Large Language Models (LLMs) achieve superior performance through Chainof-Thought (CoT) reasoning, but these token-level reasoning chains are computationally expensive and inefficient. In this paper, we introduce Compressed Latent Reasoning (CoLaR), a novel framework that dynamically compresses reasoning processes in latent space through a two-stage training approach. First, during supervised fine-tuning, CoLaR extends beyond next-token prediction by incorporating an auxiliary next compressed embedding prediction objective. This process merges embeddings of consecutive tokens using a compression factor crandomly sampled from a predefined range, and trains a specialized latent head to predict distributions of subsequent compressed embeddings. Second, we enhance CoLaR through reinforcement learning (RL) that leverages the latent head's non-deterministic nature to explore diverse reasoning paths and exploit more compact ones. This approach enables CoLaR to: i) perform reasoning at a dense latent level (i.e., silently), substantially reducing reasoning chain length, and ii) dynamically adjust reasoning speed at inference time by simply prompting the desired compression factor. Extensive experiments across four mathematical reasoning datasets demonstrate that CoLaR achieves 14.1% higher accuracy than latent-based baseline methods at comparable compression ratios, and reduces reasoning chain length by 53.3%with

Large Language Models (LLMs) achieve superior performance through Chain-of-Thought (CoT) reasoning, but these token-level reasoning chains are computationally expensive and inefficient. In this paper, we introduce Compressed Latent Reasoning (CoLaR), a novel framework that dynamically compresses reasoning processes in latent space through a two-stage training approach. First, during supervised fine-tuning, CoLaR extends beyond next-token prediction by incorporating an auxiliary next compressed embedding prediction objective. This process merges embeddings of consecutive tokens using a compression factor $c$ randomly sampled from a predefined range, and trains a specialized latent head to predict distributions of subsequent compressed embeddings. Second, we enhance CoLaR through reinforcement learning (RL) that leverages the latent head's non-deterministic nature to explore diverse reasoning paths and exploit more compact ones. This approach enables CoLaR to: i) **perform reasoning at a dense latent level** (i.e., silently), substantially reducing reasoning chain length, and ii) **dynamically adjust reasoning speed** at inference time by simply prompting the desired compression factor. Extensive experiments across four mathematical reasoning datasets demonstrate that CoLaR achieves 14.1% higher accuracy than latent-based baseline methods at comparable compression ratios, and reduces reasoning chain length by 53.3% with only 4.8% performance degradation compared to explicit CoT method. Moreover, when applied to more challenging mathematical reasoning tasks, our RL-enhanced CoLaR demonstrates performance gains of up to 5.4% while dramatically reducing latent reasoning chain length by 82.8%. The code and models will be released upon acceptance.

Neural Information Processing Systems http://nips.cc/

Diffusion-weighted MRI allows us to estimate structural brain connections in-vivo by measuring the diffusion of water molecules at different spatial directions.

Medical images are inherently high-resolution and contain locally varying structures crucial for diagnosis. Efficient compression must preserve diagnostic fidelity while minimizing redundancy. Low-rank matrix approximation (LoRMA) techniques have shown strong potential for image compression by capturing global correlations; however, they often fail to adapt to local structural variations across regions of interest. To address this, we introduce an adaptive LoRMA, which partitions a medical image into overlapping patches, groups structurally similar patches into clusters using k-means, and performs SVD within each cluster. We derive the overall compression factor accounting for patch overlap and analyze how patch size influences compression efficiency and computational cost. While applicable to any data with high local variation, we focus on medical imaging due to its pronounced local variability. We evaluate and compare our adaptive LoRMA against global SVD across four imaging modalities: MRI, ultrasound, CT scan, and chest X-ray. Results demonstrate that adaptive LoRMA effectively preserves structural integrity, edge details, and diagnostic relevance, measured by PSNR, SSIM, MSE, IoU, and EPI. Adaptive LoRMA minimizes block artifacts and residual errors, particularly in pathological regions, consistently outperforming global SVD in PSNR, SSIM, IoU, EPI, and achieving lower MSE. It prioritizes clinically salient regions while allowing aggressive compression in non-critical regions, optimizing storage efficiency. Although adaptive LoRMA requires higher processing time, its diagnostic fidelity justifies the overhead for high-compression applications.

Kolmogorov-Arnold Networks (KANs) replace scalar weights with per-edge vectors of basis coefficients, thereby boosting expressivity and accuracy but at the same time resulting in a multiplicative increase in parameters and memory. We propose MetaCluster, a framework that makes KANs highly compressible without sacrificing accuracy. Specifically, a lightweight meta-learner, trained jointly with the KAN, is used to map low-dimensional embedding to coefficient vectors, shaping them to lie on a low-dimensional manifold that is amenable to clustering. We then run K-means in coefficient space and replace per-edge vectors with shared centroids. Afterwards, the meta-learner can be discarded, and a brief fine-tuning of the centroid codebook recovers any residual accuracy loss. The resulting model stores only a small codebook and per-edge indices, exploiting the vector nature of KAN parameters to amortize storage across multiple coefficients. On MNIST, CIFAR-10, and CIFAR-100, across standard KANs and ConvKANs using multiple basis functions, MetaCluster achieves a reduction of up to 80$\times$ in parameter storage, with no loss in accuracy. Code will be released upon publication.

Deep neural networks currently demonstrate state-of-the-art performance in several domains. At the same time, models of this class are very demanding in terms of computational resources. In particular, a large amount of memory is required by commonly used fully-connected layers, making it hard to use the models on low-end devices and stopping the further increase of the model size. In this paper we convert the dense weight matrices of the fully-connected layers to the Tensor Train [17] format such that the number of parameters is reduced by a huge factor and at the same time the expressive power of the layer is preserved. In particular, for the V ery Deep VGG networks [21] we report the compression factor of the dense weight matrix of a fully-connected layer up to 200000 times leading to the compression factor of the whole network up to 7 times.

Deep neural networks currently demonstrate state-of-the-art performance in several domains.At the same time, models of this class are very demanding in terms of computational resources. In particular, a large amount of memory is required by commonly used fully-connected layers, making it hard to use the models on low-end devices and stopping the further increase of the model size. In this paper we convert the dense weight matrices of the fully-connected layers to the Tensor Train format such that the number of parameters is reduced by a huge factor and at the same time the expressive power of the layer is preserved.In particular, for the Very Deep VGG networks we report the compression factor of the dense weight matrix of a fully-connected layer up to 200000 times leading to the compression factor of the whole network up to 7 times.

--Semantic communication systems introduce a new paradigm in wireless communications, focusing on transmitting the intended meaning rather than ensuring strict bit-level accuracy. These systems often rely on Deep Neural Networks (DNNs) to learn and encode meaning directly from data, enabling more efficient communication. However, in multi-user settings where interacting agents are trained independently--without shared context or joint optimization--divergent latent representations across AI-native devices can lead to semantic mismatches, impeding mutual understanding even in the absence of traditional transmission errors. In this work, we address semantic mismatch in Multiple-Input Multiple-Output (MIMO) channels by proposing a joint physical and semantic channel equalization framework that leverages the presence of Reconfigurable Intelligent Surfaces (RIS). The semantic equalization is implemented as a sequence of transformations: (i) a pre-equalization stage at the transmitter; (ii) propagation through the RIS-aided channel; and (iii) a post-equalization stage at the receiver . We formulate the problem as a constrained Minimum Mean Squared Error (MMSE) optimization and propose two solutions: (i) a linear semantic equalization chain, and (ii) a non-linear DNN-based semantic equalizer . Both methods are designed to operate under semantic compression in the latent space and adhere to transmit power constraints. Through extensive evaluations, we show that the proposed joint equalization strategies consistently outperform conventional, disjoint approaches to physical and semantic channel equalization across a broad range of scenarios and wireless channel conditions. Index T erms --Semantic communications, latent space alignment, reconfigurable intelligent surfaces, 6G. OR the last seven decades, communication systems have been designed with the main objective of reliably transmitting symbols through noisy communication channels, typically disregarding the interpretation and impact of these symbols upon reception. Following this principle, communication networks have achieved significant advancements in bit transmission rate and reliability, fundamental metrics for data-centric applications such as video and audio streaming, where communication itself is the primary objective.

Model compression is critical for deploying deep learning models on resource-constrained devices. We introduce a novel method enhancing knowledge distillation with integrated gradients (IG) as a data augmentation strategy. Our approach overlays IG maps onto input images during training, providing student models with deeper insights into teacher models' decision-making processes. Extensive evaluation on CIFAR-10 demonstrates that our IG-augmented knowledge distillation achieves 92.6% testing accuracy with a 4.1x compression factor-a significant 1.1 percentage point improvement ($p<0.001$) over non-distilled models (91.5%). This compression reduces inference time from 140 ms to 13 ms. Our method precomputes IG maps before training, transforming substantial runtime costs into a one-time preprocessing step. Our comprehensive experiments include: (1) comparisons with attention transfer, revealing complementary benefits when combined with our approach; (2) Monte Carlo simulations confirming statistical robustness; (3) systematic evaluation of compression factor versus accuracy trade-offs across a wide range (2.2x-1122x); and (4) validation on an ImageNet subset aligned with CIFAR-10 classes, demonstrating generalisability beyond the initial dataset. These extensive ablation studies confirm that IG-based knowledge distillation consistently outperforms conventional approaches across varied architectures and compression ratios. Our results establish this framework as a viable compression technique for real-world deployment on edge devices while maintaining competitive accuracy.