Large Language Models (LLMs) hold significant potential for enhancing healthcare applications, yet their deployment is hindered by high computational and memory demands. Model compression techniques offer solutions to reduce these demands, but their impact on medical LLMs remains underexplored. In this paper, we introduce CMedBench, the first comprehensive benchmark for evaluating compressed LLMs in medical contexts. CMedBench assesses five core dimensions: Medical Knowledge Ability, Medical Application Ability, Trustworthiness Maintenance, Compression Cross Combination, and Computational Efficiency. Through extensive empirical studies, we analyze the trade-offs between model efficiency and clinical performance across diverse models, datasets, and compression strategies. Our findings highlight critical limitations in current evaluation practices and provide a robust framework for aligning compression strategies with medical requirements. CMedBench serves as a vital resource for researchers and practitioners, guiding the development of efficient, trustworthy, and clinically effective LLMs for healthcare applications.

Tree of Thoughts (ToT) enhances Large Language Model (LLM) reasoning by structuring problem-solving as a spanning tree. However, recent methods focus on search accuracy while overlooking computational efficiency. The challenges of accelerating the ToT lie in the frequent switching of reasoning focus, and the redundant exploration of suboptimal solutions. To alleviate this dilemma, we propose Dynamic Parallel Tree Search (DPTS), a novel parallelism framework that aims to dynamically optimize the reasoning path in inference. It includes the Parallelism Streamline in the generation phase to build up a flexible and adaptive parallelism with arbitrary paths by cache management and alignment. Meanwhile, the Search and Transition Mechanism filters potential candidates to dynamically maintain the reasoning focus on more possible solutions with less redundancy. Experiments on Qwen-2.5 and Llama-3 on math and code datasets show that DPTS significantly improves efficiency by 2-4\times on average while maintaining or even surpassing existing reasoning algorithms in accuracy, making ToT-based reasoning more scalable and computationally efficient. Codes are released at: https://github.com/yifu-ding/DPTS.

Although knowledge distillation (KD) is an effective approach to improve the performance of a smaller LLM (i.e., the student model) by transferring knowledge from a large LLM (i.e., the teacher model), it still suffers from high training cost. Existing LLM distillation methods ignore the difficulty difference among different samples, making the distillation of easy samples unnecessary. This leads to high distillation cost. In this paper, we propose difficulty-aware knowledge distillation (DA-KD) framework for efficient knowledge distillation, in which we dynamically adjust the distillation dataset based on the difficulty of samples. We further observe existing KD loss cannot perform well when most of samples are difficult in the distillation dataset because of unstable optimization and the neglect of hard samples. Therefore, we also propose a new KD loss called bidirectional discrepancy loss (BDL) for effective KD. Extensive experiments demonstrate that our DA-KD framework is effective and efficient. Without bells and whistles, DA-KD can outperform existing state-of-the-art KD methods by 2% with half training cost and even surpass the teacher model with 4.7$\times$ compression.

Video diffusion transformers have achieved remarkable progress in high-quality video generation, but remain computationally expensive due to the quadratic complexity of attention over high-dimensional video sequences. Recent acceleration methods enhance the efficiency by exploiting the local sparsity of attention scores; yet they often struggle with accelerating the long-range computation. To address this problem, we propose VORTA, an acceleration framework with two novel components: 1) a sparse attention mechanism that efficiently captures long-range dependencies, and 2) a routing strategy that adaptively replaces full 3D attention with specialized sparse attention variants. VORTA achieves an end-to-end speedup $1.76\times$ without loss of quality on VBench. Furthermore, it can seamlessly integrate with various other acceleration methods, such as model caching and step distillation, reaching up to speedup $14.41\times$ with negligible performance degradation. VORTA demonstrates its efficiency and enhances the practicality of video diffusion transformers in real-world settings. Codes and weights are available at https://github.com/wenhao728/VORTA.

Although deep learning based object detection is of great significance for various applications, it faces challenges when deployed on edge devices due to the computation and energy limitations. Post-training quantization (PTQ) can improve inference efficiency through integer computing. However, they suffer from severe performance degra-dation when performing full quantization due to overlooking the unique characteristics of regression tasks in ob-ject detection. In this paper, we are the first to explore regression-friendly quantization and conduct full quantization on various detectors. We reveal the intrinsic reason behind the difficulty of quantizing regressors with empir-ical and theoretical justifications, and introduce a novel Regression-specialized Post-Training Quantization (Reg- PTQ) scheme. It includes Filtered Global Loss Integration Calibration to combine the global loss with a two-step fil-tering mechanism, mitigating the adverse impact of false positive bounding boxes, and Learnable Logarithmic-Affine Quantizer tailored for the non-uniform distributed param-eters in regression structures. Extensive experiments on prevalent detectors showcase the effectiveness of the well-designed Reg-PTQ. Notably, our Reg-PTQ achieves 7.6x and 5.4x reduction in computation and storage consumption under INT4 with little performance degradation, which indicates the immense potential of fully quantized detectors in real-world object detection applications.

Segment Anything Model (SAM) has achieved impressive performance in many computer vision tasks. However, as a large-scale model, the immense memory and computation costs hinder its practical deployment. In this paper, we pro-pose a post-training quantization (PTQ)frameworkfor Segment Anything Model, namely PTQ4SAM. First, we investigate the inherent bottleneck of SAM quantization attributed to the bimodal distribution in post-Key-Linear activations. We analyze its characteristics from both per-tensor and per-channel perspectives, and propose a Bimodal Integration strategy, which utilizes a mathematically equivalent sign operation to transform the bimodal distribution into a relatively easy-quantized normal distribution offline. Second, SAM encompasses diverse attention mechanisms (i.e., self-attention and two-way cross-attention), resulting in substantial variations in the post-Softmax distributions. Therefore, we introduce an Adaptive Granularity Quantization for Softmax through searching the optimal power-of-two base, which is hardware-friendly. Extensive experimen-tal results across various vision tasks (instance segmentation, semantic segmentation and object detection), datasets and model variants show the superiority of PTQ4SAM. For example, when quantizing SAM-L to 6-bit, we achieve loss-less accuracy for instance segmentation, about 0.5% drop with theoretical3.9x acceleration. The code is available at https://github.com/chengtao-lv/PTQ4SAM.

In this paper, we introduce a novel model compression technique named Joint Sparsification and Quantization (JSQ), explicitly tailored for large language models (LLMs). Traditional methods employ either sparsification or quantization individually to compress LLMs, leading to performance degradation at high compression ratios. In contrast, our JSQ approach integrates sparsification and quantization cohesively. As sparsification tend to preserve outliers that is harmful to quantization, we introduce a novel sparsity metric to serves as a bridge between the sparsification and quantization. Moreover, it is proven outliers in LLMs have significant impact but harmful to compression. Current solutions are highly coupled with quantization process, which is not helpful to sparsification. To this end, we also introduce a search-based activation editor to automatically eliminate relatively useless outliers. Comprehensive experiments across various datasets and architectures affirm the efficacy of our JSQ framework. Notably, our JSQ achieves 7.96$\times$ computation reduction without crashing for the representative model LLaMA. This accomplishment stands in stark contrast to the limitations of most state-of-the-art LLM compression methods, which typically fail under such extreme compression ratios. Our code is released at https://github.com/uanu2002/JSQ.

Although large language models (LLMs) have demonstrated their strong intelligence ability, the high demand for computation and storage hinders their practical application. To this end, many model compression techniques are proposed to increase the efficiency of LLMs. However, current researches only validate their methods on limited models, datasets, metrics, etc, and still lack a comprehensive evaluation under more general scenarios. So it is still a question of which model compression approach we should use under a specific case. To mitigate this gap, we present the Large Language Model Compression Benchmark (LLMCBench), a rigorously designed benchmark with an in-depth analysis for LLM compression algorithms. We first analyze the actual model production requirements and carefully design evaluation tracks and metrics. Then, we conduct extensive experiments and comparison using multiple mainstream LLM compression approaches. Finally, we perform an in-depth analysis based on the evaluation and provide useful insight for LLM compression design. We hope our LLMCBench can contribute insightful suggestions for LLM compression algorithm design and serve as a foundation for future research.

Deep neural networks, such as the deep-FSMN, have been widely studied for keyword spotting (KWS) applications while suffering expensive computation and storage. Therefore, network compression technologies such as binarization are studied to deploy KWS models on edge. In this article, we present a strong yet efficient binary neural network for KWS, namely, BiFSMNv2, pushing it to the real-network accuracy performance. First, we present a dual-scale thinnable 1-bit-architecture (DTA) to recover the representation capability of the binarized computation units by dual-scale activation binarization and liberate the speedup potential from an overall architecture perspective. Second, we also construct a frequency-independent distillation (FID) scheme for KWS binarization-aware training, which distills the high- and low-frequency components independently to mitigate the information mismatch between full-precision and binarized representations. Moreover, we propose the learning propagation binarizer (LPB), a general and efficient binarizer that enables the forward and backward propagation of binary KWS networks to be continuously improved through learning. We implement and deploy BiFSMNv2 on ARMv8 real-world hardware with a novel fast bitwise computation kernel (FBCK), which is proposed to fully use registers and increase instruction throughput. Comprehensive experiments show our BiFSMNv2 outperforms the existing binary networks for KWS by convincing margins across different datasets and achieves comparable accuracy with the full-precision networks (only a tiny 1.51% drop on Speech Commands V1-12). We highlight that benefiting from the compact architecture and optimized hardware kernel, BiFSMNv2 can achieve an impressive 25.1× speedup and 20.2× storage-saving on edge hardware.

Network binarization emerges as one of the most promising compression approaches offering extraordinary computation and memory savings by minimizing the bit-width. However, recent research has shown that applying existing binarization algorithms to diverse tasks, architectures, and hardware in realistic scenarios is still not straightforward. Common challenges of binarization, such as accuracy degradation and efficiency limitation, suggest that its attributes are not fully understood. To close this gap, we present BiBench, a rigorously designed benchmark with in-depth analysis for network binarization. We first carefully scrutinize the requirements of binarization in the actual production and define evaluation tracks and metrics for a comprehensive and fair investigation. Then, we evaluate and analyze a series of milestone binarization algorithms that function at the operator level and with extensive influence. Our benchmark reveals that 1) the binarized operator has a crucial impact on the performance and deployability of binarized networks; 2) the accuracy of binarization varies significantly across different learning tasks and neural architectures; 3) binarization has demonstrated promising efficiency potential on edge devices despite the limited hardware support. The results and analysis also lead to a promising paradigm for accurate and efficient binarization. We believe that BiBench will contribute to the broader adoption of binarization and serve as a foundation for future research. The code for our BiBench is released https://github.com/htqin/BiBench .

Low-bit quantization in image super-resolution (SR) has attracted copious attention in recent research due to its ability to reduce parameters and operations significantly. However, many quantized SR models suffer from accuracy degradation compared to their full-precision counterparts, especially at ultra-low bit widths (2-4 bits), limiting their practical applications. To address this issue, we propose a novel quantized image SR network, called QuantSR, which achieves accurate and efficient SR processing under low-bit quantization. To overcome the representation homogeneity caused by quantization in the network, we introduce the Redistribution-driven Learnable Quantizer (RLQ). This is accomplished through an inference-agnostic efficient redistribution design, which adds additional information in both forward and backward passes to improve the representation ability of quantized networks. Furthermore, to achieve flexible inference and break the upper limit of accuracy, we propose the Depth-dynamic Quantized Architecture (DQA). Our DQA allows for the trade-off between efficiency and accuracy during inference through weight sharing. Our comprehensive experiments show that QuantSR outperforms existing state-of-the-art quantized SR networks in terms of accuracy while also providing more competitive computational efficiency. In addition, we demonstrate the scheme's satisfactory architecture generality by providing QuantSR-C and QuantSR-T for both convolution and Transformer versions, respectively. Our code and models are released at https://github.com/htqin/QuantSR .

Generative data-free quantization emerges as a practical compression approach that quantizes deep neural networks to low bit-width without accessing the real data. This approach generates data utilizing batch normalization (BN) statistics of the full-precision networks to quantize the networks. However, it always faces the serious challenges of accuracy degradation in practice. We first give a theoretical analysis that the diversity of synthetic samples is crucial for the data-free quantization, while in existing approaches, the synthetic data completely constrained by BN statistics experimentally exhibit severe homogenization at distribution and sample levels. This paper presents a generic Diverse Sample Generation (DSG) scheme for the generative data-free quantization, to mitigate detrimental homogenization. We first slack the statistics alignment for features in the BN layer to relax the distribution constraint. Then, we strengthen the loss impact of the specific BN layers for different samples and inhibit the correlation among samples in the generation process, to diversify samples from the statistical and spatial perspectives, respectively. Comprehensive experiments show that for large-scale image classification tasks, our DSG can consistently quantization performance on different neural architectures, especially under ultra-low bit-width. And data diversification caused by our DSG brings a general gain to various quantization-aware training and post-training quantization approaches, demonstrating its generality and effectiveness.

Action Recognition is a fundamental task in computer vision field, with a wide range of applications in autonomous driving, security monitoring, etc. However, previous action recognition approaches usually suffer from the inappropriate spatio-temporal modeling or high computational consumption (e.g., 3D CNN). In this paper, we propose a novel Spatio-Temporal Adaptive Network (STANet) with bidirectional temporal difference, consisting of a Temporal Adaptive module (TA) and a Spatial Adaptive (SA) module, to sufficiently extract the crucial motion information and model the spatial pivotal appearance information from both forward and backward perspectives, respectively. Specifically, the Temporal Adaptive module uses bidirectional temporal differences to learn valuable motion trends and balance the static semantics and dynamic motion for a certain action during information fusion; while the Spatial Adaptive module uses the bidirectional temporal difference to obtain the spatio-channel attention to stress the discriminative position-relevant and semantic-relevant appearance features. Extensive experiments conducted on widely-used action recognition benchmarks UCF-101, HMDB-51, Something-Something V1, and Kinetics-400 prove the effectiveness of the proposed methods compared to other state-of-the-art approaches.

Existing cross-domain detection methods mostly study the domain shifts where differences between domains are often caused by external environment and perceivable for humans. However, in real-world scenarios (e.g., MRI medical diagnosis, X-ray security inspection), there still exists another type of shift, named endogenous shift, where the differences between domains are mainly caused by the intrinsic factors (e.g., imaging mechanisms, hardware components, etc.), and usually inconspicuous. This shift can also severely harm the cross-domain detection performance but has been rarely studied. To support this study, we contribute the first Endogenous Domain Shift (EDS) benchmark, X-ray security inspection, where the endogenous shifts among the domains are mainly caused by different X-ray machine types with different hardware parameters, wear degrees, etc. EDS consists of 14,219 images including 31,654 common instances from three domains (X-ray machines), with bounding-box annotations from 10 categories. To handle the endogenous shift, we further introduce the Perturbation Suppression Network (PSN), motivated by the fact that this shift is mainly caused by two types of perturbations: category-dependent and category-independent ones. PSN respectively exploits local prototype alignment and global adversarial learning mechanism to suppress these two types of perturbations. The comprehensive evaluation results show that PSN outperforms SOTA methods, serving a new perspective to the cross-domain research community.

The deep neural networks, such as the Deep-FSMN, have been widely studied for keyword spotting (KWS) applications. However, computational resources for these networks are significantly constrained since they usually run on-call on edge devices. In this paper, we present BiFSMN, an accurate and extreme-efficient binary neural network for KWS. We first construct a High-frequency Enhancement Distillation scheme for the binarization-aware training, which emphasizes the high-frequency information from the full-precision network's representation that is more crucial for the optimization of the binarized network. Then, to allow the instant and adaptive accuracy-efficiency trade-offs at runtime, we also propose a Thinnable Binarization Architecture to further liberate the acceleration potential of the binarized network from the topology perspective. Moreover, we implement a Fast Bitwise Computation Kernel for BiFSMN on ARMv8 devices which fully utilizes registers and increases instruction throughput to push the limit of deployment efficiency. Extensive experiments show that BiFSMN outperforms existing binarization methods by convincing margins on various datasets and is even comparable with the full-precision counterpart (e.g., less than 3% drop on Speech Commands V1-12). We highlight that benefiting from the thinnable architecture and the optimized 1-bit implementation, BiFSMN can achieve an impressive 22.3x speedup and 15.5x storage-saving on real-world edge hardware.

Vision transformer emerges as a potential architecture for vision tasks. However, the intense computation and non-negligible delay hinder its application in the real world. As a widespread model compression technique, existing post-training quantization methods still cause severe performance drops. We find the main reasons lie in (1) the existing calibration metric is inaccurate in measuring the quantization influence for extremely low-bit representation, and (2) the existing quantization paradigm is unfriendly to the power-law distribution of Softmax. Based on these observations, we propose a novel Accurate Post-training Quantization framework for Vision Transformer, namely APQ-ViT. We first present a unified Bottom-elimination Blockwise Calibration scheme to optimize the calibration metric to perceive the overall quantization disturbance in a blockwise manner and prioritize the crucial quantization errors that influence more on the final output. Then, we design a Matthew-effect Preserving Quantization for Softmax to maintain the power-law character and keep the function of the attention mechanism. Comprehensive experiments on large-scale classification and detection datasets demonstrate that our APQ-ViT surpasses the existing post-training quantization methods by convincing margins, especially in lower bit-width settings (e.g., averagely up to 5.17% improvement for classification and 24.43% for detection on W4A4). We also highlight that APQ-ViT enjoys versatility and works well on diverse transformer variants.

Quantization has emerged as one of the most prevalent approaches to compress and accelerate neural networks. Recently, data-free quantization has been widely studied as a practical and promising solution. It synthesizes data for calibrating the quantized model according to the batch normalization (BN) statistics of FP32 ones and significantly relieves the heavy dependency on real training data in traditional quantization methods. Unfortunately, we find that in practice, the synthetic data identically constrained by BN statistics suffers serious homogenization at both distribution level and sample level and further causes a significant performance drop of the quantized model. We propose Diverse Sample Generation (DSG) scheme to mitigate the adverse effects caused by homogenization. Specifically, we slack the alignment of feature statistics in the BN layer to relax the constraint at the distribution level and design a layerwise enhancement to reinforce specific layers for different data samples. Our DSG scheme is versatile and even able to be applied to the state-of-the-art post-training quantization method like AdaRound. We evaluate the DSG scheme on the large-scale image classification task and consistently obtain significant improvements over various network architectures and quantization methods, especially when quantized to lower bits (e.g., up to 22% improvement on W4A4). Moreover, benefiting from the enhanced diversity, models calibrated with synthetic data perform close to those calibrated with real data and even outperform them on W4A4.

Few-shot learning is an interesting and challenging study, which enables machines to learn from few samples like humans. Existing studies rarely exploit auxiliary information from large amount of unlabeled data. Self-supervised learning is emerged as an efficient method to utilize unlabeled data. Existing self-supervised learning methods always rely on the combination of geometric transformations for the single sample by augmentation, while seriously neglect the endogenous correlation information among different samples that is the same important for the task. In this work, we propose a Graph-driven Clustering (GC), a novel augmentation-free method for self-supervised learning, which does not rely on any auxiliary sample and utilizes the endogenous correlation information among input samples. Besides, we propose Multi-pretext Attention Network (MAN), which exploits a specific attention mechanism to combine the traditional augmentation-relied methods and our GC, adaptively learning their optimized weights to improve the performance and enabling the feature extractor to obtain more universal representations. We evaluate our MAN extensively on miniImageNet and tieredImageNet datasets and the results demonstrate that the proposed method outperforms the state-of-the-art (SOTA) relevant methods.1