CORTEXA
← Browse
arxivcs.LGcs.AIcs.CL2026-07-02

Training Hybrid Block Diffusion Language Models with Partial Bidirectionality

Pranshu Chaturvedi, Parth Shroff, Tarun Suresh, Hangoo Kang, Kaiyue Wen

High-throughput long-context generation is one of the central challenges for large language models. Generation is typically memory-bandwidth-bound rather than compute-bound: each decoding step must stream the accumulated key/value (KV) cache from memory, so bandwidth demand grows with context length while only one token is emitted. Two parallel approaches have therefore emerged: reducing memory access with efficient attention variants and linear-time mixers such as Mamba, or increasing parallel computation by generating blocks of tokens at once. However, technical challenges arise when combining these two ideas. Earlier hybrid diffusion models such as DiffuMamba use bidirectional Mamba mixing, including a reverse-direction scan relative to causal generation. This reverse scan needs to scan the entire sequence, so its states are not prefix-only and cannot be precisely reused as a cache even when diffusion is performed block by block. We propose a BDLM Mamba--attention hybrid that addresses this challenge by restricting the reverse Mamba scan to the active denoising block, which enables exact caching across blocks. In an 87M-parameter DCLM sweep, BDLM Mamba-H achieves the best C4-en validation perplexity compared to BDLM attention and full-sequence baselines. At 350M parameters, it remains competitive with BDLM attention. For long-context inference, BDLM Mamba-H reaches 19.7x the throughput of full-sequence DiffuMamba-H at 65K tokens and 3.7x the throughput of BDLM attention at 262K, showing that Mamba hybrids are a potential long-context diffusion architecture.

View free PDFSource page

Related papers

arxivcs.LGcs.AIcs.CL2026-07-01

CausalMix: Data Mixture as Causal Inference for Language Model Training

Zinan Tang, Yukun Zhang, Shaomian Zheng, Zhuoshi Pan, Qizhi Pei, Dingnan Jin, et al.

In Large Language Model (LLM) training, data mixing plays a pivotal role in determining model performance. Recent methods optimize mixture weights via proxy models, but they rely on the assumption of static data distributions. As a result, when the underlying data pool shifts, th…

View free PDFSource page
arxivcs.CLcs.AIcs.LG2026-07-18

Trace-Based On-Policy Distillation for Masked Diffusion Language Models

Haolin Ren, Ziyang Huang, Chenhao Yuan, Jun Zhao, Kang Liu

Diffusion large language models (dLLMs) are a promising alternative to autoregressive generation. However, reasoning-oriented post-training for dLLMs remains challenging. Supervised fine-tuning (SFT) for dLLMs requires dense but often off-policy masked states, while reinforcement…

View free PDFSource page
arxivcs.LGcs.AIcs.CL2026-07-14

Accelerating Masked Diffusion Large Language Models: A Survey of Efficient Inference Techniques

Daehoon Gwak, Minhyung Lee, Junwoo Park, Jaegul Choo

Diffusion large language models (dLLMs) offer a theoretical advantage in parallel generation over standard autoregressive models. However, parallel generation alone does not guarantee practical speedups. Realizing this efficiency requires specialized inference mechanisms, such as…

View free PDFSource page
arxivcs.CLcs.AIcs.LG2026-07-16

Mask-Aware Policy Gradients for Diffusion Language Models

Haran Raajesh, Kulin Shah, Adam Klivans, Philipp Krähenbühl

Reinforcement learning has proven effective for improving reasoning in large language models, but extending it to Masked Diffusion Language Models (MDLMs) remains challenging due to the intractability of the log-likelihood estimation. Existing approaches approximate this log-like…

View free PDFSource page
arxivcs.CLcs.AIcs.LG2026-07-17

Induction in Both Directions: A Mechanistic Analysis of In-Context Learning in Masked Diffusion Language Models

Andy Catruna, Emilian Radoi

While the internal mechanisms of autoregressive (AR) transformers have been studied extensively, much less is known about diffusion language models (DLMs), an emerging alternative that generates text by iterative denoising. In this work, we study how DLMs implement induction, a m…

View free PDFSource page
arxivcs.LGcs.AIcs.CL2026-07-23

Training Large Language Models for Self-Explanation Faithfulness

Yeoktatt Cheah, María Pérez-Ortiz, Noah Y. Siegel, Oana-Maria Camburu

We propose a Reinforcement Learning (RL) method to directly optimize the faithfulness of self-explanations - the extent to which a model's generated reasoning accurately reflects its internal decision-making process. While existing work focuses on evaluating faithfulness or using…

View free PDFSource page