass of answer contributions, rather than the number of remaining valid answers. We theoretically establish that low NtN_t guarantees low conflict probability and empirically find that low NtN_t is associated with low CtC_t across all four tasks.


Details

I. What NtN_t Represents

As defined in Section 3.3, NtN_t is the inverse of the sum of squared answer contribution masses st\mathbf{s}_t. Table A illustrates how the distribution and total mass of answer contributions affect NtN_t.

Table A. NtN_t for simple predictions

Prediction at ztz_tst\mathbf{s}_tNtN_t
All samples coherently form one answer(1,0,… )(1,0,\dots)1
Samples split evenly between two answers(0.5,0.5)(0.5,0.5)2
Half of the samples form one answer, the rest mix answers(0.5,0,… )(0.5,0,\dots)4

As illustrated in Table A, NtN_t can decrease through two factors:

  • More concentrated answer contributions: When predictive mass becomes concentrated on fewer valid answers, NtN_t decreases.
  • Higher total answer contribution mass: When fewer completions mix incompatible units and more predictive mass is coherently assigned to valid answers, NtN_t decreases.

Thus, tracking changes in NtN_t during denoising helps characterize how refinement affects the risk of answer mixing under parallel completion, rather than simply counting the remaining valid answers.

II. Relation between NtN_t and Conflict Probability

To examine how NtN_t relates to answer mixing, we compare it with the conflict probability CtC_t. A completion oo is conflicting if it contains at least one individually valid unit, but no single valid answer contains all of those units. We define

Conflict⁡(o)=1[∥x(o)∥1>0 ∧ ∥b(o)∥1=0],Ct=Eo∼pθ(⋅∣zt)[Conflict⁡(o)].\operatorname{Conflict}(o)=\mathbf 1\big[\|\mathbf x(o)\|_1>0\ \wedge\ \|\mathbf b(o)\|_1=0\big],\qquad C_t=\mathbb E_{o\sim p_\theta(\cdot\mid z_t)}\big[\operatorname{Conflict}(o)\big].

CtC_t measures the probability of answer mixing when all remaining masked positions are completed independently using the current predictions.

Theoretical relation. Conflicting completions contribute no answer contribution mass, so St=∑ist,i≤1−CtS_t=\sum_i s_{t,i}\le 1-C_t. Since ∑ist,i2≤St2\sum_i s_{t,i}^2\le S_t^2, we have St≥1/NtS_t\ge 1/\sqrt{N_t}. Combining these inequalities yields

Ct≤1−St≤1−1Nt.C_t\le 1-S_t\le 1-\frac{1}{\sqrt{N_t}}.

Thus, a low NtN_t guarantees a low conflict probability under independent parallel completion.

Empirical relation. We compare NtN_t with CtC_t, estimated from 512 independent completions at pre-commitment states of top-1, top-2, and top-4 decoding across four analysis tasks.

Table B. Correlation between NtN_t and CtC_t by NtN_t range

TaskNt≤5N_t\le5Nt≤10N_t\le10Nt≤100N_t\le100
Graph0.310.480.55
Replace Random0.810.590.09
Shuffle0.790.840.57
Latin Square0.960.960.98

Table B shows positive correlations between NtN_t and CtC_t across all four tasks for Nt≤10N_t\le10, ranging from 0.48 to 0.96.

These results show that low NtN_t is associated with low conflict probability, indicating that NtN_t reflects answer-mixing risk in practice.

We added Section 4.3 to present this analysis, with further details provided in Appendix C.

[W2] Entropy reduction, valid answer uncertainty, and final correctness

Thank you for the comment regarding the anchor selection score.

To clarify its connection to NtN_t, we added a theoretical derivation of an upper bound on log⁡Nt\log N_t using predictive entropy under an explicit assumption. AnF also achieved the highest accuracy among parallel decoding methods on three of four benchmarks.

We additionally conducted an anchor-score comparison and found that entropy reduction achieves the highest average accuracy*across four benchmarks.

The theoretical analysis is presented in Section 5.1 and Appendix D, with the anchor-score comparison in Table 7 of Appendix G.

[A for W2] Under an explicit assumption, reducing predictive entropy tightens an upper bound on log⁡Nt\log N_t. AnF achieves the highest accuracy among parallel decoders on three benchmarks, while its entropy-based anchor score yields the highest average accuracy.


Details

I. Relation between Entropy Reduction and NtN_t

As derived in Appendix D, NtN_t can be expressed in terms of the model's position-wise predictions:

log⁡Nt=−∑j∈Mtlog⁡∑vpt,j(v)2−log⁡Λt,\log N_t=-\sum_{j\in\mathcal M_t}\log\sum_v p_{t,j}(v)^2-\log\Lambda_t,

where Λt=∥st∥22∏j∈Mt∑vpt,j(v)2\Lambda_t=\frac{\|\mathbf{s}_t\|_2^2}{\prod_{j\in\mathcal M_t}\sum_v p_{t,j}(v)^2} captures the relationship between the concentration of token predictions and answer contribution mass.

We assume that committing a valid anchor and refining does not decrease Λt\Lambda_t, i.e., Λt(a)≥Λt>0\Lambda_t^{(a)}\ge\Lambda_t>0.

Since −log⁡∑vp(v)2≤H(p)-\log\sum_v p(v)^2\le H(p), this yields

log⁡Nt(a)≤∑j∈Mt∖{a}H(pt,j(a))−log⁡Λt.\log N_t^{(a)}\le\sum_{j\in\mathcal M_t\setminus\{a\}}H(p_{t,j}^{(a)})-\log\Lambda_t.

Thus, reducing predictive entropy after anchor-conditioned refinement tightens an upper bound on log⁡Nt\log N_t under the stated assumption.

AnF uses normalized entropy reduction as a practical criterion for anchor selection, without requiring NtN_t to be computed during decoding.

II. Accuracy against Decoding Baselines

Table C compares the best achieved accuracy of AnF and the baselines across four generation benchmarks.

Table C. Accuracy on generation benchmarks (%)

MethodHumanEvalMBPPGSM8KIFEval
Sequential45.147.584.260.8
Top-234.139.382.352.9
Top-420.126.177.046.6
Threshold44.547.584.460.6
Fast-dLLM37.840.584.053.6
PUNT21.341.264.250.6
DAPD40.246.376.659.1
ETE36.640.181.753.6
AnF47.049.880.861.0

Table C shows that AnF achieves the highest accuracy among parallel decoders on three of four benchmarks, except GSM8K, where Threshold performs best.

III. Accuracy across Anchor Scores

To isolate the effect of the anchor score, we added a comparison of entropy reduction with alternative scores while keeping the candidate pool and other decoder settings fixed.

Table D. Accuracy and NFE across anchor scores

Anchor scoreGSM8KHumanEvalMBPPIFEvalAvg.
Acc.NFEAcc.NFEAcc.NFEAcc.NFEAcc.NFE
Random71.666.729.344.326.830.158.6174.946.679.0
Top-1 confidence75.3107.139.656.249.033.860.8195.656.298.2
ETE score77.486.541.544.449.030.360.4193.357.188.6
LoPA score79.583.747.042.747.929.057.3183.757.984.8
Entropy (AnF)79.680.946.342.747.128.660.1166.358.379.6

Table D shows that entropy-based anchor selection achieves the highest average accuracy of 58.3% across four benchmarks, with fewer or equal NFEs than the other scores.

These results demonstrate the effectiveness of entropy reduction for anchor selection. The full comparison is provided in Table 7 of Appendix G.

[W3] GPU-Forward Throughput and Memory Usage

Thank you for the suggestion. We extended the GPU-forward throughput measurements to all four benchmarks and added peak GPU memory measurements for each method.

AnF achieves the highest throughput on HumanEval and MBPP, remains competitive on GSM8K and IFEval, and uses comparable GPU memory.

[A for W3] We now report throughput and GPU memory across all four benchmarks. AnF achieves the highest throughput on the two code benchmarks while maintaining competitive performance and comparable memory usage.


Details

We measured GPU-forward throughput and peak memory using LLaDA-8B-Instruct on an NVIDIA H200, with each method evaluated at its best-performing setting reported in Table 1.

Table E. GPU-forward throughput (tokens/s) and peak memory (GB)

MethodHumanEvalMBPPGSM8KIFEvalMemory
Sequential33.636.336.021.917.2
Top-267.973.072.444.217.2
Top-4135.4145.3144.388.317.2
Threshold276.5403.7153.488.316.9
Fast-dLLM222.2363.3149.9151.518.6
PUNT30.556.588.5116.617.7
DAPD231.3171.2224.5646.516.8
ETE125.2250.4127.068.020.1
AnF362.2599.3194.7138.018.9

The results are reported in Table 1 of Section 6.1, with measurement details in Appendix E.


[W4] Evaluation on LLaDA2.0-mini

Thank you for the suggestion. We additionally evaluated AnF on LLaDA2.0-mini, a 16B MoE model, across four benchmarks without model-specific tuning.

AnF achieves the highest accuracy among parallel decoders on HumanEval and GSM8K, while remaining comparable to the official decoder on MBPP and IFEval.

The results are presented in Table 11 of Appendix I.

[A for W4] AnF performs competitively on the larger 16B MoE model without retuning, achieving the highest accuracy on two benchmarks and comparable performance to the official decoder on the other two.


Details

We evaluated each method using its recommended sampling settings for LLaDA2.0-mini, with a temperature of 0.0, a block length of 32, and 32 denoising steps per block.

Table F. Accuracy (%) and mean NFE on LLaDA2.0-mini

MethodHumanEvalMBPPGSM8KIFEval
Acc.NFEAcc.NFEAcc.NFEAcc.NFE
Sequential (baseline)73.2228.777.070.590.1276.082.6360.0
Threshold71.331.875.517.290.1116.381.9292.5
PUNT60.4161.466.578.388.9399.380.6703.0
ETE57.925.560.310.488.679.876.2176.7
AnF72.639.075.117.991.0108.981.5282.6

AnF achieves the highest accuracy among parallel decoders on HumanEval and GSM8K, with 39.0 and 108.9 NFEs. On MBPP and IFEval, AnF achieves accuracy within 0.4 percentage points of the best parallel baselines, with comparable or fewer NFEs.

Thank you again for your thoughtful review. Your comments helped us clarify the theory and strengthen the evaluation. We summarize the main changes below.

  1. Valid answer uncertainty. We clarified that NtN_t reflects both the total mass and the spread of answer contributions. We derived Ct≤1−1/NtC_t \leq 1-1/\sqrt{N_t} under independent parallel completion and found that low NtN_t was associated with low conflict probability across all four tasks.
  2. Entropy reduction and accuracy. Under our stated assumption, entropy reduction tightens an upper bound on log⁡Nt\log N_t. With the candidate pool and decoder settings fixed, it achieved the highest average accuracy across four benchmarks.
  3. Efficiency. We reported GPU-forward throughput and peak memory across all four benchmarks. AnF achieved the highest throughput on HumanEval and MBPP.
  4. Larger-model evaluation. We evaluated AnF on LLaDA2.0-mini without retuning. It achieved the highest accuracy among the tested parallel decoders on HumanEval and GSM8K.

As the discussion period draws to a close, we would be grateful to hear whether our responses have addressed your main concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.