In recent years, dense retrieval has been the focus of information retrieval (IR) research. While effective, dense retrieval produces uninterpretable dense vectors, and suffers from the drawback of large index size. Learned sparse retrieval (LSR) has emerged as promising alternative, achieving competitive retrieval performance while also being able to leverage the classical inverted index data structure for efficient retrieval. However, limited works have explored scaling LSR beyond BERT scale. In this work, we identify two challenges in training large language models (LLM) for LSR: (1) training instability during the early stage of contrastive training; (2) suboptimal performance due to pre-trained LLM's unidirectional attention. To address these challenges, we propose two corresponding techniques: (1) a lightweight adaptation training phase to eliminate training instability; (2) two model variants to enable bidirectional information. With these techniques, we are able to train LSR models with 8B scale LLM, and achieve competitive retrieval performance with reduced index size. Furthermore, we are among the first to analyze the performance-efficiency tradeoff of LLM-based LSR model through the lens of model quantization. Our findings provide insights into adapting LLMs for efficient retrieval modeling.
Layer-wise PTQ is a promising technique for compressing large language models (LLMs), due to its simplicity and effectiveness without requiring retraining. However, recent progress in this area is saturating, underscoring the need to revisit its core limitations and explore further improvements. We address this challenge by identifying a key limitation of existing layer-wise PTQ methods: the growth of quantization errors across layers significantly degrades performance, particularly in low-bit regimes. To address this fundamental issue, we propose Quantization Error Propagation (QEP), a general, lightweight, and scalable framework that enhances layer-wise PTQ by explicitly propagating quantization errors and compensating for accumulated errors. QEP also offers a tunable propagation mechanism that prevents overfitting and controls computational overhead, enabling the framework to adapt to various architectures and resource budgets. Extensive experiments on several LLMs demonstrate that QEP-enhanced layer-wise PTQ achieves substantially higher accuracy than existing methods. Notably, the gains are most pronounced in the extremely low-bit quantization regime.
The classical limit of quantum mechanics, formally investigated through frameworks like strict deformation quantization, remains a profound area of inquiry in the philosophy of physics. This paper explores a computational approach employing a neural network to emulate the emergence of classical behavior from the quantum harmonic oscillator as Planck's constant $\hbar$ approaches zero. We develop and train a neural network architecture to learn the mapping from initial expectation values and $\hbar$ to the time evolution of the expectation value of position. By analyzing the network's predictions across different regimes of hbar, we aim to provide computational insights into the nature of the quantum-classical transition. This work demonstrates the potential of machine learning as a complementary tool for exploring foundational questions in quantum mechanics and its classical limit.
Simultaneous Input and State Estimation (SISE) enables the reconstruction of unknown inputs and internal states in dynamical systems, with applications in fault detection, robotics, and control. While various methods exist for linear systems, extensions to systems with output quantization are scarce, and formal connections to limit Kalman filters in this context are lacking. This work addresses these gaps by proposing a novel SISE algorithm for linear systems with quantized output measurements that is based on a Gaussian mixture model formulation. The observation model is represented as a Gaussian sum density, leading to closed-form recursive equations in the form of a Gaussian sum filter. In the absence of input prior knowledge, the recursions converge to a limit-case SISE algorithm, implementable as a bank of linear SISE filters running in parallel. A simulation example is presented to illustrate the effectiveness of the proposed approach.
Generative masked transformers have demonstrated remarkable success across various content generation tasks, primarily due to their ability to effectively model large-scale dataset distributions with high consistency. However, in the animation domain, large datasets are not always available. Applying generative masked modeling to generate diverse instances from a single MoCap reference may lead to overfitting, a challenge that remains unexplored. In this work, we present MotionDreamer, a localized masked modeling paradigm designed to learn internal motion patterns from a given motion with arbitrary topology and duration. By embedding the given motion into quantized tokens with a novel distribution regularization method, MotionDreamer constructs a robust and informative codebook for local motion patterns. Moreover, a sliding window local attention is introduced in our masked transformer, enabling the generation of natural yet diverse animations that closely resemble the reference motion patterns. As demonstrated through comprehensive experiments, MotionDreamer outperforms the state-of-the-art methods that are typically GAN or Diffusion-based in both faithfulness and diversity. Thanks to the consistency and robustness of the quantization-based approach, MotionDreamer can also effectively perform downstream tasks such as temporal motion editing, \textcolor{update}{crowd animation}, and beat-aligned dance generation, all using a single reference motion. Visit our project page: https://motiondreamer.github.io/
Early exiting has recently emerged as a promising technique for accelerating large language models (LLMs) by effectively reducing the hardware computation and memory access. In this paper, we present SpecEE, a fast LLM inference engine with speculative early exiting. (1) At the algorithm level, we propose the speculation-based lightweight predictor design by exploiting the probabilistic correlation between the speculative tokens and the correct results and high parallelism of GPUs. (2) At the system level, we point out that not all layers need a predictor and design the two-level heuristic predictor scheduling engine based on skewed distribution and contextual similarity. (3) At the mapping level, we point out that different decoding methods share the same essential characteristics, and propose the context-aware merged mapping for predictor with efficient GPU implementations to support speculative decoding, and form a framework for various existing orthogonal acceleration techniques (e.g., quantization and sparse activation) on cloud and personal computer (PC) scenarios, successfully pushing the Pareto frontier of accuracy and speedup. It is worth noting that SpecEE can be applied to any LLM by negligible training overhead in advance without affecting the model original parameters. Extensive experiments show that SpecEE achieves 2.25x and 2.43x speedup with Llama2-7B on cloud and PC scenarios respectively.
Emergency of DeepSeek R1 and QwQ 32B have broken through performance barriers for running frontier large language models (LLMs) on home devices. While consumer hardware is getting stronger and model quantization is improving, existing end-side solutions still demand GPU clusters, large RAM/VRAM, and high bandwidth, far beyond what a common home cluster can handle. This paper introduces prima.cpp, a distributed inference system that runs 70B-scale models on everyday home devices using a mix of CPU/GPU, low RAM/VRAM, Wi-Fi, and cross-platform support. It uses mmap to manage model weights and introduces piped-ring parallelism with prefetching to hide disk loading. By modeling heterogeneity in computation, communication, disk, memory (and its management behavior), and OS, it optimally assigns model layers to each device's CPU and GPU, further reducing token latency. An elegant algorithm named Halda is proposed to solve this NP-hard assignment problem. We evaluate prima.cpp on a common four-node home cluster. It outperforms llama.cpp, exo, and dllama on 30B+ models while keeping memory pressure below 6%. This brings frontier 30B-70B models, such as Llama 3, DeepSeek R1, Qwen 2.5, and QwQ to home assistants, making advanced AI truly accessible to individuals. The code is open source and available at https://github.com/Lizonghang/prima.cpp.
In this paper, we propose methods to encrypted a pre-given dynamic controller with homomorphic encryption, without re-encrypting the control inputs. We first present a preliminary result showing that the coefficients in a pre-given dynamic controller can be scaled up into integers by the zooming-in factor in dynamic quantization, without utilizing re-encryption. However, a sufficiently small zooming-in factor may not always exist because it requires that the convergence speed of the pre-given closed-loop system should be sufficiently fast. Then, as the main result, we design a new controller approximating the pre-given dynamic controller, in which the zooming-in factor is decoupled from the convergence rate of the pre-given closed-loop system. Therefore, there always exist a (sufficiently small) zooming-in factor of dynamic quantization scaling up all the controller's coefficients to integers, and a finite modulus preventing overflow in cryptosystems. The process is asymptotically stable and the quantizer is not saturated.
Recent advances in visual synthesis have leveraged diffusion models and attention mechanisms to achieve high-fidelity artistic style transfer and photorealistic text-to-image generation. However, real-time deployment on edge devices remains challenging due to computational and memory constraints. We propose Muon-AD, a co-designed framework that integrates the Muon optimizer with attention distillation for real-time edge synthesis. By eliminating gradient conflicts through orthogonal parameter updates and dynamic pruning, Muon-AD achieves 3.2 times faster convergence compared to Stable Diffusion-TensorRT, while maintaining synthesis quality (15% lower FID, 4% higher SSIM). Our framework reduces peak memory to 7GB on Jetson Orin and enables 24FPS real-time generation through mixed-precision quantization and curriculum learning. Extensive experiments on COCO-Stuff and ImageNet-Texture demonstrate Muon-AD's Pareto-optimal efficiency-quality trade-offs. Here, we show a 65% reduction in communication overhead during distributed training and real-time 10s/image generation on edge GPUs. These advancements pave the way for democratizing high-quality visual synthesis in resource-constrained environments.
Diffusion Transformer (DiT) has driven significant progress in image generation tasks. However, DiT inferencing is notoriously compute-intensive and incurs long latency even on datacenter-scale GPUs, primarily due to its iterative nature and heavy reliance on GEMM operations inherent to its encoder-based structure. To address the challenge, prior work has explored quantization, but achieving low-precision quantization for DiT inferencing with both high accuracy and substantial speedup remains an open problem. To this end, this paper proposes MixDiT, an algorithm-hardware co-designed acceleration solution that exploits mixed Microscaling (MX) formats to quantize DiT activation values. MixDiT quantizes the DiT activation tensors by selectively applying higher precision to magnitude-based outliers, which produce mixed-precision GEMM operations. To achieve tangible speedup from the mixed-precision arithmetic, we design a MixDiT accelerator that enables precision-flexible multiplications and efficient MX precision conversions. Our experimental results show that MixDiT delivers a speedup of 2.10-5.32 times over RTX 3090, with no loss in FID.
Spiking Neural Networks (SNNs) are highly energy-efficient during inference, making them particularly suitable for deployment on neuromorphic hardware. Their ability to process event-driven inputs, such as data from dynamic vision sensors (DVS), further enhances their applicability to edge computing tasks. However, the resource constraints of edge hardware necessitate techniques like weight quantization, which reduce the memory footprint of SNNs while preserving accuracy. Despite its importance, existing quantization methods typically focus on synaptic weights quantization without taking account of other critical parameters, such as scaling neuron firing thresholds. To address this limitation, we present the first benchmark for the DVS gesture recognition task using SNNs optimized for the many-core neuromorphic chip SpiNNaker2. Our study evaluates two quantization pipelines for fixed-point computations. The first approach employs post training quantization (PTQ) with percentile-based threshold scaling, while the second uses quantization aware training (QAT) with adaptive threshold scaling. Both methods achieve accurate 8-bit on-chip inference, closely approximating 32-bit floating-point performance. Additionally, our baseline SNNs perform competitively against previously reported results without specialized techniques. These models are deployed on SpiNNaker2 using the neuromorphic intermediate representation (NIR). Ultimately, we achieve 94.13% classification accuracy on-chip, demonstrating the SpiNNaker2's potential for efficient, low-energy neuromorphic computing.
Multimodal Large Language Models (MLLMs) have demonstrated remarkable performance across diverse applications. However, their computational overhead during deployment remains a critical bottleneck. While Key-Value (KV) caching effectively trades memory for computation to enhance inference efficiency, the growing memory footprint from extensive KV caches significantly reduces throughput and restricts prolonged deployment on memory-constrained GPU devices. To address this challenge, we propose CalibQuant, a simple yet highly effective visual quantization strategy that drastically reduces both memory and computational overhead. Specifically, CalibQuant introduces an extreme 1-bit quantization scheme, complemented by novel post-scaling and calibration techniques tailored to the intrinsic patterns of KV caches, thereby ensuring high efficiency without compromising model performance. Leveraging Triton for runtime optimization, we achieve a 10x throughput increase on InternVL models. Our method is designed to be plug-and-play, seamlessly integrating with various existing MLLMs without requiring architectural changes. Extensive experiments confirm that our approach significantly reduces memory usage while maintaining computational efficiency and preserving multimodal capabilities. Codes are available at https://github.com/insuhan/calibquant.
Multi-modal sequential recommendation systems leverage auxiliary signals (e.g., text, images) to alleviate data sparsity in user-item interactions. While recent methods exploit large language models to encode modalities into discrete semantic IDs for autoregressive prediction, we identify two critical limitations: (1) Existing approaches adopt fragmented quantization, where modalities are independently mapped to semantic spaces misaligned with behavioral objectives, and (2) Over-reliance on semantic IDs disrupts inter-modal semantic coherence, thereby weakening the expressive power of multi-modal representations for modeling diverse user preferences. To address these challenges, we propose a Behavior-Bind multi-modal Quantization for Sequential Recommendation (BBQRec for short) featuring dual-aligned quantization and semantics-aware sequence modeling. First, our behavior-semantic alignment module disentangles modality-agnostic behavioral patterns from noisy modality-specific features through contrastive codebook learning, ensuring semantic IDs are inherently tied to recommendation tasks. Second, we design a discretized similarity reweighting mechanism that dynamically adjusts self-attention scores using quantized semantic relationships, preserving multi-modal synergies while avoiding invasive modifications to the sequence modeling architecture. Extensive evaluations across four real-world benchmarks demonstrate BBQRec's superiority over the state-of-the-art baselines.
Deep learning-based computer vision systems adopt complex and large architectures to improve performance, yet they face challenges in deployment on resource-constrained mobile and edge devices. To address this issue, model compression techniques such as pruning, quantization, and matrix factorization have been proposed; however, these compressed models are often highly vulnerable to adversarial attacks. We introduce the \textbf{Efficient Ensemble Defense (EED)} technique, which diversifies the compression of a single base model based on different pruning importance scores and enhances ensemble diversity to achieve high adversarial robustness and resource efficiency. EED dynamically determines the number of necessary sub-models during the inference stage, minimizing unnecessary computations while maintaining high robustness. On the CIFAR-10 and SVHN datasets, EED demonstrated state-of-the-art robustness performance compared to existing adversarial pruning techniques, along with an inference speed improvement of up to 1.86 times. This proves that EED is a powerful defense solution in resource-constrained environments.
We present a design called \emph{Proof of Gradient Optimization} (PoGO) for blockchain consensus, where miners produce verifiable evidence of training large-scale machine-learning models. Building on previous work, we incorporate \emph{quantized gradients} (4-bit precision) to reduce storage and computation requirements, while still preserving the ability of verifiers to check that real progress has been made on lowering the model's loss. Additionally, we employ Merkle proofs over the full 32-bit model to handle large parameter sets and to enable random leaf checks with minimal on-chain data. We illustrate these ideas using GPT-3 (175B parameters) as a reference example and also refer to smaller but high-performance models (e.g., \emph{Gemma~3} with 27B parameters). We provide an empirical cost analysis showing that verification is significantly cheaper than training, thanks in part to quantization and sampling. We also discuss the necessity of longer block times (potentially hours) when incorporating meaningful training steps, the trade-offs when using specialized GPU hardware, and how binary diffs may incrementally optimize updates. Finally, we note that fine-tuning can be handled in a similar manner, merely changing the dataset and the manner of sampling but preserving the overall verification flow. Our protocol allows verifiers to issue either \emph{positive} or \emph{negative} attestations; these are aggregated at finalization to either confirm the update or slash the miner.
Post-training quantization (PTQ) reduces a model's memory footprint by mapping full precision weights into low bit weights without costly retraining, but can degrade its downstream performance especially in low 2- to 3-bit settings. We develop a new mixed-precision PTQ approach, Task-Circuit Quantization (TaCQ), that draws parallels to automated circuit discovery, directly conditioning the quantization process on specific weight circuits -- which we define as sets of weights associated with downstream task performance. These weights are kept as 16-bit weights, while others are quantized, maintaining performance while only adding a marginal memory cost. Specifically, TaCQ contrasts unquantized model weights with a uniformly-quantized model to estimate the expected change in weights due to quantization and uses gradient information to predict the resulting impact on task performance, allowing us to preserve task-specific weights. We compare TaCQ-based quantization to existing mixed-precision quantization methods when conditioning both on general-purpose and task-specific data. Across QA, math reasoning, and text-to-SQL tasks for both Llama-3 and Qwen2.5, we find that TaCQ outperforms baselines using the same calibration data and a lower weight budget, achieving major improvements in the 2 and 3-bit regime. With only 3.1 bits we are able to recover 96% of Llama-3-8B-Instruct's unquantized 16-bit MMLU performance, obtaining a 5.25% absolute improvement over SPQR. We also observe consistently large gains over existing methods in the 2-bit regime, with an average gain of 14.74% over the strongest baseline, SliM-LLM. Moreover, we observe a 7.20% gain without conditioning on specific tasks, showing TaCQ's ability to identify important weights is not limited to task-conditioned settings.
High dynamic range (HDR) imaging is vital for capturing the full range of light tones in scenes, essential for computer vision tasks such as autonomous driving. Standard commercial imaging systems face limitations in capacity for well depth, and quantization precision, hindering their HDR capabilities. Modulo imaging, based on unlimited sampling (US) theory, addresses these limitations by using a modulo analog-to-digital approach that resets signals upon saturation, enabling estimation of pixel resets through neighboring pixel intensities. Despite the effectiveness of (US) algorithms in one-dimensional signals, their optimization problem for two-dimensional signals remains unclear. This work formulates the US framework as an autoregressive $\ell_2$ phase unwrapping problem, providing computationally efficient solutions in the discrete cosine domain jointly with a stride removal algorithm also based on spatial differences. By leveraging higher-order finite differences for two-dimensional images, our approach enhances HDR image reconstruction from modulo images, demonstrating its efficacy in improving object detection in autonomous driving scenes without retraining.
We present the implementation of four FPGA-accelerated convolutional neural network (CNN) models for onboard cloud detection in resource-constrained CubeSat missions, leveraging Xilinx's Vitis AI (VAI) framework and Deep Learning Processing Unit (DPU), a programmable engine with pre-implemented, parameterizable IP cores optimized for deep neural networks, on a Zynq UltraScale+ MPSoC. This study explores both pixel-wise (Pixel-Net and Patch-Net) and image-wise (U-Net and Scene-Net) models to benchmark trade-offs in accuracy, latency, and model complexity. Applying channel pruning, we achieved substantial reductions in model parameters (up to 98.6%) and floating-point operations (up to 90.7%) with minimal accuracy loss. Furthermore, the VAI tool was used to quantize the models to 8-bit precision, ensuring optimized hardware performance with negligible impact on accuracy. All models retained high accuracy post-FPGA integration, with a cumulative maximum accuracy drop of only 0.6% after quantization and pruning. The image-wise Scene-Net and U-Net models demonstrated strong real-time inference capabilities, achieving frame rates per second of 57.14 and 37.45, respectively, with power consumption of around 2.5 W, surpassing state-of-the-art onboard cloud detection solutions. Our approach underscores the potential of DPU-based hardware accelerators to expand the processing capabilities of small satellites, enabling efficient and flexible onboard CNN-based applications.
Self-supervised learning has become a core technique in speech processing, but the high dimensionality of its representations makes discretization essential for improving efficiency. However, existing discretization methods still suffer from significant information loss, resulting in a notable performance gap compared to continuous representations. To overcome these limitations, we propose two quantization-based discretization methods: Product Quantization (PQ) and Random Product Quantization (RPQ). PQ partitions the original feature space into multiple subspaces and independently quantizes each sub-vector, producing a fused set of discrete units that retain diverse information from different subspaces, thus mitigating the loss associated with single-cluster quantization. RPQ further enhances representation diversity by randomly sampling a fixed proportion of feature dimensions multiple times to construct sub-vectors, thereby better capturing the variability in the data distribution. Theoretical analysis shows that RPQ reduces the correlation coefficient rho (where 0 <= rho <= 1) between sub-quantizers. Its quantization error is lower-bounded by the product of rho and epsilon-kms, where epsilon-kms denotes the quantization error of a single K-means quantizer. Experimental results on a combined dataset built from LibriSpeech and ML-SUPERB show that PQ and RPQ outperform standard K-means discretization, achieving relative improvements of 21.8 percent and 20.0 percent in WER on LibriSpeech, and 24.1 percent and 19.6 percent in CER on ML-SUPERB, respectively. Moreover, their performance is competitive with, and in some cases even surpasses, that of continuous SSL representations.
The proliferation of Large Language Models (LLMs) accessed via black-box APIs introduces a significant trust challenge: users pay for services based on advertised model capabilities (e.g., size, performance), but providers may covertly substitute the specified model with a cheaper, lower-quality alternative to reduce operational costs. This lack of transparency undermines fairness, erodes trust, and complicates reliable benchmarking. Detecting such substitutions is difficult due to the black-box nature, typically limiting interaction to input-output queries. This paper formalizes the problem of model substitution detection in LLM APIs. We systematically evaluate existing verification techniques, including output-based statistical tests, benchmark evaluations, and log probability analysis, under various realistic attack scenarios like model quantization, randomized substitution, and benchmark evasion. Our findings reveal the limitations of methods relying solely on text outputs, especially against subtle or adaptive attacks. While log probability analysis offers stronger guarantees when available, its accessibility is often limited. We conclude by discussing the potential of hardware-based solutions like Trusted Execution Environments (TEEs) as a pathway towards provable model integrity, highlighting the trade-offs between security, performance, and provider adoption. Code is available at https://github.com/sunblaze-ucb/llm-api-audit