Researchers Decompose Linear Layers in Neural Networks Using Geometric Primitives
A study accepted at NeurIPS 2025 demonstrates that linear layers in neural networks can be decomposed into compositions of geometric objects called bivectors using Clifford algebra, requiring only O(log² d) parameters instead of the O(d²) needed by dense matrices. The work introduces a differentiable algorithm that breaks linear transformations into products of rotors — mathematical objects encoding rotations in oriented planes. This offers a more parameter-efficient and geometrically interpretable alternative for key, query, and value projections in large language model attention layers, matching the performance of established baselines.
Researchers have developed a framework for expressing the linear layers found in large neural networks as compositions of bivectors — geometric primitives from Clifford algebra that encode oriented planes. Their differentiable decomposition algorithm rewrites linear transformations as products of rotors, reducing the parameter count from O(d²) for standard dense matrices to O(log² d), a substantial compression. When applied to the key, query, and value projection matrices in transformer attention mechanisms, the rotor-based layers perform comparably to strong existing approaches such as block-Hadamard and low-rank approximations. The work is motivated by the observation that large models appear to exhibit modular, compositional behavior at a low level, yet the fundamental building blocks of that structure are not well understood. By grounding the analysis in Clifford algebra, the authors provide an algebraic and geometric lens through which the internal structure of deep models can be studied. The paper, spanning 35 pages with 11 tables and 6 figures, has been accepted to Advances in Neural Information Processing Systems 38 (NeurIPS 2025).
What's missing
The abstract does not report empirical results (e.g., specific accuracy or perplexity numbers) for the rotor-based layers versus baselines, nor does it specify which LLMs or benchmark tasks were used for evaluation. It is also unclear whether the O(log² d) parameterization introduces any training instability or convergence trade-offs relative to dense layers. The scope of 'matching performance' — whether this holds across model scales and tasks — is not addressed in the available abstract.
What different sources said
- arXiv cs.LGCenter
Composing Linear Layers from Irreducibles
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.