A newly revised arXiv preprint analyzes how quantization affects large language models deployed for machine translation in server settings. The work weighs reduced memory use and faster inference against the quality loss that compression can introduce. It aims to help practitioners pick quantization settings that balance latency and translation accuracy.
A revised arXiv paper examines tensorization, a method that reshapes a network's dense weight matrices into higher-order tensors and approximates them with low-rank tensor network decompositions. The authors argue the approach remains underused despite promising results as a model compression technique, and they highlight its potential for making networks easier to interpret. The submission appears as a replacement cross-list across arXiv's AI and machine learning categories.
A new arXiv preprint introduces a knowledge distillation method intended to help lightweight student models retain correlation knowledge between categories. The approach, called adaptive reciprocal knowledge distillation, targets the common problem that a large gap in size between teacher and student models weakens knowledge transfer. The work falls within ongoing research on model compression and efficient training.
A new arXiv paper introduces UniRank, a method for deciding how much rank to assign to each weight matrix when compressing large language models via low-rank decomposition. The authors argue that uniform or hand-tuned allocation schemes overlook differences in module importance. The work is released as a revised submission on arXiv.
Researchers propose SparseTalk, a method that compresses the semantic features stored in 3D Gaussian language fields used for 3D visual question answering. Dense scene representations can require tens of thousands of embeddings, driving up storage and memory costs. The approach aims to cut that redundancy while preserving performance on 3D VQA tasks.
A new arXiv paper argues that the action-generation backbones in Vision-Language-Action models are far larger than the task requires, since robot actions carry much less information than image pixels. The authors introduce a freeze-share-shrink strategy to cut parameters in diffusion and flow-matching policies while preserving performance. The work targets more efficient manipulation models for robotics.
A new arXiv preprint argues that AI deployment performance depends on how compression methods, compiler transformations, and serving policies interact, rather than on model architecture alone. It notes that existing benchmarks often report latency and throughput under conditions that cannot be directly compared, which limits practical conclusions. The work appears to be a cross-listed submission surveying the inference deployment stack.
A new arXiv paper tackles the mismatch between high-dimensional text embeddings and the limited input capacity of near-term quantum circuits. The authors experiment with compressing semantic representations so they can feed into hybrid quantum-classical classifiers for NLP tasks. The work offers an empirical look at how much representation reduction is viable for quantum machine learning applications.
A new paper examines how to shrink Mixture-of-Experts language models by pruning experts, targeting models specialized for particular domains. The authors find that earlier compression techniques break down badly when the trimmed models are tested on general-purpose benchmarks. The work aims to make these large, sparsely activated models easier to deploy without losing performance.
A new arXiv paper introduces LightMedSeg-ISLES, a 1.26-million-parameter pipeline for segmenting stroke lesions in T1-weighted MRI scans. The authors position it as a lighter alternative to large networks and ensembles like nnU-Net, citing an 81-fold reduction in parameters. The smaller footprint is intended to ease storage and inference demands for clinical deployment.