Prefilling-dLLM partitions prefixes into chunks, caches KV representations, and applies sparse top-K selection during decoding to cut dLLM inference complexity to quadratic in decode length only.
hub
Lavida: A large diffusion model for vision-language understanding.Advances in neural information process- ing systems, 2025b
23 Pith papers cite this work. Polarity classification is still indexing.
23
Pith papers citing it
abstract
Diffusion Language Models (DLMs) are rapidly emerging as a powerful and promising alternative to the dominant autoregressive (AR) paradigm. By generating tokens in parallel through an iterative denoising process, DLMs possess inherent advantages in reducing inference latency and capturing bidirectional context, thereby enabling fine-grained control over the generation process. While achieving a several-fold speed-up, recent advancements have allowed DLMs to show performance comparable to their autoregressive counterparts, making them a compelling choice for various natural language processing tasks. In this survey, we provide a holistic overview of the current DLM landscape. We trace its evolution and relationship with other paradigms, such as autoregressive and masked language models, and cover both foundational principles and state-of-the-art models. Our work offers an up-to-date, comprehensive taxonomy and an in-depth analysis of current techniques, from pre-training strategies to advanced post-training methods. Another contribution of this survey is a thorough review of DLM inference strategies and optimizations, including improvements in decoding parallelism, caching mechanisms, and generation quality. We also highlight the latest approaches to multimodal extensions of DLMs and delineate their applications across various practical scenarios. Furthermore, our discussion addresses the limitations and challenges of DLMs, including efficiency, long-sequence handling, and infrastructure requirements, while outlining future research directions to sustain progress in this rapidly evolving field. Project GitHub is available at https://github.com/VILA-Lab/Awesome-DLMs.
hub tools
citation-role summary
background 4
citation-polarity summary
roles
background 4polarities
background 4representative citing papers
The paper introduces Uni-E, a unified energy for DLMs that accounts for model capacity, dependency and invariance, can be computed exactly, and corrects distribution shifts from dependency and invariance.
Continuous language diffusion works by entering high-margin decoder basins where frozen T5 embeddings recover 93-96% of native decisions and linear readouts reach 97.9% agreement, implying models should be evaluated as representation-decoder systems.
Infilling extraction on diffusion language models extracts up to three times more verbatim sequences than prefix methods and achieves higher recall on redacted emails than autoregressive models.
Discrete diffusion models learn data support before frequencies because the exact reverse process decomposes edits into a dominant validity scale and a finer probability coefficient.
TAD improves the accuracy-parallelism trade-off in diffusion LLMs via temporal-aware self-distillation that applies hard labels to soon-to-be-decoded tokens and soft supervision to future tokens.
DepCap accelerates diffusion LM inference up to 5.63x by using last-block influence for adaptive block boundaries and conflict-free token selection for parallel decoding, with negligible quality loss.
Diffusion language models and a CTC-USDM joint decoder improve ASR accuracy over standard approaches.
DMax uses On-Policy Uniform Training and Soft Parallel Decoding to enable aggressive parallelism in dLLMs, raising TPF on GSM8K from 2.04 to 5.47 and on MBPP from 2.71 to 5.86 while preserving accuracy.
DiLaServe improves SLO attainment for diffusion language models by up to 56.6 percentage points and reduces latency by up to 46% with less than 1% accuracy drop via deadline-aware scheduling and dynamic reconfiguration.
Training-time augmentations in token noise, permutation, and offset categories reduce overfitting and improve minimum validation loss in multi-epoch autoregressive pretraining on fixed corpora.
AGDO improves dLLM reasoning performance by determining denoising order and emphasizing tokens based on attention-derived dependencies rather than random masking.
DFlare replaces DFlash's shared fused representation with per-draft-layer attention to distinct target-layer combinations, enabling deeper drafts and 2.4M training samples for 5-11% higher speedups than DFlash on Qwen3 and GPT-OSS models.
dMoE aggregates token expert distributions to block level in dLLMs, cutting unique experts from 69.5 to 14.6, memory by 76-80%, and latency by 1.14-1.66x while retaining 99.11% performance.
Diagnoses mask prior drift and positional attention collapse in LDVLMs and introduces two plug-and-play decoding interventions that raise long-form generation quality without retraining.
ELF applies continuous-time flow matching in embedding space for language generation and reports outperforming prior discrete and continuous diffusion language models with fewer steps.
TrajDLM applies block diffusion language models to discrete road-segment sequences with topology constraints to generate realistic trajectories up to 2.8 times faster than prior methods while supporting zero-shot transfer.
Saber improves both speed and accuracy of diffusion language models on code generation by dynamically adjusting unmasking steps and reverting low-confidence tokens via backtracking.
FS-DFM enables 1024-token generation at perplexity parity with 1024-step baselines using only 8 steps via explicit step-budget training, reliable updates, and teacher guidance.
WaveFilter applies wavelet decomposition to filter critical tokens for sparse KV caching, improving long-context performance of diffusion LLMs as a plug-and-play addition to existing methods.
Predict-then-Diffuse predicts response length for diffusion LLMs before inference, cutting FLOPs with a data-driven safety buffer while preserving output quality.
MarCos modifies transformers to perform continuous multi-step reasoning by mapping thought-level continuous states directly to next-thought distributions, achieving substantial wall-clock speedups on math problems.
Empirical test shows top-1 argmax concentration has zero precision as collapse warning in DLM LoRA training due to pre-equilibrium saturation while max gradient norm provides usable but family-specific detection on short-horizon runs.
citing papers explorer
-
Prefilling-dLLM: Predictive Prefilling for Long-Context Inference in Diffusion Language Models
Prefilling-dLLM partitions prefixes into chunks, caches KV representations, and applies sparse top-K selection during decoding to cut dLLM inference complexity to quadratic in decode length only.
-
Unified Energy for Invariant and Independent Decoding in Diffusion Language Models
The paper introduces Uni-E, a unified energy for DLMs that accounts for model capacity, dependency and invariance, can be computed exactly, and corrects distribution shifts from dependency and invariance.
-
Continuous Language Diffusion as a Decoder-Interface Problem
Continuous language diffusion works by entering high-margin decoder basins where frozen T5 embeddings recover 93-96% of native decisions and linear readouts reach 97.9% agreement, implying models should be evaluated as representation-decoder systems.
-
Extracting Training Data from Diffusion Language Models via Infilling
Infilling extraction on diffusion language models extracts up to three times more verbatim sequences than prefix methods and achieves higher recall on redacted emails than autoregressive models.
-
Support Before Frequency in Discrete Diffusion
Discrete diffusion models learn data support before frequencies because the exact reverse process decomposes edits into a dominant validity scale and a finer probability coefficient.
-
TAD: Temporal-Aware Trajectory Self-Distillation for Fast and Accurate Diffusion LLM
TAD improves the accuracy-parallelism trade-off in diffusion LLMs via temporal-aware self-distillation that applies hard labels to soon-to-be-decoded tokens and soft supervision to future tokens.
-
DepCap: Adaptive Block-Wise Parallel Decoding for Efficient Diffusion LM Inference
DepCap accelerates diffusion LM inference up to 5.63x by using last-block influence for adaptive block boundaries and conflict-free token selection for parallel decoding, with negligible quality loss.
-
Diffusion Language Models for Speech Recognition
Diffusion language models and a CTC-USDM joint decoder improve ASR accuracy over standard approaches.
-
DMax: Aggressive Parallel Decoding for dLLMs
DMax uses On-Policy Uniform Training and Soft Parallel Decoding to enable aggressive parallelism in dLLMs, raising TPF on GSM8K from 2.04 to 5.47 and on MBPP from 2.71 to 5.86 while preserving accuracy.
-
DiLaServe: High SLO Attainment Serving for Diffusion Language Models
DiLaServe improves SLO attainment for diffusion language models by up to 56.6 percentage points and reduces latency by up to 46% with less than 1% accuracy drop via deadline-aware scheduling and dynamic reconfiguration.
-
Demystifying Training-Time Augmentation for Data-Constrained Language Model Pretraining
Training-time augmentations in token noise, permutation, and offset categories reduce overfitting and improve minimum validation loss in multi-epoch autoregressive pretraining on fixed corpora.
-
Beyond Fully Random Masking: Attention-Guided Denoising and Optimization for Diffusion Language Models
AGDO improves dLLM reasoning performance by determining denoising order and emphasizing tokens based on attention-derived dependencies rather than random masking.
-
DFlare: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding
DFlare replaces DFlash's shared fused representation with per-draft-layer attention to distinct target-layer combinations, enabling deeper drafts and 2.4M training samples for 5-11% higher speedups than DFlash on Qwen3 and GPT-OSS models.
-
dMoE: dLLMs with Learnable Block Experts
dMoE aggregates token expert distributions to block level in dLLMs, cutting unique experts from 69.5 to 14.6, memory by 76-80%, and latency by 1.14-1.66x while retaining 99.11% performance.
-
Mitigating Mask Prior Drift and Positional Attention Collapse in Large Diffusion Vision-Language Models
Diagnoses mask prior drift and positional attention collapse in LDVLMs and introduces two plug-and-play decoding interventions that raise long-form generation quality without retraining.
-
ELF: Embedded Language Flows
ELF applies continuous-time flow matching in embedding space for language generation and reports outperforming prior discrete and continuous diffusion language models with fewer steps.
-
TrajDLM: Topology-Aware Block Diffusion Language Model for Trajectory Generation
TrajDLM applies block diffusion language models to discrete road-segment sequences with topology constraints to generate realistic trajectories up to 2.8 times faster than prior methods while supporting zero-shot transfer.
-
Saber: An Efficient Sampling with Adaptive Acceleration and Backtracking Enhanced Remasking for Diffusion Language Model
Saber improves both speed and accuracy of diffusion language models on code generation by dynamically adjusting unmasking steps and reverting low-confidence tokens via backtracking.
-
FS-DFM: Fast and Accurate Long Text Generation with Few-Step Diffusion Language Models
FS-DFM enables 1024-token generation at perplexity parity with 1024-step baselines using only 8 steps via explicit step-budget training, reliable updates, and teacher guidance.
-
WaveFilter: Enhancing the Long-Context Capability of Diffusion LLMs via Wavelet-Guided KV Cache Filtering
WaveFilter applies wavelet decomposition to filter critical tokens for sparse KV caching, improving long-context performance of diffusion LLMs as a plug-and-play addition to existing methods.
-
Predict-then-Diffuse: Adaptive Response Length for Compute-Budgeted Inference in Diffusion LLMs
Predict-then-Diffuse predicts response length for diffusion LLMs before inference, cutting FLOPs with a data-driven safety buffer while preserving output quality.
-
Deep Thinking by Markov Chain of Continuous Thoughts
MarCos modifies transformers to perform continuous multi-step reasoning by mapping thought-level continuous states directly to next-thought distributions, achieving substantial wall-clock speedups on math problems.
-
When Top-1 Fails: Calibrating LoRA Monitors for Masked Diffusion LMs
Empirical test shows top-1 argmax concentration has zero precision as collapse warning in DLM LoRA training due to pre-equilibrium saturation while max gradient norm provides usable but family-specific detection on short-horizon runs.