Pith. sign in

REVIEW 4 major objections 5 minor 25 references

AdaLook makes lookahead decoding adaptive by rolling out deeper exactly when candidate scores disagree, yielding a better accuracy-efficiency trade-off than one-step lookahead.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 22:40 UTC pith:2D3NRF7S

load-bearing objection Sensible adaptive lookahead with a real empirical signal, but the central variance threshold is never reported and the paper contradicts itself about how it works. the 4 major comments →

arxiv 2607.15655 v1 pith:2D3NRF7S submitted 2026-07-17 cs.CL cs.LG

Adaptive Multi-Step Lookahead Decoding for Diffusion Language Models

Yingqian Cui , Wei Deng , Lantao Mei , Hang Li , Charu C. Aggarwal , Hui Liu , Yue Xing This is my paper
classification cs.CL cs.LG
keywords diffusion language modelsmasked diffusionlookahead decodingadaptive rolloutvariance-gated stoppingbranch expansionparallel decodingaccuracy-efficiency trade-off
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Masked diffusion language models refine masked tokens in parallel, and lookahead decoding can improve the accuracy-efficiency trade-off by exploring future states before committing tokens. The paper claims that one-step lookahead is too shallow and fixed-depth multi-step rollout is wasteful, because the right rollout depth varies across examples and decoding stages. It proposes AdaLook, which after each rollout step computes the variance of candidate scores: low variance means keep looking, high variance means stop and commit. It also re-triggers lookahead from intermediate branches when a rollout state is still uncertain. On math, reasoning, and knowledge benchmarks, AdaLook achieves better accuracy for a given number of decoding steps than one-step lookahead and confidence-aware parallel decoding, with only about 7% higher per-step latency than its one-step counterpart on modern GPUs.

Core claim

The central claim is that adaptive multi-step lookahead decoding, gated by candidate-score variance, dominates both shallow one-step lookahead and naive fixed-depth rollout for masked diffusion language models. AdaLook keeps k=4 candidate trajectories alive, and after each rollout step computes a cumulative score that adds the log-confidence mass unlocked across all rollout steps (Equation 1). It continues rollout only when the variance of these scores is below a threshold tau, meaning the candidates have not yet separated into a clear winner; once variance exceeds tau, or the maximum depth Tmax=2 is reached, it commits the highest-scoring hypothesis. To handle heterogeneous intermediate sta

What carries the argument

The load-bearing mechanism is the variance gate on cumulative candidate scores. AdaLook maintains k=4 candidate trajectories; each is scored by Eq. (1), the log of the explored candidate's own confidence plus the log sum of downstream high-confidence confidence mass unlocked across R rollout steps. Before each additional rollout step the decoder computes Var({s(R)(j)}) across candidates: if it is below threshold tau, the candidates are not yet discriminative and it rolls out one more step; if it exceeds tau, it commits argmax. A second mechanism, dynamic branch expansion, re-evaluates each branch after every forward pass using the original ETE lookahead-trigger condition and handles the thre

Load-bearing premise

The load-bearing premise is that variance among the k=4 candidate rollout scores—after adding up unlocked-confidence mass over a variable number of steps—reliably indicates whether another rollout step will separate good from bad candidates, and that one fixed cut-off value (tau) can serve across examples, decoding stages, and benchmarks; the paper neither derives this threshold nor reports its value.

What would settle it

Measure the distribution of Var({s(R)(j)}) at each rollout step on a held-out set like MATH500. If the variance almost always sits above any reasonable tau after one step, AdaLook collapses into one-step lookahead; if it almost always sits below tau up to Tmax, it collapses into fixed two-step rollout. Either pattern would show the variance gate is not the source of the reported trade-off improvement.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Fixed-depth multi-step rollout is not the right extension of one-step lookahead; state-dependent depth is where the accuracy-per-step gain comes from.
  • On harder reasoning benchmarks like MATH and BBH, adaptive lookahead yields larger gains, suggesting deeper exploration matters most when local confidence is unreliable.
  • At a fixed number of decoding steps, AdaLook reaches higher maximum accuracy than one-step lookahead on MATH, MMLU, and BBH, so the method is not merely spending more computation.
  • Code generation behaves differently: both one-step and adaptive multi-step lookahead stay close to plain block diffusion sampling, indicating confidence-based lookahead signals are too local for long-range syntactic dependencies.
  • The extra latency of adaptive multi-step rollout shrinks on stronger GPUs, implying that the decoding-step savings translate into wall-clock gains as hardware batch efficiency improves.

Where Pith is reading between the lines

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

  • The paper never reports the value of the variance threshold tau, even though it is the sole new gate that separates AdaLook from a fixed-depth rollout; a reader cannot reproduce the method's stopping behavior without it.
  • If tau is chosen poorly, AdaLook degenerates: a threshold that is too high makes variance immediately exceed it (so every lookahead stops after one step, reducing to ETE), while too low a threshold keeps rolling to Tmax on every example (reducing to fixed 2-step rollout). The reported gains therefore hinge on a single number the paper leaves unstated.
  • The same variance-gating idea could be tested in other search/decoding settings—e.g., lookahead in autoregressive sampling or tree search—where candidate scores accumulate over depth and a natural stopping rule is needed.
  • A testable variant would set tau per decoding stage or per remaining-mask count rather than globally, since early decoding steps may tolerate more rollout than late ones.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes AdaLook, an adaptive multi-step lookahead decoding method for masked diffusion language models. AdaLook extends the one-step Explore-then-Exploit (ETE) framework by (i) continuing rollout while the variance of candidate cumulative scores is below a threshold tau and (ii) dynamically re-triggering or pruning branches based on the ETE confidence-trigger condition. The authors claim that this adaptive rollout and branch-expansion mechanism improves the accuracy-versus-decoding-steps trade-off compared to ETE and Fast-dLLM, and they report results on LLaDA-8B-Instruct and Dream-7B across several benchmarks, with higher maximum accuracy on MATH (43.6% vs. 42.6% for ETE). The manuscript also reports latency measurements and a discussion of code-generation tasks where lookahead provides limited benefit.

Significance. If the empirical ordering AdaLook > ETE > Fast-dLLM holds with the reported calibration discipline, the paper would provide a useful and non-obvious extension of lookahead decoding for masked diffusion LMs. The comparison against external baselines (ETE, Fast-dLLM) on public benchmarks, with two backbones, is a strength, as is the explicit latency analysis and the candid limitation regarding code generation. However, the central mechanism—the variance gate controlled by tau—is not specified, is described with contradictory stopping conditions in Section 1 vs. Section 4.2/Algorithm 1, and is never isolated through fixed-depth or tau-ablation controls. Because the adaptive early-stopping rule is the paper's core contribution, the manuscript is not reproducible as written and the claimed benefit of adaptivity is not yet supported.

major comments (4)
  1. [Section 1 vs. Section 4.2, Algorithm 1 Step 3] The variance-gate rule is described inconsistently. Section 1 says 'Only when the variance exceeds a predefined threshold does the decoder continue expanding future decoding trajectories,' while Section 4.2 states that if variance falls below tau the decoder proceeds with another rollout step, and Algorithm 1 Step 3 continues rollout when Var({s_j}) < tau. These are opposite conditions. Moreover, tau is never assigned a value: the fixed-hyperparameter list in Section 4.3 omits it, and the calibration subsection does not discuss it. Since tau is the sole quantity controlling rollout depth, the core adaptive mechanism cannot be reproduced or checked. Please correct the description, report the value(s) used, and include a sensitivity analysis around tau.
  2. [Section 5.2, Figure 3] The main curves compare AdaLook against one-step ETE and Fast-dLLM, but no fixed-depth rollout baseline is reported. The paper's own motivation is that fixed-depth rollout is ineffective (Figure 1 and L1/L2), and the adaptive gain is claimed to come from avoiding that fixed-depth failure. Without a control that uses the same k, Tmax, and branch-expansion rules but with a fixed depth (e.g., tau=0 reducing to ETE and tau=infinity giving full-depth rollout, or a fixed R=2 baseline), the reported gains cannot be attributed to the variance gate. They could instead arise simply from the extra depth and branch expansion. Please add such controls and, if possible, a tau ablation on at least one benchmark.
  3. [Eq. (1), Section 4.2] The variance-based early-stopping signal is not justified in a way that accounts for the cumulative definition of the score. The score in Eq. (1) sums downstream confidence gains over R rollout steps, so as R grows the scores are sums of more nonnegative terms. The variance of such cumulative sums can increase mechanically with R, independent of whether the candidates are becoming more or less discriminative. If the threshold is reached mainly because the accumulated sums grow with rollout depth, the gate does not measure 'whether candidates remain not sufficiently discriminative,' as claimed. Please provide a derivation, normalize the score by R or use per-step variances, and/or show empirically that the variance dynamics actually separate informative from uninformative continuation cases.
  4. [Section 5.1 and 4.3 'optimized' setting] The 'optimized' curves scan all combinations of gamma in [0.1,0.4] and calibrated N/C pairs and report 'the best Pareto-frontier results.' This is a selection procedure on the evaluation benchmark unless the Pareto frontier is chosen on a separate calibration set and then evaluated on the benchmark. The manuscript does not state explicitly which points are selected on which data, how many configurations were run, or whether the same selection rule was applied to all methods. While the standard gamma=0.2 setting is less suspect, the optimized setting as described can overstate the achievable trade-off. Please specify the selection protocol, the calibration set, and the number of runs, and report the standard-setting results with error bars or configuration counts.
minor comments (5)
  1. [Section 4.3] Typographical issue: the fixed-hyperparameter list contains a double comma: 'beta=0.01, , Tmax=2.' Please fix.
  2. [Algorithm 1] The control flow after the mixed-branch case is ambiguous. Line 35 ('j* <- argmax_j s_j; go to Commit') appears to be outside the for-loop body, but its placement relative to the Case 1/Case 2/Case 3 branches is not clear from indentation. Reformulate the pseudocode with explicit branch labels and ensure each path either commits or continues to the next iteration.
  3. [Section 5.3, Table 1] Table 1 reports per-step latency only. To support the practical-efficiency claim, it would help to also report end-to-end wall-clock time for matched accuracy levels (or total latency), since AdaLook may use more batched forward passes even at the same number of reported decoding steps. Without this, the 7% per-step overhead does not directly translate into a statement about total runtime.
  4. [Figure 3] The curves in Figure 3 are dense and the legend/marker scheme is not described. Consider labeling the standard-setting and optimized-setting curves more clearly, and possibly plotting only a subset of configurations for readability.
  5. [Section 5.4] The 'Fast Block Diffusion Sampling' baseline is mentioned in the code-generation discussion but is not defined in the experimental setup. Add a one-sentence definition or a pointer to the relevant prior work.

Circularity Check

0 steps flagged

No significant circularity: AdaLook's headline trade-off is measured against external baselines; the only suspicious item (the C-N calibration 'justification') is explicitly empirical and post hoc, not a fitted parameter disguised as a prediction.

full rationale

I walked the claimed derivation chain. The central claim—that adaptive multi-step lookahead improves the accuracy-vs-decoding-steps trade-off relative to one-step lookahead (ETE) and Fast-dLLM—is supported by public-benchmark comparisons against external prior work (Fu et al. 2025; Wu et al. 2025). No load-bearing reference is authored by the present paper's authors, so the self-citation patterns do not apply. Eq. (1) is a direct extension of the external ETE score; it is an algorithmic scoring rule, not a fitted parameter later renamed as a prediction. The hyperparameters k=4, alpha=0.1, beta=0.01, Tmax=2, cinfo=0.2 are either inherited from prior work or chosen on a calibration set, and the C-N log guideline is explicitly empirical: Section 4.3 says the 'exact threshold is chosen empirically on the calibration set,' and Appendix B states it combines 'this empirical approximation with Eq.(2)' to 'provide a theoretical justification.' That is a post-hoc rationalization of a calibration-set trend rather than a first-principles prediction forced by construction. I also flag, as non-circular but material, that the variance threshold tau is never given a value and that Section 1 ('Only when the variance exceeds a predefined threshold does the decoder continue expanding future decoding trajectories') states the opposite gate direction from Section 4.2 and Algorithm 1 ('if Var({s_j}) < tau then' rollout continues). This is a reproducibility/correctness problem, not a circularity problem, because it does not make the reported accuracy ordering equivalent to the method's inputs. The absence of fixed-depth rollout controls similarly weakens attribution but does not constitute circular reasoning. Therefore the derivation is not circular; the score is 0.

Axiom & Free-Parameter Ledger

10 free parameters · 7 axioms · 0 invented entities

The paper introduces no new physical or model entities; its inventions are algorithmic constructs (the variance gate, the branch-expansion state machine, the accumulated rollout score). The honest measure of what is pulled from the prior literature is large: the entire scoring machinery, candidate construction, and triggering conditions are inherited from ETE, while the genuinely new adaptive control depends on an unreported threshold tau and a calibration rule fit to MATH500.

free parameters (10)
  • k (lookahead beam width) = 4
    Number of candidate trajectories evaluated in parallel; fixed per Section 4.3 without reporting the calibration evidence.
  • alpha (score weight) = 0.1
    Weight on the candidate's own confidence in Eq. (1); adopted from Fu et al. (2025) per Section 4.3.
  • beta (position bias) = 0.01
    Bias favoring later positions in candidate construction; fixed per Section 4.3.
  • Tmax (max rollout depth) = 2
    Maximum lookahead depth per round; fixed per Section 4.3, meaning 'multi-step' is at most one extra forward pass per round, with longer horizons coming from re-triggering.
  • c_info (medium-confidence target) = 0.2
    Target confidence for candidate selection, taken from Fu et al. (2025) following their empirical findings (Section 3.2).
  • C (confidence threshold) = range [0.5 + 0.05 log2 N, 0.5 + 0.10 log2 N], clipped to [0.5, 0.9]
    Calibration-guided: the log-scale constants 0.05 and 0.10 are fit on a calibration set; Figure 5 uses MATH500 (Section 4.3, Appendix B).
  • N (per-block unlocking budget) = powers of two in [2, 64)
    Varied to sweep the accuracy-efficiency curve; paired with C via the range rule (Section 5.1).
  • gamma (lookahead trigger threshold) = scanned in [0.1, 0.4]; optimized curves pick per-benchmark best
    No stable calibration trend found (Section 4.3); the 'optimized' setting selects gamma on the test benchmarks themselves (Section 5.1).
  • tau (variance threshold) = unreported (no value given anywhere)
    The gate for adaptive rollout continuation (Section 4.2; Algorithm 1 line 13). Section 4.3 claims the extra adaptive hyperparameters are fixed after calibration, but tau's value is never stated - a central uncalibrated scalar.
  • N_e (minimum frontier tokens to trigger lookahead) = set equal to N
    Reported to have little impact within a reasonable range; set to N by default (Section 4.3).
axioms (7)
  • ad hoc to paper Candidate-score variance over the k=4 beam is a valid early-stopping signal for rollout continuation.
    Core of AdaLook (Section 4.2): 'If the variance falls below a threshold tau ... the decoder proceeds with another rollout step. Once the variance exceeds tau ... the decoder commits.' No derivation, calibration evidence, or reported value for tau.
  • domain assumption Accumulated downstream confidence mass (Eq. 1) remains a valid trajectory-quality proxy over multiple rollout steps.
    Extends ETE's one-step score to R steps by summing committed-token confidences; assumes confidence mass is additive across steps (Section 4.2).
  • domain assumption The ETE triggering condition (average frontier confidence below gamma with enough masked tokens) is the right criterion for re-triggering lookahead at intermediate branches.
    Borrowed from Fu et al. (2025) (Section 3.2) and reused for branch expansion (Section 4.3, Case 2) without re-validation in the multi-step setting.
  • domain assumption The log-tail probability log Pr(c >= C) is approximately linear in C over [0.2, 0.9].
    Empirical observation on MATH500 (Figure 5) used in Appendix B to 'justify' the log-scale C-N rule; the derivation is a post-hoc fit to the observed trend.
  • domain assumption Per-step token commitments are approximately independent, so the expected fraction committed after N rounds is 1 - (1 - p(C))^N.
    Appendix B; ignores within-sequence token correlation, which is known to be strong in language.
  • domain assumption Raw model argmax probabilities are a reliable confidence signal for triggering, scoring, and gating.
    Every mechanism (triggering, scoring, variance gate) is a function of raw model probabilities (Section 3.2); miscalibration would shift all thresholds.
  • domain assumption Hyperparameters calibrated on one small set transfer across benchmarks (MMLU, GSM8K, BBH) and backbones (LLaDA-8B, Dream-7B).
    Section 4.3 claims trends are 'generally consistent', but Figure 5 is evaluated with MATH500 only; transfer is asserted, not demonstrated per dataset.

pith-pipeline@v1.3.0-alltime-deepseek · 10754 in / 20489 out tokens · 184339 ms · 2026-08-01T22:40:35.441384+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Adaptive Multi-Step Lookahead Decoding for Diffusion Language Models." pith.science (2026). https://pith.science/paper/2D3NRF7S

@misc{pith2026260715655,
  author       = {Pith},
  title        = {Pith review of: Adaptive Multi-Step Lookahead Decoding for Diffusion Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2D3NRF7S}},
  note         = {Machine review of arXiv:2607.15655}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Masked diffusion language models (DLMs) enable parallel text generation by iteratively refining masked tokens, offering a promising alternative to autoregressive decoding. Recent lookahead-based decoding methods improve the accuracy--efficiency trade-off by exploring future decoding states before committing token updates. However, existing approaches mainly rely on shallow one-step lookahead, which optimizes immediate information gain but can be suboptimal for longer-horizon decoding trajectories. Meanwhile, we find that a naive extension for deeper lookahead is also ineffective, as fixed-depth rollout introduces additional computation and cannot adapt to heterogeneous intermediate decoding states. Thus, in this work, we propose AdaLook, an adaptive lookahead framework for DLM decoding. AdaLook dynamically determines whether to continue rollout based on candidate-score variance and further enables branch expansion when intermediate rollout states require additional exploration. This design avoids unnecessary deep rollout while allowing the decoder to re-trigger lookahead from informative intermediate states. Experiments on various benchmarks and models demonstrate that AdaLook achieves a better accuracy--decoding steps trade-off than existing one-step lookahead decoding methods.

Figures

Figures reproduced from arXiv: 2607.15655 by Charu C. Aggarwal, Hang Li, Hui Liu, Lantao Mei, Wei Deng, Yingqian Cui, Yue Xing.

Figure 1
Figure 1. Figure 1: Accuracy vs Decoding Steps (evaluated on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the adaptive multi-step lookahead mechanism (when the maximum lookahead depth is 2). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy vs decoding steps across different benchmarks [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance of different methods in HumanEval end-to-end generation time, including model for￾ward passes, score computation, candidate selec￾tion, branch evaluation, and other decoding over￾head, and divide it by the average number of for￾ward steps. Therefore, the reported value should be interpreted as the average time corresponding to one decoding step reported in [PITH_FULL_IMAGE:figures/full_fig_p00… view at source ↗
Figure 5
Figure 5. Figure 5: Relationship between log(p(c)) and C (evalu￾ated with MATH500). We provide an additional justification for why the preferred confidence threshold C increases ap￾proximately logarithmically with the per-block un￾locking budget N. For a fixed confidence threshold C, the expected fraction of tokens that can be committed in one refinement step is p(C) = Pr(c ≥ C). After N rounds the expected fraction of tokens… view at source ↗
Figure 6
Figure 6. Figure 6: Accuracy vs decoding steps across different benchmarks with DREAM-v0-Instruct-7B [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗

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

25 extracted references · 12 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems , volume=

    Simple and effective masked diffusion language models , author=. Advances in Neural Information Processing Systems , volume=

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    Large language diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    arXiv preprint arXiv:2511.05563 , year=

    Lookahead unmasking elicits accurate decoding in diffusion language models , author=. arXiv preprint arXiv:2511.05563 , year=

    arXiv

  4. [4]

    arXiv preprint arXiv:2505.16990 , year=

    Dimple: Discrete diffusion multimodal large language model with parallel decoding , author=. arXiv preprint arXiv:2505.16990 , year=

    Pith/arXiv arXiv

  5. [5]

    Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    Challenging big-bench tasks and whether chain-of-thought can solve them , author=. Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    2023

  6. [6]

    arXiv preprint arXiv:2310.16834 , year=

    Discrete diffusion modeling by estimating the ratios of the data distribution , author=. arXiv preprint arXiv:2310.16834 , year=

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2009.09761 , year=

    Diffwave: A versatile diffusion model for audio synthesis , author=. arXiv preprint arXiv:2009.09761 , year=

    Pith/arXiv arXiv 2009

  8. [8]

    International Conference on Learning Representations , volume=

    Your absorbing discrete diffusion secretly models the conditional distributions of clean data , author=. International Conference on Learning Representations , volume=

  9. [9]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    Diffusionbert: Improving generative masked language models with diffusion models , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

  10. [10]

    0: Scaling up diffusion language models to 100b , author=

    Llada2. 0: Scaling up diffusion language models to 100b , author=. arXiv preprint arXiv:2512.15745 , year=

    Pith/arXiv arXiv

  11. [11]

    Advances in neural information processing systems , volume=

    Simplified and generalized masked diffusion for discrete data , author=. Advances in neural information processing systems , volume=

  12. [12]

    Advances in neural information processing systems , volume=

    Structured denoising diffusion models in discrete state-spaces , author=. Advances in neural information processing systems , volume=

  13. [13]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  14. [14]

    arXiv preprint arXiv:2011.13456 , year=

    Score-based generative modeling through stochastic differential equations , author=. arXiv preprint arXiv:2011.13456 , year=

    Pith/arXiv arXiv 2011

  15. [15]

    arXiv preprint arXiv:2511.21103 , year=

    From bits to rounds: Parallel decoding with exploration for diffusion language models , author=. arXiv preprint arXiv:2511.21103 , year=

    arXiv

  16. [16]

    arXiv e-prints , pages=

    Accelerating diffusion large language models with slowfast: The three golden principles , author=. arXiv e-prints , pages=

  17. [17]

    arXiv preprint arXiv:2505.22618 , year=

    Fast-dllm: Training-free acceleration of diffusion llm by enabling kv cache and parallel decoding , author=. arXiv preprint arXiv:2505.22618 , year=

    Pith/arXiv arXiv

  18. [18]

    arXiv preprint arXiv:2110.14168 , year=

    Training verifiers to solve math word problems , author=. arXiv preprint arXiv:2110.14168 , year=

    Pith/arXiv arXiv

  19. [19]

    arXiv preprint arXiv:2009.03300 , year=

    Measuring massive multitask language understanding , author=. arXiv preprint arXiv:2009.03300 , year=

    Pith/arXiv arXiv 2009

  20. [20]

    arXiv preprint arXiv:2107.03374 , year=

    Evaluating large language models trained on code , author=. arXiv preprint arXiv:2107.03374 , year=

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2103.03874 , year=

    Measuring mathematical problem solving with the math dataset , author=. arXiv preprint arXiv:2103.03874 , year=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2506.06295 , year=

    dllm-cache: Accelerating diffusion large language models with adaptive caching , author=. arXiv preprint arXiv:2506.06295 , year=

    Pith/arXiv arXiv

  23. [23]

    Advances in Neural Information Processing Systems , volume=

    Accelerated sampling from masked diffusion models via entropy bounded unmasking , author=. Advances in Neural Information Processing Systems , volume=

  24. [24]

    arXiv preprint arXiv:2508.13021 , year=

    Pc-sampler: Position-aware calibration of decoding bias in masked diffusion models , author=. arXiv preprint arXiv:2508.13021 , year=

    arXiv

  25. [25]

    arXiv preprint arXiv:2508.15487 , year=

    Dream 7b: Diffusion large language models , author=. arXiv preprint arXiv:2508.15487 , year=

    Pith/arXiv arXiv