FOCUS: New Inference System Improves Diffusion Language Model Efficiency by Up to 3.5x
Researchers have developed FOCUS, an inference system that addresses computational inefficiencies in Diffusion Large Language Models (DLLMs) by dynamically focusing computation on decodable tokens. The system works by identifying which tokens can actually be decoded at each diffusion step and evicting non-decodable ones, reducing wasted computation. This approach achieves up to 3.52x throughput improvement in large-batch settings while maintaining or improving generation quality, potentially making DLLMs more practical alternatives to traditional auto-regressive language models.
Researchers have developed FOCUS, an inference optimization system designed to address a fundamental inefficiency in Diffusion Large Language Models (DLLMs). The key problem identified is that while DLLM decoding parallelizes computation across token blocks, only a small fraction of tokens are actually decodable at each diffusion step, causing most computational resources to be wasted on non-decodable tokens. The researchers discovered a strong correlation between attention-derived token importance scores and token-wise decoding probability, which informed their solution. FOCUS dynamically focuses computation on decodable tokens while evicting non-decodable ones during inference, effectively increasing the batch size and alleviating compute bottlenecks. Empirical evaluations demonstrate throughput improvements of up to 3.52x compared to the production-grade LMDeploy engine in large-batch settings, while preserving or improving generation quality across multiple benchmarks. The work is accepted as a camera-ready version for ICML 2026.
What's missing
The paper does not discuss potential limitations of the approach, such as scenarios where it may underperform, computational overhead of the dynamic focusing mechanism itself, or how performance scales in small-batch settings where the inefficiency may be less pronounced.
What different sources said
- arXiv cs.LGCenter
FOCUS: DLLMs Know How to Tame Their Compute Bound
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.