Jinyang Guo

Xiuying Wei, Yunchen Zhang, Yuhang Li et al. · ibm-research

Post-training quantization~(PTQ) of transformer language models faces significant challenges due to the existence of detrimental outliers in activations. We observe that these outliers are concentrated in specific channels and are asymmetric across channels. To address this issue, we propose the Outlier Suppression+~(OS+) framework, which contains the channel-wise shifting for asymmetry and channel-wise scaling for concentration. We show that these operations can be seamlessly migrated into subsequent modules while maintaining equivalence. Second, we propose a fast and stable scheme to calculate effective shifting and scaling values. The channel-wise shifting aligns the center of each channel for removal of outlier asymmetry. The channel-wise scaling quantitatively evaluates changes brought by migration and quantization for better quantization burden balance. We validate our OS+ under both standard and fine-grained quantization settings with models including BERT, OPT, BLOOM, BLOOMZ, and LLaMA. Comprehensive results across various tasks demonstrate the superiority of our approach. Especially, with standard quantization, OS+ can achieve near-floating-point performance on both small models and large language models on 8-bit and 6-bit. Besides, we establish a new state-of-the-art for 4-bit BERT with 15.5\% improvement. Our code is available at \url{https://github.com/ModelTC/Outlier_Suppression_Plus}.

Qiuhong Shen, Lei Qiao, Jinyang Guo et al.

Unsupervised learning has been popular in various computer vision tasks, including visual object tracking. However, prior unsupervised tracking approaches rely heavily on spatial supervision from template-search pairs and are still unable to track objects with strong variation over a long time span. As unlimited self-supervision signals can be obtained by tracking a video along a cycle in time, we investigate evolving a Siamese tracker by tracking videos forward-backward. We present a novel unsupervised tracking framework, in which we can learn temporal correspondence both on the classification branch and regression branch. Specifically, to propagate reliable template feature in the forward propagation process so that the tracker can be trained in the cycle, we first propose a consistency propagation transformation. We then identify an ill-posed penalty problem in conventional cycle training in backward propagation process. Thus, a differentiable region mask is proposed to select features as well as to implicitly penalize tracking errors on intermediate frames. Moreover, since noisy labels may degrade training, we propose a mask-guided loss reweighting strategy to assign dynamic weights based on the quality of pseudo labels. In extensive experiments, our tracker outperforms preceding unsupervised methods by a substantial margin, performing on par with supervised methods on large-scale datasets such as TrackingNet and LaSOT. Code is available at https://github.com/FlorinShum/ULAST.

Jinyang Du, Shenghao Jin, Ziqian Xu et al.

Large video diffusion models achieve strong visual quality but remain expensive to deploy because each sample requires many denoising steps and a large resident parameter footprint. This paper studies a deployment-oriented compression pipeline for Wan2.2-T2V-A14B by combining few-step distribution-matching distillation with low-bit quantization. The pipeline follows the model's dual-expert denoising route, calibrates the high-noise and low-noise branches separately, protects sensitive entrance layers, and uses HiF4-style low-bit representation to improve dynamic-range coverage. Quantization is calibrated on the distilled few-step student rather than on the original long-step trajectory, reducing activation-distribution mismatch during inference. The proposed co-design keeps the quantized model close to the same-step full-precision model and surpasses the original full-precision baseline at 8 and 20 steps on average. The 20-step setting gives the best quality-efficiency trade-off in the tested configurations.

Yisong Xiao, Aishan Liu, Tianyuan Zhang et al.

Quantization has emerged as an essential technique for deploying deep neural networks (DNNs) on devices with limited resources. However, quantized models exhibit vulnerabilities when exposed to various noises in real-world applications. Despite the importance of evaluating the impact of quantization on robustness, existing research on this topic is limited and often disregards established principles of robustness evaluation, resulting in incomplete and inconclusive findings. To address this gap, we thoroughly evaluated the robustness of quantized models against various noises (adversarial attacks, natural corruptions, and systematic noises) on ImageNet. The comprehensive evaluation results empirically provide valuable insights into the robustness of quantized models in various scenarios, for example: (1) quantized models exhibit higher adversarial robustness than their floating-point counterparts, but are more vulnerable to natural corruptions and systematic noises; (2) in general, increasing the quantization bit-width results in a decrease in adversarial robustness, an increase in natural robustness, and an increase in systematic robustness; (3) among corruption methods, \textit{impulse noise} and \textit{glass blur} are the most harmful to quantized models, while \textit{brightness} has the least impact; (4) among systematic noises, the \textit{nearest neighbor interpolation} has the highest impact, while bilinear interpolation, cubic interpolation, and area interpolation are the three least harmful. Our research contributes to advancing the robust quantization of models and their deployment in real-world scenarios.

Zhihao Hu, Guo Lu, Jinyang Guo et al.

The previous deep video compression approaches only use the single scale motion compensation strategy and rarely adopt the mode prediction technique from the traditional standards like H.264/H.265 for both motion and residual compression. In this work, we first propose a coarse-to-fine (C2F) deep video compression framework for better motion compensation, in which we perform motion estimation, compression and compensation twice in a coarse to fine manner. Our C2F framework can achieve better motion compensation results without significantly increasing bit costs. Observing hyperprior information (i.e., the mean and variance values) from the hyperprior networks contains discriminant statistical information of different patches, we also propose two efficient hyperprior-guided mode prediction methods. Specifically, using hyperprior information as the input, we propose two mode prediction networks to respectively predict the optimal block resolutions for better motion coding and decide whether to skip residual information from each block for better residual coding without introducing additional bit cost while bringing negligible extra computation cost. Comprehensive experimental results demonstrate our proposed C2F video compression framework equipped with the new hyperprior-guided mode prediction methods achieves the state-of-the-art performance on HEVC, UVG and MCL-JCV datasets.

Ruihao Gong, Yifu Ding, Zining Wang et al.

Large language models (LLMs) have achieved remarkable advancements in natural language processing, showcasing exceptional performance across various tasks. However, the expensive memory and computational requirements present significant challenges for their practical deployment. Low-bit quantization has emerged as a critical approach to mitigate these challenges by reducing the bit-width of model parameters, activations, and gradients, thus decreasing memory usage and computational demands. This paper presents a comprehensive survey of low-bit quantization methods tailored for LLMs, covering the fundamental principles, system implementations, and algorithmic strategies. An overview of basic concepts and new data formats specific to low-bit LLMs is first introduced, followed by a review of frameworks and systems that facilitate low-bit LLMs across various hardware platforms. Then, we categorize and analyze techniques and toolkits for efficient low-bit training and inference of LLMs. Finally, we conclude with a discussion of future trends and potential advancements of low-bit LLMs. Our systematic overview from basic, system, and algorithm perspectives can offer valuable insights and guidelines for future works to enhance the efficiency and applicability of LLMs through low-bit quantization.

Jiaheng Liu, Haoyu Qin, Yichao Wu et al.

Knowledge distillation is an effective method to improve the performance of a lightweight neural network (i.e., student model) by transferring the knowledge of a well-performed neural network (i.e., teacher model), which has been widely applied in many computer vision tasks, including face recognition. Nevertheless, the current face recognition distillation methods usually utilize the Feature Consistency Distillation (FCD) (e.g., L2 distance) on the learned embeddings extracted by the teacher and student models for each sample, which is not able to fully transfer the knowledge from the teacher to the student for face recognition. In this work, we observe that mutual relation knowledge between samples is also important to improve the discriminative ability of the learned representation of the student model, and propose an effective face recognition distillation method called CoupleFace by additionally introducing the Mutual Relation Distillation (MRD) into existing distillation framework. Specifically, in MRD, we first propose to mine the informative mutual relations, and then introduce the Relation-Aware Distillation (RAD) loss to transfer the mutual relation knowledge of the teacher model to the student model. Extensive experimental results on multiple benchmark datasets demonstrate the effectiveness of our proposed CoupleFace for face recognition. Moreover, based on our proposed CoupleFace, we have won the first place in the ICCV21 Masked Face Recognition Challenge (MS1M track).

Jiaheng Liu, Chenchen Zhang, Jinyang Guo et al.

Despite the advanced intelligence abilities of large language models (LLMs) in various applications, they still face significant computational and storage demands. Knowledge Distillation (KD) has emerged as an effective strategy to improve the performance of a smaller LLM (i.e., the student model) by transferring knowledge from a high-performing LLM (i.e., the teacher model). Prevailing techniques in LLM distillation typically use a black-box model API to generate high-quality pretrained and aligned datasets, or utilize white-box distillation by altering the loss function to better transfer knowledge from the teacher LLM. However, these methods ignore the knowledge differences between the student and teacher LLMs across domains. This results in excessive focus on domains with minimal performance gaps and insufficient attention to domains with large gaps, reducing overall performance. In this paper, we introduce a new LLM distillation framework called DDK, which dynamically adjusts the composition of the distillation dataset in a smooth manner according to the domain performance differences between the teacher and student models, making the distillation process more stable and effective. Extensive evaluations show that DDK significantly improves the performance of student models, outperforming both continuously pretrained baselines and existing knowledge distillation methods by a large margin.

Li Kang, Heng Zhou, Xiufeng Song et al.

Recent advancements in multimodal large language models and vision-languageaction models have significantly driven progress in Embodied AI. As the field transitions toward more complex task scenarios, multi-agent system frameworks are becoming essential for achieving scalable, efficient, and collaborative solutions. This shift is fueled by three primary factors: increasing agent capabilities, enhancing system efficiency through task delegation, and enabling advanced human-agent interactions. To address the challenges posed by multi-agent collaboration, we propose the Multi-Agent Robotic System (MARS) Challenge, held at the NeurIPS 2025 Workshop on SpaVLE. The competition focuses on two critical areas: planning and control, where participants explore multi-agent embodied planning using vision-language models (VLMs) to coordinate tasks and policy execution to perform robotic manipulation in dynamic environments. By evaluating solutions submitted by participants, the challenge provides valuable insights into the design and coordination of embodied multi-agent systems, contributing to the future development of advanced collaborative AI systems.

Shilong Tian, Hong Chen, Chengtao Lv et al.

Recently, video diffusion models (VDMs) have garnered significant attention due to their notable advancements in generating coherent and realistic video content. However, processing multiple frame features concurrently, coupled with the considerable model size, results in high latency and extensive memory consumption, hindering their broader application. Post-training quantization (PTQ) is an effective technique to reduce memory footprint and improve computational efficiency. Unlike image diffusion, we observe that the temporal features, which are integrated into all frame features, exhibit pronounced skewness. Furthermore, we investigate significant inter-channel disparities and asymmetries in the activation of video diffusion models, resulting in low coverage of quantization levels by individual channels and increasing the challenge of quantization. To address these issues, we introduce the first PTQ strategy tailored for video diffusion models, dubbed QVD. Specifically, we propose the High Temporal Discriminability Quantization (HTDQ) method, designed for temporal features, which retains the high discriminability of quantized features, providing precise temporal guidance for all video frames. In addition, we present the Scattered Channel Range Integration (SCRI) method which aims to improve the coverage of quantization levels across individual channels. Experimental validations across various models, datasets, and bit-width settings demonstrate the effectiveness of our QVD in terms of diverse metrics. In particular, we achieve near-lossless performance degradation on W8A8, outperforming the current methods by 205.12 in FVD.

Hong Chen, Chengtao Lv, Liang Ding et al.

Large language models (LLMs) have significantly advanced the field of natural language processing, while the expensive memory and computation consumption impede their practical deployment. Quantization emerges as one of the most effective methods for improving the computational efficiency of LLMs. However, existing ultra-low-bit quantization always causes severe accuracy drops. In this paper, we empirically relieve the micro and macro characteristics of ultra-low bit quantization and present a novel Dual-Binarization method for LLMs, namely DB-LLM. For the micro-level, we take both the accuracy advantage of 2-bit-width and the efficiency advantage of binarization into account, introducing Flexible Dual Binarization (FDB). By splitting 2-bit quantized weights into two independent sets of binaries, FDB ensures the accuracy of representations and introduces flexibility, utilizing the efficient bitwise operations of binarization while retaining the inherent high sparsity of ultra-low bit quantization. For the macro-level, we find the distortion that exists in the prediction of LLM after quantization, which is specified as the deviations related to the ambiguity of samples. We propose the Deviation-Aware Distillation (DAD) method, enabling the model to focus differently on various samples. Comprehensive experiments show that our DB-LLM not only significantly surpasses the current State-of-The-Art (SoTA) in ultra-low bit quantization (eg, perplexity decreased from 9.64 to 7.23), but also achieves an additional 20\% reduction in computational consumption compared to the SOTA method under the same bit-width. Our code will be released soon.

Haotong Qin, Xudong Ma, Xianglong Liu et al.

Low-bit quantization is widely used to compress super-resolution (SR) models and reduce storage and computation costs for deployment on resource-limited devices. However, when SR models are pushed to ultra-low precision (2-4 bits), performance can drop sharply due to diminished representational capacity and the detail-sensitive nature of SR. To address these issues, we propose QuantSR+, a unified framework that improves quantization operators, network design, and training optimization, achieving better trade-offs between accuracy and efficiency than prior low-bit SR methods. QuantSR+ mainly relies on three technical contributions: (1) Redistribution-driven Bit Determination (RBD), which reshapes quantization distributions in both forward and backward passes to preserve representation fidelity; (2) Quantized Slimmable Architecture (QSA), which begins with an over-parameterized model and progressively prunes less critical blocks to meet efficiency budgets while pushing the accuracy performance; and (3) Slimming-guided Function-localized Distillation (SFD), which enforces block-aware feature alignment via a direct loss and a progressive, function-local training schedule to capture quantization effects better and speed up convergence. Extensive experiments show that QuantSR+ achieves state-of-the-art performance against both specialized quantized SR methods and generic quantization approaches. For SwinIR-S on Urban100 (x4), it improves PSNR by 0.29 dB over the 2-bit SOTA baseline. Meanwhile, it delivers strong efficiency gains at 2-bit, reducing operations by up to 87.9% and storage by 89.4%. QuantSR+ is effective for both convolutional and transformer-based SR models, indicating broad applicability.

Xingyu Zheng, Yuye Li, Haoran Chu et al.

The Qwen series has emerged as a leading family of open-source Large Language Models (LLMs), demonstrating remarkable capabilities in natural language understanding tasks. With the recent release of Qwen3, which exhibits superior performance across diverse benchmarks, there is growing interest in deploying these models efficiently in resource-constrained environments. Low-bit quantization presents a promising solution, yet its impact on Qwen3's performance remains underexplored. This study conducts a systematic evaluation of Qwen3's robustness under various quantization settings, aiming to uncover both opportunities and challenges in compressing this state-of-the-art model. We rigorously assess 5 existing classic post-training quantization techniques applied to Qwen3, spanning bit-widths from 1 to 8 bits, and evaluate their effectiveness across multiple datasets. Our findings reveal that while Qwen3 maintains competitive performance at moderate bit-widths, it experiences notable degradation in linguistic tasks under ultra-low precision, underscoring the persistent hurdles in LLM compression. These results emphasize the need for further research to mitigate performance loss in extreme quantization scenarios. We anticipate that this empirical analysis will provide actionable insights for advancing quantization methods tailored to Qwen3 and future LLMs, ultimately enhancing their practicality without compromising accuracy. Our project is released on https://github.com/Efficient-ML/Qwen3-Quantization and https://huggingface.co/collections/Efficient-ML/qwen3-quantization-68164450decb1c868788cb2b.

Shukai Liu, Jian Yang, Bo Jiang et al.

Agents based on large language models have recently shown strong potential on real-world software engineering (SWE) tasks that require long-horizon interaction with repository-scale codebases. However, most existing agents rely on append-only context maintenance or passively triggered compression heuristics, which often lead to context explosion, semantic drift, and degraded reasoning in long-running interactions. We propose CAT, a new context management paradigm that elevates context maintenance to a callable tool integrated into the decision-making process of agents. CAT formalizes a structured context workspace consisting of stable task semantics, condensed long-term memory, and high-fidelity short-term interactions, and enables agents to proactively compress historical trajectories into actionable summaries at appropriate milestones. To support context management for SWE-agents, we propose a trajectory-level supervision framework, CAT-GENERATOR, based on an offline data construction pipeline that injects context-management actions into complete interaction trajectories. Using this framework, we train a context-aware model, SWE-Compressor. Experiments on SWE-Bench-Verified demonstrate that SWE-Compressor reaches a 57.6% solved rate and significantly outperforms ReAct-based agents and static compression baselines, while maintaining stable and scalable long-horizon reasoning under a bounded context budget.

Ge Yang, Changyi He, Jinyang Guo et al.

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. Our code is available at https://github.com/AboveParadise/LLMCBench.

Zining Wnag, Jinyang Guo, Ruihao Gong et al.

With the increased attention to model efficiency, post-training sparsity (PTS) has become more and more prevalent because of its effectiveness and efficiency. However, there remain questions on better practice of PTS algorithms and the sparsification ability of models, which hinders the further development of this area. Therefore, a benchmark to comprehensively investigate the issues above is urgently needed. In this paper, we propose the first comprehensive post-training sparsity benchmark called PTSBench towards algorithms and models. We benchmark 10+ PTS general-pluggable fine-grained techniques on 3 typical tasks using over 40 off-the-shelf model architectures. Through extensive experiments and analyses, we obtain valuable conclusions and provide several insights from both algorithms and model aspects. Our PTSBench can provide (1) new observations for a better understanding of the PTS algorithms, (2) in-depth and comprehensive evaluations for the sparsification ability of models, and (3) a well-structured and easy-integrate open-source framework. We hope this work will provide illuminating conclusions and advice for future studies of post-training sparsity methods and sparsification-friendly model design. The code for our PTSBench is released at \href{https://github.com/ModelTC/msbench}{https://github.com/ModelTC/msbench}.

Tianbo Wang, Yuqing Ma, Kewei Liao et al.

Large Vision-Language Models (LVLMs) have achieved substantial progress in cross-modal tasks. However, due to language bias, LVLMs are susceptible to object hallucination, which can be primarily divided into category, attribute, and relation hallucination, significantly impeding the trustworthy AI applications. Editing the internal activations of LVLMs has shown promising effectiveness in mitigating hallucinations with minimal cost. However, previous editing approaches neglect the effective guidance offered by factual textual semantics, thereby struggling to explicitly mitigate language bias. To address these issues, we propose Adaptive Factual-guided Visual-Textual Editing for hallucination mitigation (AFTER), which comprises Factual-Augmented Activation Steering (FAS) and Query-Adaptive Offset Optimization (QAO), to adaptively guides the original biased activations towards factual semantics. Specifically, FAS is proposed to provide factual and general guidance for activation editing, thereby explicitly modeling the precise visual-textual associations. Subsequently, QAO introduces a query-aware offset estimator to establish query-specific editing from the general steering vector, enhancing the diversity and granularity of editing. Extensive experiments on standard hallucination benchmarks across three widely adopted LVLMs validate the efficacy of the proposed AFTER, notably achieving up to a 16.3% reduction of hallucination over baseline on the AMBER benchmark. Our code and data will be released for reproducibility.

Xingyu Zheng, Xianglong Liu, Yichen Bian et al.

Diffusion models (DMs) have been significantly developed and widely used in various applications due to their excellent generative qualities. However, the expensive computation and massive parameters of DMs hinder their practical use in resource-constrained scenarios. As one of the effective compression approaches, quantization allows DMs to achieve storage saving and inference acceleration by reducing bit-width while maintaining generation performance. However, as the most extreme quantization form, 1-bit binarization causes the generation performance of DMs to face severe degradation or even collapse. This paper proposes a novel method, namely BiDM, for fully binarizing weights and activations of DMs, pushing quantization to the 1-bit limit. From a temporal perspective, we introduce the Timestep-friendly Binary Structure (TBS), which uses learnable activation binarizers and cross-timestep feature connections to address the highly timestep-correlated activation features of DMs. From a spatial perspective, we propose Space Patched Distillation (SPD) to address the difficulty of matching binary features during distillation, focusing on the spatial locality of image generation tasks and noise estimation networks. As the first work to fully binarize DMs, the W1A1 BiDM on the LDM-4 model for LSUN-Bedrooms 256$\times$256 achieves a remarkable FID of 22.74, significantly outperforming the current state-of-the-art general binarization methods with an FID of 59.44 and invalid generative samples, and achieves up to excellent 28.0 times storage and 52.7 times OPs savings. The code is available at https://github.com/Xingyu-Zheng/BiDM .

Xingyu Zheng, Xianglong Liu, Haotong Qin et al.

With the advancement of diffusion models (DMs) and the substantially increased computational requirements, quantization emerges as a practical solution to obtain compact and efficient low-bit DMs. However, the highly discrete representation leads to severe accuracy degradation, hindering the quantization of diffusion models to ultra-low bit-widths. This paper proposes a novel weight binarization approach for DMs, namely BinaryDM, pushing binarized DMs to be accurate and efficient by improving the representation and optimization. From the representation perspective, we present an Evolvable-Basis Binarizer (EBB) to enable a smooth evolution of DMs from full-precision to accurately binarized. EBB enhances information representation in the initial stage through the flexible combination of multiple binary bases and applies regularization to evolve into efficient single-basis binarization. The evolution only occurs in the head and tail of the DM architecture to retain the stability of training. From the optimization perspective, a Low-rank Representation Mimicking (LRM) is applied to assist the optimization of binarized DMs. The LRM mimics the representations of full-precision DMs in low-rank space, alleviating the direction ambiguity of the optimization process caused by fine-grained alignment. Comprehensive experiments demonstrate that BinaryDM achieves significant accuracy and efficiency gains compared to SOTA quantization methods of DMs under ultra-low bit-widths. With 1-bit weight and 4-bit activation (W1A4), BinaryDM achieves as low as 7.74 FID and saves the performance from collapse (baseline FID 10.87). As the first binarization method for diffusion models, W1A4 BinaryDM achieves impressive 15.2x OPs and 29.2x model size savings, showcasing its substantial potential for edge deployment. The code is available at https://github.com/Xingyu-Zheng/BinaryDM.

Zhuguanyu Wu, Jiayi Zhang, Jiaxin Chen et al.

Vision Transformers (ViTs) have become one of the most commonly used backbones for vision tasks. Despite their remarkable performance, they often suffer significant accuracy drops when quantized for practical deployment, particularly by post-training quantization (PTQ) under ultra-low bits. Recently, reconstruction-based PTQ methods have shown promising performance in quantizing Convolutional Neural Networks (CNNs). However, they fail when applied to ViTs, primarily due to the inaccurate estimation of output importance and the substantial accuracy degradation in quantizing post-GELU activations. To address these issues, we propose \textbf{APHQ-ViT}, a novel PTQ approach based on importance estimation with Average Perturbation Hessian (APH). Specifically, we first thoroughly analyze the current approximation approaches with Hessian loss, and propose an improved average perturbation Hessian loss. To deal with the quantization of the post-GELU activations, we design an MLP Reconstruction (MR) method by replacing the GELU function in MLP with ReLU and reconstructing it by the APH loss on a small unlabeled calibration set. Extensive experiments demonstrate that APHQ-ViT using linear quantizers outperforms existing PTQ methods by substantial margins in 3-bit and 4-bit across different vision tasks. The source code is available at https://github.com/GoatWu/APHQ-ViT.

Junzhu Mao, Yang Shen, Jinyang Guo et al.

Token compression is essential for reducing the computational and memory requirements of transformer models, enabling their deployment in resource-constrained environments. In this work, we propose an efficient and hardware-compatible token compression method called Prune and Merge. Our approach integrates token pruning and merging operations within transformer models to achieve layer-wise token compression. By introducing trainable merge and reconstruct matrices and utilizing shortcut connections, we efficiently merge tokens while preserving important information and enabling the restoration of pruned tokens. Additionally, we introduce a novel gradient-weighted attention scoring mechanism that computes token importance scores during the training phase, eliminating the need for separate computations during inference and enhancing compression efficiency. We also leverage gradient information to capture the global impact of tokens and automatically identify optimal compression structures. Extensive experiments on the ImageNet-1k and ADE20K datasets validate the effectiveness of our approach, achieving significant speed-ups with minimal accuracy degradation compared to state-of-the-art methods. For instance, on DeiT-Small, we achieve a 1.64$\times$ speed-up with only a 0.2\% drop in accuracy on ImageNet-1k. Moreover, by compressing segmenter models and comparing with existing methods, we demonstrate the superior performance of our approach in terms of efficiency and effectiveness. Code and models have been made available at https://github.com/NUST-Machine-Intelligence-Laboratory/prune_and_merge.

Yifu Ding, Xinhao Zhang, Jinyang Guo

Transformer-based large language models (LLMs) have demonstrated remarkable performance across a wide range of real-world tasks, but their inference cost remains prohibitively high due to the quadratic complexity of attention and the memory bandwidth limitations of high-precision operations. In this work, we present a low-bit mixed-precision attention kernel using the microscaling floating-point (MXFP) data format, utilizing the computing capability on next-generation GPU architectures. Our Diagonal-Tiled Mixed-Precision Attention (DMA) incorporates two kinds of low-bit computation at the tiling-level, and is a delicate fused kernel implemented using Triton, exploiting hardware-level parallelism and memory efficiency to enable fast and efficient inference without compromising model performance. Extensive empirical evaluations on NVIDIA B200 GPUs show that our kernel maintains generation quality with negligible degradation, and meanwhile achieves significant speedup by kernel fusion. We release our code at https://github.com/yifu-ding/MP-Sparse-Attn.

Kun Wang, Guibin Zhang, ManKit Ye et al.

The past two years have witnessed the meteoric rise of Large Language Model (LLM)-powered multi-agent systems (MAS), which harness collective intelligence and exhibit a remarkable trajectory toward self-evolution. This paradigm has rapidly progressed from manually engineered systems that require bespoke configuration of prompts, tools, roles, and communication protocols toward frameworks capable of automated orchestration. Yet, dominant automatic multi-agent systems, whether generated by external modules or a single LLM agent, largely adhere to a rigid ``\textit{generate-once-and-deploy}'' paradigm, rendering the resulting systems brittle and ill-prepared for the dynamism and uncertainty of real-world environments. To transcend this limitation, we introduce MAS$^2$, a paradigm predicated on the principle of recursive self-generation: a multi-agent system that autonomously architects bespoke multi-agent systems for diverse problems. Technically, we devise a ``\textit{generator-implementer-rectifier}'' tri-agent team capable of dynamically composing and adaptively rectifying a target agent system in response to real-time task demands. Collaborative Tree Optimization is proposed to train and specialize these meta-agents. Extensive evaluation across seven benchmarks reveals that MAS$^2$ achieves performance gains of up to $19.6\%$ over state-of-the-art MAS in complex scenarios such as deep research and code generation. Moreover, MAS$^2$ exhibits superior cross-backbone generalization, effectively leveraging previously unseen LLMs to yield improvements of up to $15.1\%$. Crucially, these gains are attained without incurring excessive token costs, as MAS$^2$ consistently resides on the Pareto frontier of cost-performance trade-offs. The source codes are available at https://github.com/yeyeyeah2/MAS2.

Chengtao Lv, Bilang Zhang, Yang Yong et al.

Large Vision-Language Models (VLMs) exhibit impressive multi-modal capabilities but suffer from prohibitive computational and memory demands, due to their long visual token sequences and massive parameter sizes. To address these issues, recent works have proposed training-free compression methods. However, existing efforts often suffer from three major limitations: (1) Current approaches do not decompose techniques into comparable modules, hindering fair evaluation across spatial and temporal redundancy. (2) Evaluation confined to simple single-turn tasks, failing to reflect performance in realistic scenarios. (3) Isolated use of individual compression techniques, without exploring their joint potential. To overcome these gaps, we introduce LLMC+, a comprehensive VLM compression benchmark with a versatile, plug-and-play toolkit. LLMC+ supports over 20 algorithms across five representative VLM families and enables systematic study of token-level and model-level compression. Our benchmark reveals that: (1) Spatial and temporal redundancies demand distinct technical strategies. (2) Token reduction methods degrade significantly in multi-turn dialogue and detail-sensitive tasks. (3) Combining token and model compression achieves extreme compression with minimal performance loss. We believe LLMC+ will facilitate fair evaluation and inspire future research in efficient VLM. Our code is available at https://github.com/ModelTC/LightCompress.

Ruihao Gong, Yang Yong, Zining Wang et al.

Neural network sparsity has attracted many research interests due to its similarity to biological schemes and high energy efficiency. However, existing methods depend on long-time training or fine-tuning, which prevents large-scale applications. Recently, some works focusing on post-training sparsity (PTS) have emerged. They get rid of the high training cost but usually suffer from distinct accuracy degradation due to neglect of the reasonable sparsity rate at each layer. Previous methods for finding sparsity rates mainly focus on the training-aware scenario, which usually fails to converge stably under the PTS setting with limited data and much less training cost. In this paper, we propose a fast and controllable post-training sparsity (FCPTS) framework. By incorporating a differentiable bridge function and a controllable optimization objective, our method allows for rapid and accurate sparsity allocation learning in minutes, with the added assurance of convergence to a predetermined global sparsity rate. Equipped with these techniques, we can surpass the state-of-the-art methods by a large margin, e.g., over 30\% improvement for ResNet-50 on ImageNet under the sparsity rate of 80\%. Our plug-and-play code and supplementary materials are open-sourced at https://github.com/ModelTC/FCPTS.

Chengtao Lv, Hong Chen, Jinyang Guo et al.

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 propose a post-training quantization (PTQ) framework for 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 experimental 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 lossless accuracy for instance segmentation, about 0.5\% drop with theoretical 3.9$\times$ acceleration. The code is available at \url{https://github.com/chengtao-lv/PTQ4SAM}.

Jinyang Guo, Weichen Zhang, Wanli Ouyang et al.

In this work, we propose a simple but effective channel pruning framework called Progressive Channel Pruning (PCP) to accelerate Convolutional Neural Networks (CNNs). In contrast to the existing channel pruning methods that prune channels only once per layer in a layer-by-layer fashion, our new progressive framework iteratively prunes a small number of channels from several selected layers, which consists of a three-step attempting-selecting-pruning pipeline in each iteration. In the attempting step, we attempt to prune a pre-defined number of channels from one layer by using any existing channel pruning methods and estimate the accuracy drop for this layer based on the labelled samples in the validation set. In the selecting step, based on the estimated accuracy drops for all layers, we propose a greedy strategy to automatically select a set of layers that will lead to less overall accuracy drop after pruning these layers. In the pruning step, we prune a small number of channels from these selected layers. We further extend our PCP framework to prune channels for the deep transfer learning methods like Domain Adversarial Neural Network (DANN), in which we effectively reduce the data distribution mismatch in the channel pruning process by using both labelled samples from the source domain and pseudo-labelled samples from the target domain. Our comprehensive experiments on two benchmark datasets demonstrate that our PCP framework outperforms the existing channel pruning approaches under both supervised learning and transfer learning settings.

Yifu Ding, Wentao Jiang, Shunyu Liu et al.

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 fine-grained cache management and alignment. Meanwhile, the Search and Transition Mechanism filters potential candidates to dynamically maintain the reasoning focus on more possible solutions and have less redundancy. Experiments on Qwen-2.5 and Llama-3 with Math500 and GSM8K datasets show that DPTS significantly improves efficiency by 2-4x on average while maintaining or even surpassing existing reasoning algorithms in accuracy, making ToT-based reasoning more scalable and computationally efficient.

Haocheng Huang, Jiaxin Chen, Jinyang Guo et al.

Diffusion models have achieved remarkable success in the image and video generation tasks. Nevertheless, they often require a large amount of memory and time overhead during inference, due to the complex network architecture and considerable number of timesteps for iterative diffusion. Recently, the post-training quantization (PTQ) technique has proved a promising way to reduce the inference cost by quantizing the float-point operations to low-bit ones. However, most of them fail to tackle with the large variations in the distribution of activations across distinct channels and timesteps, as well as the inconsistent of input between quantization and inference on diffusion models, thus leaving much room for improvement. To address the above issues, we propose a novel method dubbed Timestep-Channel Adaptive Quantization for Diffusion Models (TCAQ-DM). Specifically, we develop a timestep-channel joint reparameterization (TCR) module to balance the activation range along both the timesteps and channels, facilitating the successive reconstruction procedure. Subsequently, we employ a dynamically adaptive quantization (DAQ) module that mitigate the quantization error by selecting an optimal quantizer for each post-Softmax layers according to their specific types of distributions. Moreover, we present a progressively aligned reconstruction (PAR) strategy to mitigate the bias caused by the input mismatch. Extensive experiments on various benchmarks and distinct diffusion models demonstrate that the proposed method substantially outperforms the state-of-the-art approaches in most cases, especially yielding comparable FID metrics to the full precision model on CIFAR-10 in the W6A6 setting, while enabling generating available images in the W4A4 settings.

Yifu Ding, Xianglong Liu, Shenghao Jin et al.

Ultra low-bit quantization brings substantial efficiency for Transformer-based models, but the accuracy degradation and limited GPU support hinder its wide usage. In this paper, we analyze zero-point distortion in binarization and propose a Binary Weights & Ternary Activations (BWTA) quantization scheme, which projects tiny values to zero and preserves the accuracy of extremely low-bit models. For training, we propose Smooth Multi-Stage Quantization, combining a Levelwise Degradation Strategy and a Magnitude-Alignment Projection Factor to enable stable and fast convergence. For inference, we develop a BWTA MatMul CUDA kernel with instruction-level parallel bit-packing and comprehensive binary/ternary MatMul implementations for both linear and attention operators, allowing seamless integration across Transformer architectures. Experiments show that BWTA approaches full-precision performance for BERT, with an average 3.5% drop on GLUE and less than 2% drop on five tasks, and achieves comparable perplexity and accuracy for LLMs. In efficiency, it delivers 16 to 24 times kernel-level speedup over FP16 on NVIDIA GPUs, and 216 to 330 tokens/s end-to-end prefill speedup with lower memory footprint on LLMs. As an algorithm-hardware co-design, BWTA demonstrates practical, low-latency ultra-low-bit inference without sacrificing model quality.

Junli Deng, Ping Shi, Qipei Li et al.

Reconstructing intricate, ever-changing environments remains a central ambition in computer vision, yet existing solutions often crumble before the complexity of real-world dynamics. We present DynaSplat, an approach that extends Gaussian Splatting to dynamic scenes by integrating dynamic-static separation and hierarchical motion modeling. First, we classify scene elements as static or dynamic through a novel fusion of deformation offset statistics and 2D motion flow consistency, refining our spatial representation to focus precisely where motion matters. We then introduce a hierarchical motion modeling strategy that captures both coarse global transformations and fine-grained local movements, enabling accurate handling of intricate, non-rigid motions. Finally, we integrate physically-based opacity estimation to ensure visually coherent reconstructions, even under challenging occlusions and perspective shifts. Extensive experiments on challenging datasets reveal that DynaSplat not only surpasses state-of-the-art alternatives in accuracy and realism but also provides a more intuitive, compact, and efficient route to dynamic scene reconstruction.

Yushi Huang, Zining Wang, Zhihang Yuan et al.

Mixture-of-Experts (MoE) Multimodal large language models (MLLMs) excel at vision-language tasks, but they suffer from high computational inefficiency. To reduce inference overhead, expert skipping methods have been proposed to deactivate redundant experts based on the current input tokens. However, we find that applying these methods-originally designed for unimodal large language models (LLMs)-to MLLMs results in considerable performance degradation. This is primarily because such methods fail to account for the heterogeneous contributions of experts across MoE layers and modality-specific behaviors of tokens within these layers. Motivated by these findings, we propose MoDES, the first training-free framework that adaptively skips experts to enable efficient and accurate MoE MLLM inference. It incorporates a globally-modulated local gating (GMLG) mechanism that integrates global layer-wise importance into local routing probabilities to accurately estimate per-token expert importance. A dual-modality thresholding (DMT) method is then applied, which processes tokens from each modality separately, to derive the skipping schedule. To set the optimal thresholds, we introduce a frontier search algorithm that exploits monotonicity properties, cutting convergence time from several days to a few hours. Extensive experiments for 3 model series across 13 benchmarks demonstrate that MoDES far outperforms previous approaches. For instance, when skipping 88% experts for Qwen3-VL-MoE-30B-A3B-Instruct, the performance boost is up to 10.67% (97.33% vs. 86.66%). Furthermore, MoDES significantly enhances inference speed, improving the prefilling time by 2.16$\times$ and the decoding time by 1.26$\times$.

Jiacheng Wang, Yejun Zeng, Jinyang Guo et al.

Despite the growing interest in Small Language Models (SLMs) as resource-efficient alternatives to Large Language Models (LLMs), their deployment on edge devices remains challenging due to unresolved efficiency gaps in model compression. While quantization has proven effective for LLMs, its applicability to SLMs is significantly underexplored, with critical questions about differing quantization bottlenecks and efficiency profiles. This paper introduces SLMQuant, the first systematic benchmark for evaluating LLM compression techniques when applied to SLMs. Through comprehensive multi-track evaluations across diverse architectures and tasks, we analyze how state-of-the-art quantization methods perform on SLMs. Our findings reveal fundamental disparities between SLMs and LLMs in quantization sensitivity, demonstrating that direct transfer of LLM-optimized techniques leads to suboptimal results due to SLMs' unique architectural characteristics and training dynamics. We identify key factors governing effective SLM quantization and propose actionable design principles for SLM-tailored compression. SLMQuant establishes a foundational framework for advancing efficient SLM deployment on low-end devices in edge applications, and provides critical insights for deploying lightweight language models in resource-constrained scenarios.

Xingyu Zheng, Haotong Qin, Yuye Li et al.

Post-training quantization (PTQ) offers an efficient approach to compressing large language models (LLMs), significantly reducing memory access and computational costs. Existing compensation-based weight calibration methods often rely on a second-order Taylor expansion to model quantization error, under the assumption that the first-order term is negligible in well-trained full-precision models. However, we reveal that the progressive compensation process introduces accumulated first-order deviations between latent weights and their full-precision counterparts, making this assumption fundamentally flawed. To address this, we propose FOEM, a novel PTQ method that explicitly incorporates first-order gradient terms to improve quantization error compensation. FOEM approximates gradients by performing a first-order Taylor expansion around the pre-quantization weights. This yields an approximation based on the difference between latent and full-precision weights as well as the Hessian matrix. When substituted into the theoretical solution, the formulation eliminates the need to explicitly compute the Hessian, thereby avoiding the high computational cost and limited generalization of backpropagation-based gradient methods. This design introduces only minimal additional computational overhead. Extensive experiments across a wide range of models and benchmarks demonstrate that FOEM consistently outperforms the classical GPTQ method. In 3-bit weight-only quantization, FOEM reduces the perplexity of Llama3-8B by 17.3% and increases the 5-shot MMLU accuracy from 53.8% achieved by GPTAQ to 56.1%. Moreover, FOEM can be seamlessly combined with advanced techniques such as SpinQuant, delivering additional gains under the challenging W4A4KV4 setting and further narrowing the performance gap with full-precision baselines, surpassing existing state-of-the-art methods.

Hui Xie, Yuhe Liu, Shaoqi Yang et al.

While deep spiking neural networks (SNNs) demonstrate superior performance, their deployment on resource-constrained neuromorphic hardware still remains challenging. Network pruning offers a viable solution by reducing both parameters and synaptic operations (SynOps) to facilitate the edge deployment of SNNs, among which search-based pruning methods search for the SNNs structure after pruning. However, existing search-based methods fail to directly use SynOps as the constraint because it will dynamically change in the searching process, resulting in the final searched network violating the expected SynOps target. In this paper, we introduce a novel SNN pruning framework called SPEAR, which leverages reinforcement learning (RL) technique to directly use SynOps as the searching constraint. To avoid the violation of SynOps requirements, we first propose a SynOps prediction mechanism called LRE to accurately predict the final SynOps after search. Observing SynOps cannot be explicitly calculated and added to constrain the action in RL, we propose a novel reward called TAR to stabilize the searching. Extensive experiments show that our SPEAR framework can effectively compress SNN under specific SynOps constraint.

Jinyang Guo, Dong Xu, Guo Lu

In this paper, we propose a new deep image compression framework called Complexity and Bitrate Adaptive Network (CBANet), which aims to learn one single network to support variable bitrate coding under different computational complexity constraints. In contrast to the existing state-of-the-art learning based image compression frameworks that only consider the rate-distortion trade-off without introducing any constraint related to the computational complexity, our CBANet considers the trade-off between the rate and distortion under dynamic computational complexity constraints. Specifically, to decode the images with one single decoder under various computational complexity constraints, we propose a new multi-branch complexity adaptive module, in which each branch only takes a small portion of the computational budget of the decoder. The reconstructed images with different visual qualities can be readily generated by using different numbers of branches. Furthermore, to achieve variable bitrate decoding with one single decoder, we propose a bitrate adaptive module to project the representation from a base bitrate to the expected representation at a target bitrate for transmission. Then it will project the transmitted representation at the target bitrate back to that at the base bitrate for the decoding process. The proposed bit adaptive module can significantly reduce the storage requirement for deployment platforms. As a result, our CBANet enables one single codec to support multiple bitrate decoding under various computational complexity constraints. Comprehensive experiments on two benchmark datasets demonstrate the effectiveness of our CBANet for deep image compression.

Jinyang Guo, Wanli Ouyang, Dong Xu

In this work, we propose a new layer-by-layer channel pruning method called Channel Pruning guided by classification Loss and feature Importance (CPLI). In contrast to the existing layer-by-layer channel pruning approaches that only consider how to reconstruct the features from the next layer, our approach additionally take the classification loss into account in the channel pruning process. We also observe that some reconstructed features will be removed at the next pruning stage. So it is unnecessary to reconstruct these features. To this end, we propose a new strategy to suppress the influence of unimportant features (i.e., the features will be removed at the next pruning stage). Our comprehensive experiments on three benchmark datasets, i.e., CIFAR-10, ImageNet, and UCF-101, demonstrate the effectiveness of our CPLI method.