FOCUS: New Inference System Improves Diffusion Language Model Efficiency by Up to 3.5x
Researchers have proposed FOCUS, an inference system for Diffusion Large Language Models (DLLMs) that dynamically concentrates computation on tokens likely to be decoded at each diffusion step, discarding the rest. The work identifies that in standard DLLM decoding, most compute is wasted on tokens that are not yet ready to be decoded, and leverages a correlation between attention-derived token importance and decodability to address this. Accepted at ICML 2026, the system achieves up to 3.52× throughput improvement over the production engine LMDeploy without sacrificing generation quality.
Diffusion Large Language Models present an alternative to the dominant auto-regressive paradigm for text generation, but their practical deployment has been hampered by high decoding costs. The FOCUS paper identifies a core inefficiency: DLLM decoding parallelizes computation across entire token blocks, yet at any given diffusion step only a small fraction of tokens is actually ready to be decoded, meaning the majority of compute is effectively wasted. The authors observe a strong empirical correlation between attention-derived token importance scores and per-token decoding probability, which they exploit to dynamically route computation toward decodable tokens and evict non-decodable ones on the fly. This selective focus increases the effective batch size processed per step, alleviating the compute-bound bottleneck and enabling scalable throughput scaling. Empirical evaluations show up to 3.52× throughput gains over LMDeploy in large-batch settings, with generation quality preserved or improved across multiple benchmarks. The work has been accepted as a camera-ready paper at ICML 2026.
What's missing
The paper does not detail the specific benchmarks used to evaluate generation quality, the range of model sizes tested, or how FOCUS performs in low-batch or latency-sensitive (single-request) settings. It is also unclear how the approach generalizes across different DLLM architectures beyond those evaluated.
What different sources said
- arXiv cs.LGCenter
FOCUS: DLLMs Know How to Tame Their Compute Bound
Related
Gut Bacteria Enzyme Found to Break Down Heat-Processed Food Compounds, Producing Novel Biogenic Amines
Researchers have discovered that an enzyme in common gut bacteria can degrade N-epsilon-carboxymethyllysine (CML), a compound formed during thermal food processing, producing previously unknown biogenic amines. The enzyme, ornithine decarboxylase SpeC from enterobacteria, acts on CML and related modified lysine derivatives through a low-level 'underground' catalytic activity. This finding suggests a previously unrecognized communication axis between thermally processed dietary compounds and gut microbial physiology, with potential implications for host health.
Full-Length Gene Sequencing Reveals Two Distinct Bacterial Communities in Black-Legged Ticks Expanding Into Canada
Researchers used Oxford Nanopore full-length 16S rRNA gene sequencing to characterize the microbiome of Ixodes scapularis black-legged ticks collected in Nova Scotia, Canada, distinguishing between tick-adapted bacteria and environmentally acquired bacteria. The study comes as I. scapularis — the primary vector of Lyme disease — is rapidly expanding northward into Canada due to climate change. The findings suggest that environmentally derived bacteria in tick microbiomes are not mere contamination, which has implications for how tick microbiome data is collected and interpreted across surveillance studies.
Study Identifies Metabolic Link Between Cell Envelope Stress and Biofilm Formation in Bacteria
Researchers have discovered that the metabolite acetyl-CoA directly inhibits enzymes that degrade the bacterial signaling molecule c-di-GMP, connecting cell envelope biosynthesis stress to biofilm formation in Pseudomonas aeruginosa. The study found that sub-inhibitory concentrations of antibiotics targeting early peptidoglycan biosynthesis — but not other antibiotic classes — elevate c-di-GMP levels by reducing phosphodiesterase activity, with acetyl-CoA competing for the enzyme active site. Because the relevant enzyme domain is broadly conserved across bacterial species, this checkpoint mechanism may be widespread and could have implications for understanding antibiotic-induced biofilm responses.