Computer Science > Computation and Language
arXiv:2607.24841 (cs)
Submitted on 24 Jul 2026
Title: Neuromorphic Diffusion Language Models: Addressing Compute and Memory Bottlenecks via Sparsity and Block Denoising
Authors: Dengyu Wu, Clement Ruah, Jiechen Chen, Bipin Rajendran, Osvaldo Simeone
Abstract
Autoregressive (AR) large language models (LLMs) suffer from inherent inefficiency during inference because each generated token requires accessing the full set of model parameters, resulting in low operational intensity and high energy consumption. Masked diffusion language models (MDLMs) partially address this limitation in memory-bound settings by enabling multiple tokens to be generated per parameter access.
To further enhance inference efficiency on modern platforms with extensive on-chip memory, this work introduces neuromorphic MDLMs (N-MDLMs), which integrate block diffusion with spike-based neuromorphic computation. This combination jointly improves both throughput and energy efficiency. While block diffusion increases token throughput by producing multiple tokens per parameter access, spike-induced sparsity reduces effective parameter traffic and computational load by skipping inactive channels.
To analyze the synergistic effect of sparsity and diffusion, we develop a token-level, roofline-inspired model that captures the combined impact of block-parallel generation and spike sparsity on decoding efficiency. Experimental results on translation tasks demonstrate that, thanks to spike-induced sparsity, N-MDLMs achieve substantial improvements in energy efficiency and throughput—even in compute-bound platforms where MDLMs alone would fail to outperform AR-LLMs.
Additional Information
- Comments: Accepted for presentation at the 2026 IEEE Workshop on Signal Processing Systems (SiPS)
- Subjects: Computation and Language (cs.CL); Machine Learning (cs.LG); Signal Processing (eess.SP)
- Cite as: arXiv:2607.24841 [cs.CL]
- DOI: https://doi.org/10.48550/arXiv.2607.24841
Context (2026)
As of 2026, the landscape of large language model inference continues to be constrained by memory bandwidth and energy consumption, particularly in edge and real-time applications. This work proposes a neuromorphic approach that leverages spike-based computation and block denoising to overcome these bottlenecks, offering a promising direction for efficient LLM deployment on next-generation hardware platforms.
via ArXiv CL