arXiv:2609.24150v1 Announce Type: new Abstract: Speculative decoding accelerates large language model (LLM) inference by using a lightweight draft model to generate multiple candidate tokens that are...
arXiv:2605. 13026v2 Announce Type: replace-cross Abstract: Masked diffusion models (MDMs) have emerged as a promising alternative to autoregressive models (ARMs) for language modeling.
arXiv:2608. 13925v1 Announce Type: cross Abstract: Diffusion large language models (dLLMs) accelerate language generation by predicting multiple masks in a single forward pass.
The paper introduces a new framework for adaptive parallel sampling of discrete vectors, where a deterministic policy reveals coordinates round‑by‑round based on previously observed values and samples the remaining coordinates from their exact conditional marginals. The authors prove an exact identity linking the forward Kullback‑Leibler divergence of any policy to the expected conditional total correlation accumulated during the sampling process, establishing conditional total correlation as the precise information cost of within‑round parallelism. Using this identity, they derive zero‑error schedules for finite‑order Markov chains, characterize the serial depth of Bernoulli walks, and demonstrate separations between different reveal orders, permutation strategies, and string structures, thereby revealing how conditional dependence governs parallelizability. whyItMatters":"The results provide a principled, information‑theoretic measure of parallel sampling efficiency that can guide the design of decoding rules for masked diffusion models and other generative systems."
The paper presents a theoretical analysis of score mismatch in discrete diffusion models, deriving a KL divergence upper bound that quantifies the bias introduced when the training and target distributions differ. It examines partial-observation constraints such as text infilling and sentence completion, showing that the Conditional SEDD sampler can incur a bias that grows with dimension, vocabulary size, and compatibility. To mitigate this, the authors propose a Randomized Hard Clamping (RHC) sampler for coordinatewise set constraints, demonstrating through experiments that it effectively reduces bias and improves sampling performance.
The paper introduces the Expected Decoding Rounds (EDR) objective, a theoretical framework that models speculative decoding as a Markov reward process and directly optimizes the number of decoding rounds. EDR weights local rejection costs by state occupancies, eliminating auxiliary hyperparameters and enabling unbiased stochastic optimization via temporal‑difference gradients. Applying EDR to fine‑tune state‑of‑the‑art draft models (DSpark and DFly) improves mean accepted length and outperforms prior training objectives across nine benchmarks in math reasoning, code generation, and chat.
arXiv:2606. 11552v1 Announce Type: cross Abstract: Large language models (LLMs) achieve remarkable performance across a wide range of tasks, but their autoregressive decoding process incurs substantial inference costs due to inherently sequential token generation.
The paper introduces Fork-dLLM, a hybrid sampling strategy for diffusion language models that combines parallel token generation with autoregressive ordering only at uncertain fallback steps, thereby avoiding the flexibility trap that reduces generation diversity. It also presents ForkGRPO, an extension that applies the GRPO objective selectively during post‑training rollouts, preserving policy‑likelihood ratios while cutting rollout and optimization costs. Experiments show Fork-dLLM achieves pass@k scaling comparable to autoregressive sampling but 2–3× more efficient, and ForkGRPO matches or surpasses AR‑based GRPO baselines at significantly lower training expense.
The paper introduces the Single-Increment Rejection Sampler (SIRS) to improve query efficiency in parallel masked diffusion models (MDMs). SIRS estimates the log acceptance probability using only one randomly sampled increment, then applies a Bernoulli factory to decide accept/reject outcomes. By Rao‑Blackwellizing each increment and bounding the KL‑divergence, SIRS reduces bias while maintaining parallel decoding performance.
arXiv:2609.37974v1 Announce Type: cross Abstract: Masked diffusion models (MDMs) generate text by unmasking several tokens per step, but they are trained and sampled under different conditions. The m...
arXiv:2605. 04215v3 Announce Type: replace-cross Abstract: Diffusion-based Large Language Models (D-LLMs) represent a promising frontier in generative AI, offering fully parallel token generation that can lead to significant throughput advantages and superior GPU utilization over the traditional autoregressive paradigm.
arXiv:2505.16990v3 Announce Type: replace Abstract: In this work, we propose Dimple, the first Discrete Diffusion Multimodal Large Language Model (DMLLM). We observe that training with a purely discr...