TL;DR
DMax enables aggressive parallel decoding in diffusion language models through On-Policy Uniform Training and Soft Parallel Decoding, addressing error accumulation. It improves tokens per forward from 2.8 to 6.2 while preserving accuracy, achieving over 1000 tokens/second throughput.
35
已证实
0
证据不足
0
无法验证
N/A
可复现性
置信度
86%

核心问题

How can diffusion language models achieve aggressive parallel decoding without suffering from error accumulation that degrades accuracy?

核心方法

{'approach': "The paper proposes DMax with two components: On-Policy Uniform Training (OPUT) extends pretrained masked diffusion models into self-corrective uniform diffusion models by sampling noisy sequences on-policy from the model's predictive distribution. Soft Parallel Decoding (SPD) represents intermediate decoding states as hybrid soft embeddings interpolated between predicted token and mask embeddings based on prediction confidence.", 'key_components': ['MDLMs train models to recover clean tokens from masked positions through a discrete denoising process.', 'UDLMs train models to denoise from arbitrary vocabulary tokens, enabling self-correction at all positions.', 'MDLMs suffer from error accumulation because decoded tokens become fixed context that cannot be revised.', "The proposed approach combines stable masked initialization with UDLM's self-revising capability."], 'section_ids': ['sec_5', 'sec_6']}

论点验证

已证实 (85%) we propose DMax, a novel paradigm that reformulates the binary maskto-token decoding process into a self-revising transformation in the embedding space
The paper provides detailed description of DMax's architecture (p_9, p_33-p_50), mathematical formulation of hybrid embeddings (Eqs. 7-10), and experimental validation demonstrating the self-revising capability through iterative refinement with soft
已证实 (85%) Central to our approach is On-Policy Uniform Training (OPUT), a training recipe that efficiently extends a pretrained masked diffusion language model into a self-corrective uniform diffusion language model while preserving its original mask denoising capability
OPUT is described in detail (p_23-p_30) with mathematical formulation (Eqs. 4-6), training procedure, and experimental validation. The dual supervision on both masked and predicted sequences demonstrates preservation of mask denoising while adding se
已证实 (85%) we further present Soft Parallel Decoding (SPD) for inference. Instead of treating decoded tokens as discrete and irrevocable commitments, SPD represents each intermediate decoding state as a hybrid soft embedding, formed by interpolating between the predicted token embedding and the mask embedding according to the model's prediction confidence
SPD is fully specified with mathematical formulation (Eqs. 7-10 in p_37-p_40), decoding algorithm (p_34-p_48), and experimental validation. The hybrid embedding interpolation between predicted token and mask embedding based on confidence is clearly d
已证实 (75%) On the mathematical reasoning benchmark GSM8K, our method increases tokens per forward (TPF) from 2.04 to 5.48 with only minimal accuracy degradation relative to the original model
Specific TPF numbers (2.04 to 5.48) are provided in p_10 and referenced to Table 1. The claim is directly stated with quantitative evidence. Minor gap: 'minimal accuracy degradation' is qualitative without specific accuracy numbers for this compariso
已证实 (75%) On the code generation benchmark MBPP, it improves TPF from 2.71 to 5.86 while maintaining comparable performance
Specific TPF numbers (2.71 to 5.86) are provided in p_10 for MBPP benchmark. The claim is directly stated with quantitative evidence. Similar to claim 4, 'comparable performance' is qualitative but the main TPF improvement is quantified.
已证实 (80%) we propose to unify the strengths of both paradigms. Specifically, we retain a fully masked sequence as the initialization of UDLM decoding to preserve stability, while continuing to re-predict all tokens that have been decoded from [MASK] at every subsequent step
The unification approach is clearly described in p_18, explaining how masked initialization preserves stability while token-to-token denoising enables self-correction. The approach is validated through experiments comparing to both MDLM and UDLM base
已证实 (85%) we propose On-Policy Uniform Training (OPUT), a simple yet effective method for equipping MDLMs with self-corrective denoising capability. The core idea is to construct training inputs using noisy sequences sampled on-policy from the model's own predictive distribution, rather than from a uniform vocabulary distribution, thereby bridging the train-inference gap
OPUT's core mechanism is clearly specified in p_23-p_30 with mathematical formulation. The on-policy sampling from model's own distribution (Eq. 4) and dual supervision (Eq. 6) are well-described. Experimental validation shows OPUT outperforms conven
已证实 (90%) At each training iteration, we first sample a corruption level t ∼ Uniform(t l , t h ), where t l and t h denote the lower and upper bounds of the noise level, respectively. Given a clean sequence x 0 ∼ D, we construct a masked noisy sequence x (m) t by independently replacing each token with [MASK] with probability t
This is a clearly specified design choice in p_24 with mathematical notation. The corruption level sampling and masked sequence construction procedure are fully described.
已证实 (90%) x (p) t is sampled using the current model parameters at each iteration, making this a strictly on-policy rollout process
The on-policy nature is explicitly stated in p_26: 'x(p)_t is sampled using the current model parameters at each iteration, making this a strictly on-policy rollout process.' This is a clear specification of the training procedure.
已证实 (90%) we perform two forward passes, using the masked noisy sequence x (m) t and the predicted noisy sequence x (p) t as inputs, respectively
The two forward passes are specified in p_27 with clear description of using masked noisy sequence and predicted noisy sequence as inputs respectively.
已证实 (90%) we supervise both outputs against the original clean sequence x 0 using cross-entropy loss over all token positions, regardless of whether a position is masked
The loss function is specified in p_28-p_29 with cross-entropy over all token positions regardless of masking status. The final training objective (Eq. 6) combines both losses.
已证实 (75%) On LLaDA-2.0-mini, our method improves GSM8K accuracy from 78% to 90% under confidence-threshold decoding with a threshold of 0.5, while also delivering faster decoding
Specific accuracy numbers (78% to 90%) are provided in p_30 for GSM8K with threshold 0.5. The claim includes both accuracy improvement and faster decoding. Self-reported result without error bars.
已证实 (85%) we propose soft parallel decoding. The central idea is to preserve predictive uncertainty from earlier iterations and explicitly propagate it to later refinement steps. Concretely, instead of treating intermediate decoding states as discrete tokens, we represent each decoded token as a soft embedding interpolated between the predicted token embedding and the mask embedding
Soft parallel decoding is fully described in p_33-p_50 with mathematical formulation of hybrid embeddings (Eqs. 7-10). The approach is validated through experiments and ablations showing its effectiveness.
已证实 (90%) At each decoding step, we use an aggressive confidence threshold τ dec to promote some mask positions into token positions. Specifically, we scan the masked region from left to right and promote only its longest contiguous prefix whose confidence exceeds τ dec
The confidence threshold promotion strategy is clearly specified in p_34 with algorithmic description of scanning masked region left-to-right and promoting longest contiguous prefix exceeding threshold.
已证实 (80%) This design keeps the masked region contiguous and prevents unreliable future tokens on the right from interfering with mask predictions on the left
The rationale for contiguous masked region is provided in p_34. Additionally, p_61 mentions 'Maintaining the non-masked region as a contiguous prefix further improves performance' in the ablation study, providing empirical support for this design cho
已证实 (90%) For each token position j ∈ T (t), we construct a hybrid embedding from the top-1 prediction at the previous step t -1 as the model input
The hybrid embedding construction from top-1 prediction is specified in p_37 with mathematical notation.
已证实 (85%) we renormalize the hybrid embedding so that its norm matches the probability-weighted sum of the component norms
The renormalization procedure is described in p_39-p_40 with rationale about avoiding norm collapse. The mathematical formulation is provided.
已证实 (90%) We regard a block as having converged to a stable state if either of the following conditions holds: (1) the top-1 predictions at all positions remain unchanged for two consecutive decoding steps, or (2) the confidence of every position in the block exceeds a high acceptance threshold τ acc
The convergence criteria are clearly specified in p_48 with two conditions: (1) unchanged predictions for two consecutive steps, or (2) all confidences exceeding τ_acc.
已证实 (75%) On OPUT-trained LLaDA-2.0-mini, under the highly aggressive setting of τ dec = 0, soft parallel decoding improves GSM8K accuracy from 68% to 90% while achieving a higher speedup
Specific accuracy numbers (68% to 90%) are provided in p_50 for GSM8K with τ_dec=0. This demonstrates SPD's effectiveness under highly aggressive parallel decoding. Self-reported without error bars.
已证实 (85%) soft parallel decoding must be used together with OPUT-trained models. OPUT trains the model to recover the correct target not only from masked inputs, but also from its own sampled predictions
The prerequisite relationship is explained in p_51 with rationale that OPUT trains consistent mapping from both mask and predicted embeddings. Empirical evidence in p_61 (Table 3) shows SPD causes collapse without OPUT.

... 共 35 个论点

可复现性评估

较低可复现性 (0%)

缺失的复现细节

局限性（作者自述）

本分析由 PDF 阅读助手 自动生成，仅供参考，不构成学术评审意见。验证结论和可复现性评估基于论文文本自动分析，可能存在偏差。原始论文请参阅 arXiv。

分析时间：2026-04-26T07:24:12+00:00 · 数据来源：Paper Collector