pith. sign in

arxiv: 2606.06031 · v1 · pith:IO4PICALnew · submitted 2026-06-04 · 💻 cs.CL

NAVIRA: Decoupled Stochastic Remasking for Masked Diffusion Language Models

Andrey Fomenko , Maksim Kryzhanovskiy , Svetlana Glazyrina , Roman Ischenko This is my paper

Pith reviewed 2026-06-28 01:40 UTC · model grok-4.3

classification 💻 cs.CL
keywords masked diffusionlanguage modelsremaskinginference policystochastic samplingtext generationquality scoringparallel decoding
0
0 comments X

The pith

Decoupling token quality scoring from regeneration improves fluency in masked diffusion language models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Masked diffusion models generate text by unmasking many tokens in parallel steps, but early errors in one step can corrupt later context because all tokens in a step are drawn from marginal distributions. NAVIRA addresses the correction problem by running two separate forward passes: the first scores each token for quality, selected low-quality tokens are masked, and the second pass regenerates replacements from the now-cleaned context. Stochastic selection of which tokens to remask, controlled by temperature, avoids repeatedly correcting the same positions and trades off fluency against output diversity. In experiments with a 170M model this decoupled policy produces higher fluency and stronger LLM-judge scores than coupled remasking methods when extra forward passes are available.

Core claim

NAVIRA decouples the quality-scoring and regeneration operations plus temperature-controlled stochastic remasking. A first forward pass produces token quality scores; unreliable tokens are masked; a second forward pass then computes replacement logits from the cleaned context. Temperature-controlled stochastic remasking reduces repeated correction of the same positions and balances fluency against diversity. In controlled experiments decoupling improves fluency while scheduled stochastic remasking preserves entropy and achieves stronger LLM-judge scores under larger forward-pass budgets.

What carries the argument

Decoupled two-pass inference with temperature-controlled stochastic remasking, where scoring and logit computation occur in separate forward passes.

If this is right

  • Regeneration occurs without the erroneous tokens still in context, reducing error propagation.
  • Stochastic rather than deterministic remasking keeps output entropy from collapsing.
  • LLM-judge scores rise when the budget allows the extra forward pass required by decoupling.
  • Remasking policy itself, not only the learned quality signal, becomes a central lever for generation quality.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same separation of scoring and regeneration passes could be tested on other iterative parallel decoding schemes that suffer from early local errors.
  • Gains may depend on how well the quality scorer generalizes across domains or prompt lengths.
  • If the two-pass overhead is small, the method could be combined with larger models without retraining.

Load-bearing premise

Token quality scores from the first forward pass reliably identify positions whose regeneration in the second pass will improve the final sequence.

What would settle it

Run the same 170M masked diffusion model with and without the second regeneration pass on identical prompts and check whether LLM-judge preference for the decoupled outputs disappears.

Figures

Figures reproduced from arXiv: 2606.06031 by Andrey Fomenko, Maksim Kryzhanovskiy, Roman Ischenko, Svetlana Glazyrina.

Figure 1
Figure 1. Figure 1: Overview of the proposed NAVIRA step. A first forward pass computes quality scores on the current state, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Temperature-controlled stochastic remasking. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between PRISM, NAVIRA-DET, and basic Dream-style remasking baselines on the 170M MDM. NAVIRA-DET achieves the best overall quality–diversity trade-off, while the entropy- and margin-based heuristics exhibit strong entropy collapse as the number of forward passes grows. 64 128 256 512 1024 2048 Forward passes 10 1 Perplexity ↓ Perplexity under stochastic remasking 64 128 256 512 1024 2048 Forward… view at source ↗
Figure 4
Figure 4. Figure 4: Deterministic versus stochastic remasking without temperature scheduling. Stochastic [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: LLM-as-a-judge evaluation using Qwen3-235B-A22B-Instruct-2507. Stochastic [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
read the original abstract

Masked diffusion language models generate text by iteratively unmasking many tokens in parallel, but this speed comes with a correction problem: tokens generated in the same step are predicted from marginal distributions, and early local dependency errors can later contaminate the context. PRISM addresses this by learning token-level quality scores and remasking unreliable tokens, but its inference rule is coupled: the same forward pass both detects low-quality tokens and computes logits for their replacements, so the erroneous tokens still condition regeneration. We propose NAVIRA, an inference-time decoding policy that separates these two operations and samples remasking positions stochastically. A first forward pass scores tokens; selected tokens are masked; a second forward pass regenerates from the cleaned context. Temperature-controlled remasking reduces repeated correction of the same positions and balances fluency against diversity. In controlled experiments with a 170M masked diffusion language model, decoupling improves fluency, while scheduled stochastic remasking preserves entropy and achieves stronger LLM-judge scores under larger forward-pass budgets. These results show that remasking policy, not only the learned quality signal, is central to reliable masked-diffusion text generation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper proposes NAVIRA, an inference-time decoding policy for masked diffusion language models that decouples token quality scoring from regeneration via two separate forward passes, with temperature-controlled stochastic remasking of low-quality tokens. This addresses contamination from early local errors in parallel unmasking (as in PRISM) by regenerating from cleaned context. Controlled experiments on a 170M model claim that decoupling improves fluency, while the stochastic schedule preserves entropy and yields stronger LLM-judge scores under increased forward-pass budgets, showing that remasking policy matters beyond the learned quality signal.

Significance. If the empirical results hold, the work highlights that inference-time remasking policies can meaningfully improve generation quality in masked diffusion LMs without retraining. The controlled experimental setup with a fixed 170M model and focus on LLM-judge metrics under varying budgets is a strength, as is the emphasis on balancing fluency and diversity via temperature scheduling. No machine-checked proofs or parameter-free derivations are present.

major comments (1)
  1. [Experiments] The central claim that decoupling plus stochastic remasking produces net gains (stronger LLM-judge scores) rests on the untested assumption that the first-pass quality scores identify positions where regeneration from cleaned context yields improvement rather than trading one set of marginal predictions for another. The manuscript provides no correlation analysis between quality scores and post-remask delta, nor a random-remasking control, which is load-bearing for attributing gains to the quality signal rather than extra compute.
minor comments (1)
  1. [Abstract] The abstract states performance gains but supplies no quantitative results, baseline details, statistical tests, or ablation tables; adding these (with exact metrics and model details) would improve clarity.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the single major comment below and indicate planned revisions.

read point-by-point responses

  1. Referee: [Experiments] The central claim that decoupling plus stochastic remasking produces net gains (stronger LLM-judge scores) rests on the untested assumption that the first-pass quality scores identify positions where regeneration from cleaned context yields improvement rather than trading one set of marginal predictions for another. The manuscript provides no correlation analysis between quality scores and post-remask delta, nor a random-remasking control, which is load-bearing for attributing gains to the quality signal rather than extra compute.

    Authors: We agree that the current experiments do not include an explicit correlation between quality scores and post-remask improvement deltas, nor a random-remasking ablation, leaving open the possibility that gains partly reflect extra compute rather than the quality signal. Our reported results compare NAVIRA against PRISM under matched forward-pass budgets on the same 170M model, isolating the effect of the second forward pass on cleaned context; the LLM-judge gains and fluency improvements are therefore tied to the decoupling mechanism rather than raw budget alone. Nevertheless, a random-remasking control would provide stronger causal evidence. We will add both the requested correlation analysis and a random-remasking baseline in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents NAVIRA as an inference-time decoding policy that decouples token quality scoring from regeneration via two separate forward passes plus stochastic remasking. All reported gains are measured empirically against baselines in controlled experiments on a 170M model; no equations, fitted parameters, or self-citations are invoked that would make the claimed improvements equivalent to the inputs by construction. The method description is procedural and externally falsifiable via the LLM-judge and fluency metrics.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the method is presented as a policy change rather than a new learned component.

pith-pipeline@v0.9.1-grok · 5737 in / 1093 out tokens · 42638 ms · 2026-06-28T01:40:49.057213+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 17 canonical work pages · 8 internal anchors

  1. [1]

    Advances in Neural Information Processing Systems , volume =

    Structured Denoising Diffusion Models in Discrete State-Spaces , author =. Advances in Neural Information Processing Systems , volume =. 2021 , url =

    2021

  2. [2]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution , author =. Proceedings of the 41st International Conference on Machine Learning , pages =. 2024 , editor =

    2024

  3. [3]

    Advances in Neural Information Processing Systems , volume =

    Simple and Effective Masked Diffusion Language Models , author =. Advances in Neural Information Processing Systems , volume =. 2024 , doi =

    2024

  4. [4]

    Large Language Diffusion Models

    Large Language Diffusion Models , author =. 2025 , eprint =. doi:10.48550/arXiv.2502.09992 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.09992 2025

  5. [5]

    Dream 7B: Diffusion Large Language Models

    Ye, Jiacheng and Xie, Zhihui and Zheng, Lin and Gao, Jiahui and Wu, Zirui and Jiang, Xin and Li, Zhenguo and Kong, Lingpeng , year =. doi:10.48550/arXiv.2508.15487 , url =. 2508.15487 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2508.15487

  6. [6]

    Mercury: Ultra-Fast Language Models Based on Diffusion

    Mercury: Ultra-Fast Language Models Based on Diffusion , author =. 2025 , eprint =. doi:10.48550/arXiv.2506.17298 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.17298 2025

  7. [7]

    CoRR , volume =

    Gong, Shansan and Zhang, Ruixiang and Zheng, Huangjie and Gu, Jiatao and Jaitly, Navdeep and Kong, Lingpeng and Zhang, Yizhe , year =. doi:10.48550/arXiv.2506.20639 , url =. 2506.20639 , archivePrefix =

    work page doi:10.48550/arxiv.2506.20639

  8. [8]

    doi:10.48550/arXiv.2602.01326 , url =

    Wu, Zirui and Zheng, Lin and Xie, Zhihui and Ye, Jiacheng and Gao, Jiahui and Gong, Shansan and Feng, Yansong and Li, Zhenguo and Bi, Wei and Zhou, Guorui and Kong, Lingpeng , year =. doi:10.48550/arXiv.2602.01326 , url =. 2602.01326 , archivePrefix =

    work page doi:10.48550/arxiv.2602.01326

  9. [9]

    Seed Diffusion: A Large-Scale Diffusion Language Model with High-Speed Inference

    Seed Diffusion: A Large-Scale Diffusion Language Model with High-Speed Inference , author =. 2025 , eprint =. doi:10.48550/arXiv.2508.02193 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2508.02193 2025

  10. [10]

    doi:10.48550/arXiv.2503.00307 , url =

    Wang, Guanghan and Schiff, Yair and Sahoo, Subham Sekhar and Kuleshov, Volodymyr , year =. doi:10.48550/arXiv.2503.00307 , url =. 2503.00307 , archivePrefix =

    work page doi:10.48550/arxiv.2503.00307

  11. [11]

    Fine-Tuning Masked Diffusion for Provable Self-Correction

    Fine-Tuning Masked Diffusion for Provable Self-Correction , author =. 2025 , eprint =. doi:10.48550/arXiv.2510.01384 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2510.01384 2025

  12. [12]

    Gokaslan, Aaron and Cohen, Vanya and Pavlick, Ellie and Tellex, Stefanie , year =

  13. [13]

    Qwen2.5 Technical Report

    2024 , eprint =. doi:10.48550/arXiv.2412.15115 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115 2024

  14. [14]

    2026 , eprint =

    Gumbel Distillation for Parallel Text Generation , author =. 2026 , eprint =. doi:10.48550/arXiv.2603.22216 , url =

    work page doi:10.48550/arxiv.2603.22216 2026

  15. [15]

    Advances in Neural Information Processing Systems , volume =

    Simplified and Generalized Masked Diffusion for Discrete Data , author =. Advances in Neural Information Processing Systems , volume =. 2024 , doi =

    2024

  16. [16]

    Mask-Predict: Parallel Decoding of Conditional Masked Language Models , author =. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , month =. 2019 , address =. doi:10.18653/v1/D19-1633 , url =

    work page doi:10.18653/v1/d19-1633 2019

  17. [17]

    Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data

    Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data , author =. 2024 , eprint =. doi:10.48550/arXiv.2406.03736 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.03736 2024

  18. [18]

    2024 , eprint =

    Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling , author =. 2024 , eprint =. doi:10.48550/arXiv.2409.02908 , url =

    work page doi:10.48550/arxiv.2409.02908 2024

  19. [19]

    Advances in Neural Information Processing Systems , volume =

    A Continuous Time Framework for Discrete Denoising Models , author =. Advances in Neural Information Processing Systems , volume =. 2022 , url =

    2022

  20. [20]

    Advances in Neural Information Processing Systems , volume =

    Discrete Flow Matching , author =. Advances in Neural Information Processing Systems , volume =. 2024 , doi =

    2024

  21. [21]

    Computer Vision -- ECCV 2022 , series =

    Improved Masked Image Generation with Token-Critic , author =. Computer Vision -- ECCV 2022 , series =. 2022 , doi =

    2022

  22. [22]

    2025 , eprint =

    Don't Settle Too Early: Self-Reflective Remasking for Diffusion Language Models , author =. 2025 , eprint =. doi:10.48550/arXiv.2509.23653 , url =

    work page doi:10.48550/arxiv.2509.23653 2025

  23. [23]

    2024 , eprint =

    Informed Correctors for Discrete Diffusion Models , author =. 2024 , eprint =. doi:10.48550/arXiv.2407.21243 , url =

    work page doi:10.48550/arxiv.2407.21243 2024

  24. [24]

    Proceedings of the 42nd International Conference on Machine Learning , series =

    Generalized Interpolating Discrete Diffusion , author =. Proceedings of the 42nd International Conference on Machine Learning , series =. 2025 , url =

    2025

  25. [25]

    2025 , eprint =

    Path Planning for Masked Diffusion Model Sampling , author =. 2025 , eprint =. doi:10.48550/arXiv.2502.03540 , url =

    work page doi:10.48550/arxiv.2502.03540 2025

  26. [26]

    Plan for Speed: Dilated Scheduling for Masked Diffusion Language Models

    Plan for Speed: Dilated Scheduling for Masked Diffusion Language Models , author =. 2025 , eprint =. doi:10.48550/arXiv.2506.19037 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.19037 2025

  27. [27]

    International Conference on Learning Representations , year =

    The Curious Case of Neural Text Degeneration , author =. International Conference on Learning Representations , year =

  28. [28]

    and Zhang, Hao and Gonzalez, Joseph E

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric P. and Zhang, Hao and Gonzalez, Joseph E. and Stoica, Ion , booktitle =. Judging. 2023 , doi =

    2023