Geng Yuan

Zhengang Li, Mengshu Sun, Alec Lu et al. · meta-ai

Vision transformers (ViTs) are emerging with significantly improved accuracy in computer vision tasks. However, their complex architecture and enormous computation/storage demand impose urgent needs for new hardware accelerator design methodology. This work proposes an FPGA-aware automatic ViT acceleration framework based on the proposed mixed-scheme quantization. To the best of our knowledge, this is the first FPGA-based ViT acceleration framework exploring model quantization. Compared with state-of-the-art ViT quantization work (algorithmic approach only without hardware acceleration), our quantization achieves 0.47% to 1.36% higher Top-1 accuracy under the same bit-width. Compared with the 32-bit floating-point baseline FPGA accelerator, our accelerator achieves around 5.6x improvement on the frame rate (i.e., 56.8 FPS vs. 10.0 FPS) with 0.71% accuracy drop on ImageNet dataset for DeiT-base.

Zhenglun Kong, Haoyu Ma, Geng Yuan et al. · harvard, meta-ai

Vision transformers (ViTs) have recently obtained success in many applications, but their intensive computation and heavy memory usage at both training and inference time limit their generalization. Previous compression algorithms usually start from the pre-trained dense models and only focus on efficient inference, while time-consuming training is still unavoidable. In contrast, this paper points out that the million-scale training data is redundant, which is the fundamental reason for the tedious training. To address the issue, this paper aims to introduce sparsity into data and proposes an end-to-end efficient training framework from three sparse perspectives, dubbed Tri-Level E-ViT. Specifically, we leverage a hierarchical data redundancy reduction scheme, by exploring the sparsity under three levels: number of training examples in the dataset, number of patches (tokens) in each example, and number of connections between tokens that lie in attention weights. With extensive experiments, we demonstrate that our proposed technique can noticeably accelerate training for various ViT architectures while maintaining accuracy. Remarkably, under certain ratios, we are able to improve the ViT accuracy rather than compromising it. For example, we can achieve 15.2% speedup with 72.6% (+0.4) Top-1 accuracy on Deit-T, and 15.7% speedup with 79.9% (+0.1) Top-1 accuracy on Deit-S. This proves the existence of data redundancy in ViT.

Xuan Shen, Zhenglun Kong, Minghai Qin et al. · harvard

The conventional lottery ticket hypothesis (LTH) claims that there exists a sparse subnetwork within a dense neural network and a proper random initialization method called the winning ticket, such that it can be trained from scratch to almost as good as the dense counterpart. Meanwhile, the research of LTH in vision transformers (ViTs) is scarcely evaluated. In this paper, we first show that the conventional winning ticket is hard to find at the weight level of ViTs by existing methods. Then, we generalize the LTH for ViTs to input data consisting of image patches inspired by the input dependence of ViTs. That is, there exists a subset of input image patches such that a ViT can be trained from scratch by using only this subset of patches and achieve similar accuracy to the ViTs trained by using all image patches. We call this subset of input patches the em winning tickets, which represent a significant amount of information in the input data. We use a ticket selector to generate the winning tickets based on the informativeness of patches for various types of ViT, including DeiT, LV-ViT, and Swin Transformers. The experiments show that there is a clear difference between the performance of models trained with winning tickets and randomly selected subsets, which verifies our proposed theory. We elaborate on the analogical similarity between our proposed Data-LTH-ViTs and the conventional LTH to further verify the integrity of our theory. The Source codes are available at https://github.com/shawnricecake/vit-lottery-ticket-input.

Sizhe Chen, Geng Yuan, Xinwen Cheng et al.

As data becomes increasingly vital, a company would be very cautious about releasing data, because the competitors could use it to train high-performance models, thereby posing a tremendous threat to the company's commercial competence. To prevent training good models on the data, we could add imperceptible perturbations to it. Since such perturbations aim at hurting the entire training process, they should reflect the vulnerability of DNN training, rather than that of a single model. Based on this new idea, we seek perturbed examples that are always unrecognized (never correctly classified) in training. In this paper, we uncover them by model checkpoints' gradients, forming the proposed self-ensemble protection (SEP), which is very effective because (1) learning on examples ignored during normal training tends to yield DNNs ignoring normal examples; (2) checkpoints' cross-model gradients are close to orthogonal, meaning that they are as diverse as DNNs with different architectures. That is, our amazing performance of ensemble only requires the computation of training one model. By extensive experiments with 9 baselines on 3 datasets and 5 architectures, SEP is verified to be a new state-of-the-art, e.g., our small $\ell_\infty=2/255$ perturbations reduce the accuracy of a CIFAR-10 ResNet18 from 94.56% to 14.68%, compared to 41.35% by the best-known method. Code is available at https://github.com/Sizhe-Chen/SEP.

Yanyu Li, Geng Yuan, Yang Wen et al.

Vision Transformers (ViT) have shown rapid progress in computer vision tasks, achieving promising results on various benchmarks. However, due to the massive number of parameters and model design, \textit{e.g.}, attention mechanism, ViT-based models are generally times slower than lightweight convolutional networks. Therefore, the deployment of ViT for real-time applications is particularly challenging, especially on resource-constrained hardware such as mobile devices. Recent efforts try to reduce the computation complexity of ViT through network architecture search or hybrid design with MobileNet block, yet the inference speed is still unsatisfactory. This leads to an important question: can transformers run as fast as MobileNet while obtaining high performance? To answer this, we first revisit the network architecture and operators used in ViT-based models and identify inefficient designs. Then we introduce a dimension-consistent pure transformer (without MobileNet blocks) as a design paradigm. Finally, we perform latency-driven slimming to get a series of final models dubbed EfficientFormer. Extensive experiments show the superiority of EfficientFormer in performance and speed on mobile devices. Our fastest model, EfficientFormer-L1, achieves $79.2\%$ top-1 accuracy on ImageNet-1K with only $1.6$ ms inference latency on iPhone 12 (compiled with CoreML), which runs as fast as MobileNetV2$\times 1.4$ ($1.6$ ms, $74.7\%$ top-1), and our largest model, EfficientFormer-L7, obtains $83.3\%$ accuracy with only $7.0$ ms latency. Our work proves that properly designed transformers can reach extremely low latency on mobile devices while maintaining high performance.

Geng Yuan, Yanyu Li, Sheng Li et al. · harvard

Recently, sparse training has emerged as a promising paradigm for efficient deep learning on edge devices. The current research mainly devotes efforts to reducing training costs by further increasing model sparsity. However, increasing sparsity is not always ideal since it will inevitably introduce severe accuracy degradation at an extremely high sparsity level. This paper intends to explore other possible directions to effectively and efficiently reduce sparse training costs while preserving accuracy. To this end, we investigate two techniques, namely, layer freezing and data sieving. First, the layer freezing approach has shown its success in dense model training and fine-tuning, yet it has never been adopted in the sparse training domain. Nevertheless, the unique characteristics of sparse training may hinder the incorporation of layer freezing techniques. Therefore, we analyze the feasibility and potentiality of using the layer freezing technique in sparse training and find it has the potential to save considerable training costs. Second, we propose a data sieving method for dataset-efficient training, which further reduces training costs by ensuring only a partial dataset is used throughout the entire training process. We show that both techniques can be well incorporated into the sparse training algorithm to form a generic framework, which we dub SpFDE. Our extensive experiments demonstrate that SpFDE can significantly reduce training costs while preserving accuracy from three dimensions: weight sparsity, layer freezing, and dataset sieving.

Kaiyuan Deng, Bo Hui, Gen Li et al.

The widespread adoption of text-to-image (T2I) diffusion models has raised concerns about their potential to generate copyrighted, inappropriate, or sensitive imagery. As a practical solution, machine unlearning aims to erase unwanted concepts without retraining from scratch. While most existing methods are effective for single-concept unlearning, they often struggle when removing multiple concepts, causing significant challenges in unlearning effectiveness, generation quality, and sensitivity to hyperparameters and datasets. We take a unique perspective on multi-concept unlearning by leveraging model sparsity and propose the Forget It All (FIA) framework. FIA first introduces Contrastive Concept Saliency to quantify each weight connection's contribution to a target concept. It then identifies Concept Sensitive Neurons by combining temporal and spatial information, ensuring that only neurons consistently responsive to the target concept are selected. Finally, FIA constructs masks from the identified neurons and fuses them into a unified multi-concept mask, where Concept Agnostic Neurons that broadly support general content generation are preserved while concept-specific neurons are pruned to remove the targets. FIA is training-free and requires minimal hyperparameter tuning for new tasks, enabling plug-and-play use. Extensive experiments across three distinct unlearning tasks demonstrate that FIA achieves more reliable multi-concept unlearning, improving forgetting effectiveness while maintaining generation fidelity and quality. Code is available at https://github.com/kaiyuan02415/Forget-It-All

Arman Akbari, Ci Zhang, Arash Akbari et al.

World-action models (WAMs) jointly generate future video and robot actions through iterative diffusion, achieving strong performance on manipulation benchmarks but requiring tens of denoising steps, a cost that precludes real-time control. Step distillation has emerged as the natural remedy, but off-the-shelf methods break down in the joint video-action setting because video and action streams use different SNR-shifted noise schedules and reach training with substantially different marginal noise distributions, an asymmetry that single-modality distillation methods cannot accommodate. We introduce \textbf{Flash-WAM}, a modality-aware step-distillation framework inspired by consistency distillation that selects the consistency function for each modality to match its noise regime: a linear-gradient-scaling parametrization for the action stream's low-noise regime, paired with a variance-preserving parametrization for the video stream's high-noise regime, grounded in a structural analysis of the consistency-function family that characterizes the achievable gradient scaling under the consistency boundary condition. Instantiated on LingBot-VA, Flash-WAM compresses inference to a single step in each modality. On RoboTwin 2.0, this reduces per-chunk latency from $8.1$ seconds to $348$ ms on NVIDIA L40S, a $23{\times}$ speedup that enables real-time inference. Flash-WAM preserves task success on simulation benchmarks ($85.5\%$ RoboTwin 2.0, $95.7\%$ LIBERO) and substantially recovers real-world performance ($60\%$ average on a Unitree G1 humanoid robot), while naive consistency distillation drops to $24\%$ at the same step budget.

Fei Dou, Jin Ye, Geng Yuan et al.

Artificial General Intelligence (AGI), possessing the capacity to comprehend, learn, and execute tasks with human cognitive abilities, engenders significant anticipation and intrigue across scientific, commercial, and societal arenas. This fascination extends particularly to the Internet of Things (IoT), a landscape characterized by the interconnection of countless devices, sensors, and systems, collectively gathering and sharing data to enable intelligent decision-making and automation. This research embarks on an exploration of the opportunities and challenges towards achieving AGI in the context of the IoT. Specifically, it starts by outlining the fundamental principles of IoT and the critical role of Artificial Intelligence (AI) in IoT systems. Subsequently, it delves into AGI fundamentals, culminating in the formulation of a conceptual framework for AGI's seamless integration within IoT. The application spectrum for AGI-infused IoT is broad, encompassing domains ranging from smart grids, residential environments, manufacturing, and transportation to environmental monitoring, agriculture, healthcare, and education. However, adapting AGI to resource-constrained IoT settings necessitates dedicated research efforts. Furthermore, the paper addresses constraints imposed by limited computing resources, intricacies associated with large-scale IoT communication, as well as the critical concerns pertaining to security and privacy.

Zifeng Wang, Zheng Zhan, Yifan Gong et al.

Existing work in continual learning (CL) focuses on mitigating catastrophic forgetting, i.e., model performance deterioration on past tasks when learning a new task. However, the training efficiency of a CL system is under-investigated, which limits the real-world application of CL systems under resource-limited scenarios. In this work, we propose a novel framework called Sparse Continual Learning(SparCL), which is the first study that leverages sparsity to enable cost-effective continual learning on edge devices. SparCL achieves both training acceleration and accuracy preservation through the synergy of three aspects: weight sparsity, data efficiency, and gradient sparsity. Specifically, we propose task-aware dynamic masking (TDM) to learn a sparse network throughout the entire CL process, dynamic data removal (DDR) to remove less informative training data, and dynamic gradient masking (DGM) to sparsify the gradient updates. Each of them not only improves efficiency, but also further mitigates catastrophic forgetting. SparCL consistently improves the training efficiency of existing state-of-the-art (SOTA) CL methods by at most 23X less training FLOPs, and, surprisingly, further improves the SOTA accuracy by at most 1.7%. SparCL also outperforms competitive baselines obtained from adapting SOTA sparse training methods to the CL setting in both efficiency and accuracy. We also evaluate the effectiveness of SparCL on a real mobile phone, further indicating the practical potential of our method.

Yanyu Li, Pu Zhao, Geng Yuan et al.

Neural architecture search (NAS) and network pruning are widely studied efficient AI techniques, but not yet perfect. NAS performs exhaustive candidate architecture search, incurring tremendous search cost. Though (structured) pruning can simply shrink model dimension, it remains unclear how to decide the per-layer sparsity automatically and optimally. In this work, we revisit the problem of layer-width optimization and propose Pruning-as-Search (PaS), an end-to-end channel pruning method to search out desired sub-network automatically and efficiently. Specifically, we add a depth-wise binary convolution to learn pruning policies directly through gradient descent. By combining the structural reparameterization and PaS, we successfully searched out a new family of VGG-like and lightweight networks, which enable the flexibility of arbitrary width with respect to each layer instead of each stage. Experimental results show that our proposed architecture outperforms prior arts by around $1.0\%$ top-1 accuracy under similar inference speed on ImageNet-1000 classification task. Furthermore, we demonstrate the effectiveness of our width search on complex tasks including instance segmentation and image translation. Code and models are released.

Jun Liu, Pu Zhao, Zhenglun Kong et al. · harvard

Embodied robotic systems increasingly rely on large language model (LLM)-based agents to support high-level reasoning, planning, and decision-making during interactions with the environment. However, invoking LLM reasoning introduces substantial computational latency and resource overhead, which can interrupt action execution and reduce system reliability. Excessive reasoning may delay actions, while insufficient reasoning often leads to incorrect decisions and task failures. This raises a fundamental question for embodied agents: when should the agent reason, and when should it act? In this work, we propose RARRL (Resource-Aware Reasoning via Reinforcement Learning), a hierarchical framework for resource-aware orchestration of embodied agents. Rather than learning low-level control policies, RARRL learns a high-level orchestration policy that operates at the agent's decision-making layer. This policy enables the agent to adaptively determine whether to invoke reasoning, which reasoning role to employ, and how much computational budget to allocate based on current observations, execution history, and remaining resources. Extensive experiments, including evaluations with empirical latency profiles derived from the ALFRED benchmark, show that RARRL consistently improves task success rates while reducing execution latency and enhancing robustness compared with fixed or heuristic reasoning strategies. These results demonstrate that adaptive reasoning control is essential for building reliable and efficient embodied robotic agents.

Zhengang Li, Geng Yuan, Tomoharu Yamauchi et al.

Adiabatic Quantum-Flux-Parametron (AQFP) is a superconducting logic with extremely high energy efficiency. By employing the distinct polarity of current to denote logic `0' and `1', AQFP devices serve as excellent carriers for binary neural network (BNN) computations. Although recent research has made initial strides toward developing an AQFP-based BNN accelerator, several critical challenges remain, preventing the design from being a comprehensive solution. In this paper, we propose SupeRBNN, an AQFP-based randomized BNN acceleration framework that leverages software-hardware co-optimization to eventually make the AQFP devices a feasible solution for BNN acceleration. Specifically, we investigate the randomized behavior of the AQFP devices and analyze the impact of crossbar size on current attenuation, subsequently formulating the current amplitude into the values suitable for use in BNN computation. To tackle the accumulation problem and improve overall hardware performance, we propose a stochastic computing-based accumulation module and a clocking scheme adjustment-based circuit optimization method. We validate our SupeRBNN framework across various datasets and network architectures, comparing it with implementations based on different technologies, including CMOS, ReRAM, and superconducting RSFQ/ERSFQ. Experimental results demonstrate that our design achieves an energy efficiency of approximately 7.8x10^4 times higher than that of the ReRAM-based BNN framework while maintaining a similar level of model accuracy. Furthermore, when compared with superconductor-based counterparts, our framework demonstrates at least two orders of magnitude higher energy efficiency.

Timothy Rupprecht, Pu Zhao, Amir Taherin et al.

This comprehensive report distinguishes prior works by the cognitive functions they innovate. Many works claim an almost "human-like" cognitive capability in their world models. To evaluate these claims requires a proper grounding in first principles in Cognitive Architecture Theory (CAT). We present a conceptual unified framework for world models that fully incorporates all the cognitive functions associated with CAT (i.e. memory, perception, language, reasoning, imagining, motivation, and meta-cognition) and identify gaps in the research as a guide for future states of the art. In particular, we find that motivation (especially intrinsic motivation) and meta-cognition remain drastically under-researched, and we propose concrete directions informed by active inference and global workspace theory to address them. We further introduce Epistemic World Models, a new category encompassing agent frameworks for scientific discovery that operate over structured knowledge. Our taxonomy, applied across video, embodied, and epistemic world models, suggests research directions where prior taxonomies have not.

Huayu Li, Ana S. Carreon-Rascon, Xiwen Chen et al.

Medical time series data are indispensable in healthcare, providing critical insights for disease diagnosis, treatment planning, and patient management. The exponential growth in data complexity, driven by advanced sensor technologies, has presented challenges related to data labeling. Self-supervised learning (SSL) has emerged as a transformative approach to address these challenges, eliminating the need for extensive human annotation. In this study, we introduce a novel framework for Medical Time Series Representation Learning, known as MTS-LOF. MTS-LOF leverages the strengths of contrastive learning and Masked Autoencoder (MAE) methods, offering a unique approach to representation learning for medical time series data. By combining these techniques, MTS-LOF enhances the potential of healthcare applications by providing more sophisticated, context-rich representations. Additionally, MTS-LOF employs a multi-masking strategy to facilitate occlusion-invariant feature learning. This approach allows the model to create multiple views of the data by masking portions of it. By minimizing the discrepancy between the representations of these masked patches and the fully visible patches, MTS-LOF learns to capture rich contextual information within medical time series datasets. The results of experiments conducted on diverse medical time series datasets demonstrate the superiority of MTS-LOF over other methods. These findings hold promise for significantly enhancing healthcare applications by improving representation learning. Furthermore, our work delves into the integration of joint-embedding SSL and MAE techniques, shedding light on the intricate interplay between temporal and structural dependencies in healthcare data. This understanding is crucial, as it allows us to grasp the complexities of healthcare data analysis.

Chao Wu, Yifan Gong, Liangkai Liu et al.

Object detection on the edge (Edge-OD) is in growing demand thanks to its ever-broad application prospects. However, the development of this field is rigorously restricted by the deployment dilemma of simultaneously achieving high accuracy, excellent power efficiency, and meeting strict real-time requirements. To tackle this dilemma, we propose AyE-Edge, the first-of-this-kind development tool that explores automated algorithm-device deployment space search to realize Accurate yet power-Efficient real-time object detection on the Edge. Through a collaborative exploration of keyframe selection, CPU-GPU configuration, and DNN pruning strategy, AyE-Edge excels in extensive real-world experiments conducted on a mobile device. The results consistently demonstrate AyE-Edge's effectiveness, realizing outstanding real-time performance, detection accuracy, and notably, a remarkable 96.7% reduction in power consumption, compared to state-of-the-art (SOTA) competitors.

Yanyu Li, Xuan Shen, Geng Yuan et al.

In this work we demonstrate real-time portrait stylization, specifically, translating self-portrait into cartoon or anime style on mobile devices. We propose a latency-driven differentiable architecture search method, maintaining realistic generative quality. With our framework, we obtain $10\times$ computation reduction on the generative model and achieve real-time video stylization on off-the-shelf smartphone using mobile GPUs.

Qitao Tan, Xiaoying Song, Arman Akbari et al.

Current safety alignment of foundation models largely follows a \emph{one-size-fits-all} paradigm, applying the same refusal policy across users and contexts. As a result, models may refuse requests that are unsafe for general users but legitimate for authorized professionals, limiting helpfulness in specialized professional settings. Existing approaches either require costly realignment or rely on inference-time steering that suffers from imprecise control and added latency. To this end, we propose \textsc{Palette}, a modular, controllable, and efficient framework that selectively relaxes refusal behavior on authorized target domains while preserving standard safety elsewhere. Our method identifies a refusal direction via multi-objective search and internalizes it into the model through lightweight adaptation. \textsc{Palette} further supports modular composition: it learns domain-specific safety controls independently and composes them through parameter merging, enabling on-demand multi-domain authorization without retraining. Experiments across four safety benchmarks, multiple model variants, and both LLMs and VLMs show that \textsc{Palette} delivers precise safety control without sacrificing general utility, offering a practical path toward foundation models that adapt to diverse professional needs.

Jun Liu, Geng Yuan, Weihao Zeng et al.

In research findings, co-deletion of the 1p/19q gene is associated with clinical outcomes in low-grade gliomas. The ability to predict 1p19q status is critical for treatment planning and patient follow-up. This study aims to utilize a specially MRI-based convolutional neural network for brain cancer detection. Although public networks such as RestNet and AlexNet can effectively diagnose brain cancers using transfer learning, the model includes quite a few weights that have nothing to do with medical images. As a result, the diagnostic results are unreliable by the transfer learning model. To deal with the problem of trustworthiness, we create the model from the ground up, rather than depending on a pre-trained model. To enable flexibility, we combined convolution stacking with a dropout and full connect operation, it improved performance by reducing overfitting. During model training, we also supplement the given dataset and inject Gaussian noise. We use three--fold cross-validation to train the best selection model. Comparing InceptionV3, VGG16, and MobileNetV2 fine-tuned with pre-trained models, our model produces better results. On an validation set of 125 codeletion vs. 31 not codeletion images, the proposed network achieves 96.37\% percent F1-score, 97.46\% percent precision, and 96.34\% percent recall when classifying 1p/19q codeletion and not codeletion images.

Qitao Tan, Jun Liu, Zheng Zhan et al.

Large language models (LLMs) excel across various tasks, but standard first-order (FO) fine-tuning demands considerable memory, significantly limiting real-world deployment. Recently, zeroth-order (ZO) optimization stood out as a promising memory-efficient training paradigm, avoiding backward passes and relying solely on forward passes for gradient estimation, making it attractive for resource-constrained scenarios. However, ZO method lags far behind FO method in both convergence speed and accuracy. To bridge the gap, we introduce a novel layer-wise divergence analysis that uncovers the distinct update pattern of FO and ZO optimization. Aiming to resemble the learning capacity of FO method from the findings, we propose Divergence-driven Zeroth-Order (DiZO) optimization. DiZO conducts divergence-driven layer adaptation by incorporating projections to ZO updates, generating diverse-magnitude updates precisely scaled to layer-wise individual optimization needs. Our results demonstrate that DiZO significantly reduces the needed iterations for convergence without sacrificing throughput, cutting training GPU hours by up to 48\% on various datasets. Moreover, DiZO consistently outperforms the representative ZO baselines in fine-tuning RoBERTa-large, OPT-series, and Llama-series on downstream tasks and, in some cases, even surpasses memory-intensive FO fine-tuning. Our code is released at https://github.com/Skilteee/DiZO.

Pu Zhao, Juyi Lin, Timothy Rupprecht et al.

World simulators can provide safe and scalable environments for training Physical AI systems before real-world deployment. Large video generation models are emerging as a promising basis for such simulators because they can generate diverse and realistic visual futures. However, using them as world simulators requires physically faithful video continuations, namely, generated videos that preserve the physical state implied by the conditioning input, and evolve in ways consistent with basic physical principles. We propose PhyWorld, a video generation world model designed to produce temporally coherent and physically faithful scene continuations through two-stage post-training. In the first stage, we improve video-to-video continuation with flow matching fine-tuning, encouraging stable visual attributes and coherent motion dynamics across frames. In the second stage, we align generated dynamics with physical principles using Direct Preference Optimization (DPO) over physics preference pairs, guiding the model toward outputs with higher physical plausibility. To evaluate PhyWorld, we use both standard video-quality benchmarks and a dedicated physical-faithfulness benchmark with per-law scoring. Experiments show that PhyWorld improves video consistency, achieving an average score of 0.769 on VBench compared with 0.756 or below for state-of-the-art baselines. PhyWorld also improves physical plausibility, reaching an average score of 3.09 on our physical-faithfulness benchmark compared with 2.99 for the strongest baseline. These results suggest that post-training large video generation models with continuation and physics-preference signals can make them more effective world simulators for Physical AI.

Arash Akbari, Arman Akbari, Masih Eskandar et al.

Vision-Language-Action (VLA) models exhibit remarkable action generation for embodied intelligence, but their heavy compute make deployment on edge platforms impractical. Aggressive, sub-4-bit weight quantization is the natural solution, yet existing post-training quantization (PTQ) methods suffer severe performance degradation in this regime. To address this, we introduce ActQuant, an action-guided mixed-precision PTQ framework that operates in two stages: (1) an inter-tensor bit allocator that assigns each weight matrix a single bit-width based on how much it contributes to predicting the agent's actions; (2) an intra-tensor scale optimizer tunes per-block quantization scales using action-aware curvature, so that dynamic range is concentrated on the weights most influential for control. To deliver the on-device benefits of our aggressive quantization, we further introduce OmniModel.cpp, an agentic conversion pipeline that ports architectures into a native C/C++ runtime with efficient low-bit kernels. We evaluate ActQuant both in simulation and on a real-world 6-DoF UR3 arm, with all models deployed through OmniModel.cpp. On the LIBERO benchmark, ActQuant is the only method that operates at or below 3 bits-per-weight, retaining 95.0% on OpenVLA-OFT and 94.8% on $π_{0.5}$. Pushed further, ActQuant reaches 2.5 bpw at 90.1% on OpenVLA-OFT, compressing the backbone from 14.3 GB to 2.7 GB (5.3$\times$). On the physical UR3 arm, $π_{0.5}$ quantized with ActQuant retains the baseline's success rate while reducing the memory footprint by 2.5$\times$.

Ying Zhang, Ningxi Cheng, Yizhu Gao et al.

Q-matrices are a cornerstone of theory-driven assessment and learning analytics, making item demands and students' underlying knowledge components and misconceptions explicit and actionable. However, Q-matrices are typically crafted by experts, making them time-consuming to build, prone to subjectivity, and difficult to validate empirically. We propose a framework for human-AI Q-matrix refinement in which large language models (LLMs) generate candidate Q-matrices using structured, misconception-aware prompting, and NeuralCDM provides an empirical evaluation layer to compare candidates based on how well they explain student response data. We apply the framework to a thermodynamics assessment dataset and benchmark locally deployed LLMs against cloud-served models. Results show that iteratively refined LLM-generated Q-matrices can exceed expert-baseline model fit (AUC 0.780 vs. 0.717), and that locally deployed models achieve comparable performance to cloud APIs, supporting privacy-preserving deployment.

Qitao Tan, Xiaoying Song, Ningxi Cheng et al.

Public large language models (LLMs) are typically safety-aligned during pretraining, yet task-specific fine-tuning required for deployment often erodes this alignment and introduces safety risks. Existing defenses either embed safety recovery into fine-tuning or rely on fine-tuning-derived priors for post-hoc correction, leaving safety recovery tightly coupled with training and incurring high computational overhead and a complex workflow. To address these challenges, we propose \texttt{Q-realign}, a post-hoc defense method based on post-training quantization, guided by an analysis of representational structure. By reframing quantization as a dual-objective procedure for compression and safety, \texttt{Q-realign} decouples safety alignment from fine-tuning and naturally piggybacks into modern deployment pipelines. Experiments across multiple models and datasets demonstrate that our method substantially reduces unsafe behaviors while preserving task performance, with significant reductions in memory usage and GPU hours. Notably, our approach can recover the safety alignment of a fine-tuned 7B LLM on a single RTX 4090 within 40 minutes. Overall, our work provides a practical, turnkey solution for safety-aware deployment.

Zhengxiao He, Huayu Li, Geng Yuan et al.

Methods: We developed a self-supervised deep learning model that extracts meaningful patterns from multi-modal signals (Electroencephalography (EEG), Electrocardiography (ECG), and respiratory signals). The model was trained on data from 4,398 participants. Projection scores were derived by contrasting embeddings from individuals with and without CVD outcomes. External validation was conducted in an independent cohort with 1,093 participants. The source code is available on https://github.com/miraclehetech/sleep-ssl. Results: The projection scores revealed distinct and clinically meaningful patterns across modalities. ECG-derived features were predictive of both prevalent and incident cardiac conditions, particularly CVD mortality. EEG-derived features were predictive of incident hypertension and CVD mortality. Respiratory signals added complementary predictive value. Combining these projection scores with the Framingham Risk Score consistently improved predictive performance, achieving area under the curve values ranging from 0.607 to 0.965 across different outcomes. Findings were robustly replicated and validated in the external testing cohort. Conclusion: Our findings demonstrate that the proposed framework can generate individualized CVD risk scores directly from PSG data. The resulting projection scores have the potential to be integrated into clinical practice, enhancing risk assessment and supporting personalized care.

Geng Yuan, Xiaolong Ma, Wei Niu et al.

Recently, a new trend of exploring sparsity for accelerating neural network training has emerged, embracing the paradigm of training on the edge. This paper proposes a novel Memory-Economic Sparse Training (MEST) framework targeting for accurate and fast execution on edge devices. The proposed MEST framework consists of enhancements by Elastic Mutation (EM) and Soft Memory Bound (&S) that ensure superior accuracy at high sparsity ratios. Different from the existing works for sparse training, this current work reveals the importance of sparsity schemes on the performance of sparse training in terms of accuracy as well as training speed on real edge devices. On top of that, the paper proposes to employ data efficiency for further acceleration of sparse training. Our results suggest that unforgettable examples can be identified in-situ even during the dynamic exploration of sparsity masks in the sparse training process, and therefore can be removed for further training speedup on edge devices. Comparing with state-of-the-art (SOTA) works on accuracy, our MEST increases Top-1 accuracy significantly on ImageNet when using the same unstructured sparsity scheme. Systematical evaluation on accuracy, training speed, and memory footprint are conducted, where the proposed MEST framework consistently outperforms representative SOTA works. A reviewer strongly against our work based on his false assumptions and misunderstandings. On top of the previous submission, we employ data efficiency for further acceleration of sparse training. And we explore the impact of model sparsity, sparsity schemes, and sparse training algorithms on the number of removable training examples. Our codes are publicly available at: https://github.com/boone891214/MEST.

Xiaolong Ma, Geng Yuan, Xuan Shen et al.

There have been long-standing controversies and inconsistencies over the experiment setup and criteria for identifying the "winning ticket" in literature. To reconcile such, we revisit the definition of lottery ticket hypothesis, with comprehensive and more rigorous conditions. Under our new definition, we show concrete evidence to clarify whether the winning ticket exists across the major DNN architectures and/or applications. Through extensive experiments, we perform quantitative analysis on the correlations between winning tickets and various experimental factors, and empirically study the patterns of our observations. We find that the key training hyperparameters, such as learning rate and training epochs, as well as the architecture characteristics such as capacities and residual connections, are all highly correlated with whether and when the winning tickets can be identified. Based on our analysis, we summarize a guideline for parameter settings in regards of specific architecture characteristics, which we hope to catalyze the research progress on the topic of lottery ticket hypothesis. Our codes are publicly available at: https://github.com/boone891214/sanity-check-LTH.

Qing Jin, Jian Ren, Oliver J. Woodford et al.

Generative Adversarial Networks (GANs) have achieved huge success in generating high-fidelity images, however, they suffer from low efficiency due to tremendous computational cost and bulky memory usage. Recent efforts on compression GANs show noticeable progress in obtaining smaller generators by sacrificing image quality or involving a time-consuming searching process. In this work, we aim to address these issues by introducing a teacher network that provides a search space in which efficient network architectures can be found, in addition to performing knowledge distillation. First, we revisit the search space of generative models, introducing an inception-based residual block into generators. Second, to achieve target computation cost, we propose a one-step pruning algorithm that searches a student architecture from the teacher model and substantially reduces searching cost. It requires no l1 sparsity regularization and its associated hyper-parameters, simplifying the training procedure. Finally, we propose to distill knowledge through maximizing feature similarity between teacher and student via an index named Global Kernel Alignment (GKA). Our compressed networks achieve similar or even better image fidelity (FID, mIoU) than the original models with much-reduced computational cost, e.g., MACs. Code will be released at https://github.com/snap-research/CAT.

Yuxuan Cai, Hongjia Li, Geng Yuan et al.

The rapid development and wide utilization of object detection techniques have aroused attention on both accuracy and speed of object detectors. However, the current state-of-the-art object detection works are either accuracy-oriented using a large model but leading to high latency or speed-oriented using a lightweight model but sacrificing accuracy. In this work, we propose YOLObile framework, a real-time object detection on mobile devices via compression-compilation co-design. A novel block-punched pruning scheme is proposed for any kernel size. To improve computational efficiency on mobile devices, a GPU-CPU collaborative scheme is adopted along with advanced compiler-assisted optimizations. Experimental results indicate that our pruning scheme achieves 14$\times$ compression rate of YOLOv4 with 49.0 mAP. Under our YOLObile framework, we achieve 17 FPS inference speed using GPU on Samsung Galaxy S20. By incorporating our proposed GPU-CPU collaborative scheme, the inference speed is increased to 19.1 FPS, and outperforms the original YOLOv4 by 5$\times$ speedup. Source code is at: \url{https://github.com/nightsnack/YOLObile}.

Kaiyuan Deng, Hangyu Zheng, Minghai Qing et al.

Deploying models, especially large language models (LLMs), is becoming increasingly attractive to a broader user base, including those without specialized expertise. However, due to the resource constraints of certain hardware, maintaining high accuracy with larger model while meeting the hardware requirements remains a significant challenge. Model quantization technique helps mitigate memory and compute bottlenecks, yet the added complexities of tuning and deploying quantized models further exacerbates these challenges, making the process unfriendly to most of the users. We introduce the Hardware-Aware Quantization Agent (HAQA), an automated framework that leverages LLMs to streamline the entire quantization and deployment process by enabling efficient hyperparameter tuning and hardware configuration, thereby simultaneously improving deployment quality and ease of use for a broad range of users. Our results demonstrate up to a 2.3x speedup in inference, along with increased throughput and improved accuracy compared to unoptimized models on Llama. Additionally, HAQA is designed to implement adaptive quantization strategies across diverse hardware platforms, as it automatically finds optimal settings even when they appear counterintuitive, thereby reducing extensive manual effort and demonstrating superior adaptability. Code will be released.

Sheng Li, Geng Yuan, Yue Dai et al.

There has been a proliferation of artificial intelligence applications, where model training is key to promising high-quality services for these applications. However, the model training process is both time-intensive and energy-intensive, inevitably affecting the user's demand for application efficiency. Layer freezing, an efficient model training technique, has been proposed to improve training efficiency. Although existing layer freezing methods demonstrate the great potential to reduce model training costs, they still remain shortcomings such as lacking generalizability and compromised accuracy. For instance, existing layer freezing methods either require the freeze configurations to be manually defined before training, which does not apply to different networks, or use heuristic freezing criteria that is hard to guarantee decent accuracy in different scenarios. Therefore, there lacks a generic and smart layer freezing method that can automatically perform ``in-situation'' layer freezing for different networks during training processes. To this end, we propose a generic and efficient training framework (SmartFRZ). The core proposed technique in SmartFRZ is attention-guided layer freezing, which can automatically select the appropriate layers to freeze without compromising accuracy. Experimental results show that SmartFRZ effectively reduces the amount of computation in training and achieves significant training acceleration, and outperforms the state-of-the-art layer freezing approaches.

Zheng Zhan, Yushu Wu, Yifan Gong et al. · harvard

The rapid progress in artificial intelligence-generated content (AIGC), especially with diffusion models, has significantly advanced development of high-quality video generation. However, current video diffusion models exhibit demanding computational requirements and high peak memory usage, especially for generating longer and higher-resolution videos. These limitations greatly hinder the practical application of video diffusion models on standard hardware platforms. To tackle this issue, we present a novel, training-free framework named Streamlined Inference, which leverages the temporal and spatial properties of video diffusion models. Our approach integrates three core components: Feature Slicer, Operator Grouping, and Step Rehash. Specifically, Feature Slicer effectively partitions input features into sub-features and Operator Grouping processes each sub-feature with a group of consecutive operators, resulting in significant memory reduction without sacrificing the quality or speed. Step Rehash further exploits the similarity between adjacent steps in diffusion, and accelerates inference through skipping unnecessary steps. Extensive experiments demonstrate that our approach significantly reduces peak memory and computational overhead, making it feasible to generate high-quality videos on a single consumer GPU (e.g., reducing peak memory of AnimateDiff from 42GB to 11GB, featuring faster inference on 2080Ti).

Jun Liu, Zhenglun Kong, Pu Zhao et al. · harvard

Structured pruning for large language models (LLMs) has garnered significant academic interest due to its ability to efficiently compress and accelerate LLMs by eliminating redundant weight groups at a coarse-grained granularity. Current structured pruning methods for LLMs typically depend on a singular granularity for assessing weight importance, resulting in notable performance degradation in downstream tasks. Intriguingly, our empirical investigations reveal that utilizing unstructured pruning, which achieves better performance retention by pruning weights at a finer granularity, \emph{i.e.}, individual weights, yields significantly varied sparse LLM structures when juxtaposed to structured pruning. This suggests that evaluating both holistic and individual assessment for weight importance is essential for LLM pruning. Building on this insight, we introduce the Hybrid-grained Weight Importance Assessment (HyWIA), a novel method that merges fine-grained and coarse-grained evaluations of weight importance for the pruning of LLMs. Leveraging an attention mechanism, HyWIA adaptively determines the optimal blend of granularity in weight importance assessments in an end-to-end pruning manner. Extensive experiments on LLaMA-V1/V2, Vicuna, Baichuan, and Bloom across various benchmarks demonstrate the effectiveness of HyWIA in pruning LLMs. For example, HyWIA surpasses the cutting-edge LLM-Pruner by an average margin of 2.82% in accuracy across seven downstream tasks when pruning LLaMA-7B by 50%.

Jun Liu, Zhenglun Kong, Peiyan Dong et al. · harvard

Fine-tuning helps large language models (LLM) recover degraded information and enhance task performance. Although Low-Rank Adaptation (LoRA) is widely used and effective for fine-tuning, we have observed that its scaling factor can limit or even reduce performance as the rank size increases. To address this issue, we propose RoRA (Rank-adaptive Reliability Optimization), a simple yet effective method for optimizing LoRA's scaling factor. By replacing $α/r$ with $α/\sqrt{r}$, RoRA ensures improved performance as rank size increases. Moreover, RoRA enhances low-rank adaptation in fine-tuning uncompressed models and excels in the more challenging task of accuracy recovery when fine-tuning pruned models. Extensive experiments demonstrate the effectiveness of RoRA in fine-tuning both uncompressed and pruned models. RoRA surpasses the state-of-the-art (SOTA) in average accuracy and robustness on LLaMA-7B/13B, LLaMA2-7B, and LLaMA3-8B, specifically outperforming LoRA and DoRA by 6.5% and 2.9% on LLaMA-7B, respectively. In pruned model fine-tuning, RoRA shows significant advantages; for SHEARED-LLAMA-1.3, a LLaMA-7B with 81.4% pruning, RoRA achieves 5.7% higher average accuracy than LoRA and 3.9% higher than DoRA.

Jun Liu, Zhenglun Kong, Pu Zhao et al. · harvard

Autonomous driving platforms encounter diverse driving scenarios, each with varying hardware resources and precision requirements. Given the computational limitations of embedded devices, it is crucial to consider computing costs when deploying on target platforms like the NVIDIA\textsuperscript{\textregistered} DRIVE PX 2. Our objective is to customize the semantic segmentation network according to the computing power and specific scenarios of autonomous driving hardware. We implement dynamic adaptability through a three-tier control mechanism -- width multiplier, classifier depth, and classifier kernel -- allowing fine-grained control over model components based on hardware constraints and task requirements. This adaptability facilitates broad model scaling, targeted refinement of the final layers, and scenario-specific optimization of kernel sizes, leading to improved resource allocation and performance. Additionally, we leverage Bayesian Optimization with surrogate modeling to efficiently explore hyperparameter spaces under tight computational budgets. Our approach addresses scenario-specific and task-specific requirements through automatic parameter search, accommodating the unique computational complexity and accuracy needs of autonomous driving. It scales its Multiply-Accumulate Operations (MACs) for Task-Specific Learning Adaptation (TSLA), resulting in alternative configurations tailored to diverse self-driving tasks. These TSLA customizations maximize computational capacity and model accuracy, optimizing hardware utilization.

Bingbing Li, Geng Yuan, Zigeng Wang et al.

Resistive Random Access Memory (ReRAM) has emerged as a promising platform for deep neural networks (DNNs) due to its support for parallel in-situ matrix-vector multiplication. However, hardware failures, such as stuck-at-fault defects, can result in significant prediction errors during model inference. While additional crossbars can be used to address these failures, they come with storage overhead and are not efficient in terms of space, energy, and cost. In this paper, we propose a fault protection mechanism that incurs zero space cost. Our approach includes: 1) differentiable structure pruning of rows and columns to reduce model redundancy, 2) weight duplication and voting for robust output, and 3) embedding duplicated most significant bits (MSBs) into the model weight. We evaluate our method on nine tasks of the GLUE benchmark with the BERT model, and experimental results prove its effectiveness.

Qitao Tan, Sung-En Chang, Rui Xia et al.

Zeroth-order (ZO) optimization is an emerging deep neural network (DNN) training paradigm that offers computational simplicity and memory savings. However, this seemingly promising approach faces a significant and long-ignored challenge. ZO requires generating a substantial number of Gaussian random numbers, which poses significant difficulties and even makes it infeasible for hardware platforms, such as FPGAs and ASICs. In this paper, we identify this critical issue, which arises from the mismatch between algorithm and hardware designers. To address this issue, we proposed PeZO, a perturbation-efficient ZO framework. Specifically, we design random number reuse strategies to significantly reduce the demand for random number generation and introduce a hardware-friendly adaptive scaling method to replace the costly Gaussian distribution with a uniform distribution. Our experiments show that PeZO reduces the required LUTs and FFs for random number generation by 48.6\% and 12.7\%, and saves at maximum 86\% power consumption, all without compromising training performance, making ZO optimization feasible for on-device training. To the best of our knowledge, we are the first to explore the potential of on-device ZO optimization, providing valuable insights for future research.

Jun Liu, Zhenglun Kong, Changdi Yang et al. · harvard

Multi-agent large language model (LLM) systems have shown strong potential in complex reasoning and collaborative decision-making tasks. However, most existing coordination schemes rely on static or full-context routing strategies, which lead to excessive token consumption, redundant memory exposure, and limited adaptability across interaction rounds. We introduce RCR-Router, a modular and role-aware context routing framework designed to enable efficient, adaptive collaboration in multi-agent LLMs. To our knowledge, this is the first routing approach that dynamically selects semantically relevant memory subsets for each agent based on its role and task stage, while adhering to a strict token budget. A lightweight scoring policy guides memory selection, and agent outputs are iteratively integrated into a shared memory store to facilitate progressive context refinement. To better evaluate model behavior, we further propose an Answer Quality Score metric that captures LLM-generated explanations beyond standard QA accuracy. Experiments on three multi-hop QA benchmarks -- HotPotQA, MuSiQue, and 2WikiMultihop -- demonstrate that RCR-Router reduces token usage (up to 30%) while improving or maintaining answer quality. These results highlight the importance of structured memory routing and output-aware evaluation in advancing scalable multi-agent LLM systems.

Zhendong Mi, Qitao Tan, Xiaodong Yu et al.

Large language models (LLMs) have demonstrated impressive capabilities across numerous NLP tasks. Nevertheless, conventional first-order fine-tuning techniques impose heavy memory demands, creating practical obstacles to real-world applications. Zeroth-order (ZO) optimization has recently emerged as a promising memory-efficient alternative, as it circumvents the need for backpropagation by estimating gradients solely through forward passes--making it particularly suitable for resource-limited environments. Despite its efficiency, ZO optimization suffers from gradient estimation bias, which significantly hinders convergence speed. To address this, we analytically identify and characterize the lower-order bias introduced during ZO-based gradient estimation in LLM fine-tuning. Motivated by tools in mathematical physics, we introduce a kernel-function-based ZO framework aimed at mitigating this bias and improving optimization stability. KerZOO achieves comparable or superior performance to existing ZO baselines in both full-parameter and parameter-efficient fine-tuning settings of LLMs, while significantly reducing the number of iterations required to reach convergence. For example, KerZOO reduces total GPU training hours by as much as 74% and 44% on WSC and MultiRC datasets in fine-tuning OPT-2.7B model and can exceed the MeZO baseline by 2.9% and 2.6% in accuracy. We show that the kernel function is an effective avenue for reducing estimation bias in ZO methods.

Jun Liu, Zhenglun Kong, Peiyan Dong et al. · harvard

Large language models (LLMs) exhibit strong capabilities as decision-making agents by interleaving reasoning and actions, as seen in ReAct-style frameworks. Yet, their practical deployment is constrained by high inference costs and large model sizes. We propose Structured Agent Distillation, a framework that compresses large LLM-based agents into smaller student models while preserving both reasoning fidelity and action consistency. Unlike standard token-level distillation, our method segments trajectories into {[REASON]} and {[ACT]} spans, applying segment-specific losses to align each component with the teacher's behavior. This structure-aware supervision enables compact agents to better replicate the teacher's decision process. Experiments on ALFWorld, HotPotQA-ReAct, and WebShop show that our approach consistently outperforms token-level and imitation learning baselines, achieving significant compression with minimal performance drop. Scaling and ablation results further highlight the importance of span-level alignment for efficient and deployable agents.

Zhendong Mi, Zhenglun Kong, Geng Yuan et al. · harvard

With the rapid expansion of large language models (LLMs), the demand for memory and computational resources has grown significantly. Recent advances in LLM pruning aim to reduce the size and computational cost of these models. However, existing methods often suffer from either suboptimal pruning performance or low time efficiency during the pruning process. In this work, we propose an efficient and effective pruning method that simultaneously achieves high pruning performance and fast pruning speed with improved calibration efficiency. Our approach introduces two key innovations: (1) An activation cosine similarity loss-guided pruning metric, which considers the angular deviation of the output activation between the dense and pruned models. (2) An activation variance-guided pruning metric, which helps preserve semantic distinctions in output activations after pruning, enabling effective pruning with shorter input sequences. These two components can be readily combined to enhance LLM pruning in both accuracy and efficiency. Experimental results show that our method achieves up to an 18% reduction in perplexity and up to 63% decrease in pruning time on prevalent LLMs such as LLaMA, LLaMA-2, and OPT.

Huayu Li, Zhengxiao He, Xiwen Chen et al.

Large language models (LLMs) have shown promising capabilities in visually interpreting medical time-series data. However, their general-purpose design can limit domain-specific precision, and the proprietary nature of many models poses challenges for fine-tuning on specialized clinical datasets. Conversely, small specialized models (SSMs) offer strong performance on focused tasks but lack the broader reasoning needed for complex medical decision-making. To address these complementary limitations, we introduce \ConMIL{} (Conformalized Multiple Instance Learning), a novel decision-support framework distinctively synergizes three key components: (1) a new Multiple Instance Learning (MIL) mechanism, QTrans-Pooling, designed for per-class interpretability in identifying clinically relevant physiological signal segments; (2) conformal prediction, integrated with MIL to generate calibrated, set-valued outputs with statistical reliability guarantees; and (3) a structured approach for these interpretable and uncertainty-quantified SSM outputs to enhance the visual inspection capabilities of LLMs. Our experiments on arrhythmia detection and sleep stage classification demonstrate that \ConMIL{} can enhance the accuracy of LLMs such as ChatGPT4.0, Qwen2-VL-7B, and MiMo-VL-7B-RL. For example, \ConMIL{}-supported Qwen2-VL-7B and MiMo-VL-7B-RL both achieves 94.92% and 96.82% precision on confident samples and (70.61% and 78.02%)/(78.10% and 71.98%) on uncertain samples for the two tasks, compared to 46.13% and 13.16% using the LLM alone. These results suggest that integrating task-specific models with LLMs may offer a promising pathway toward more interpretable and trustworthy AI-driven clinical decision support.

Sheng Li, Geng Yuan, Yue Dai et al.

Emerging applications, such as robot-assisted eldercare and object recognition, generally employ deep learning neural networks (DNNs) and naturally require: i) handling streaming-in inference requests and ii) adapting to possible deployment scenario changes. Online model fine-tuning is widely adopted to satisfy these needs. However, an inappropriate fine-tuning scheme could involve significant energy consumption, making it challenging to deploy on edge devices. In this paper, we propose EdgeOL, an edge online learning framework that optimizes inference accuracy, fine-tuning execution time, and energy efficiency through both inter-tuning and intra-tuning optimizations. Experimental results show that, on average, EdgeOL reduces overall fine-tuning execution time by 64%, energy consumption by 52%, and improves average inference accuracy by 1.75% over the immediate online learning strategy

Weihang You, Qingchan Zhu, David Liu et al.

Chain-of-Thought (CoT) reasoning excels in language models but struggles in vision-language models due to premature visual-to-text conversion that discards continuous information such as geometry and spatial layout. While recent methods enhance CoT through static enumeration or attention-based selection, they remain passive, i.e., processing pre-computed inputs rather than actively seeking task-relevant details. Inspired by human active perception, we introduce ViThinker, a framework that enables vision-language models to autonomously generate decision (query) tokens triggering the synthesis of expert-aligned visual features on demand. ViThinker internalizes vision-expert capabilities during training, performing generative mental simulation during inference without external tool calls. Through a two-stage curriculum: first distilling frozen experts into model parameters, then learning task-driven querying via sparsity penalties, i.e., ViThinker discovers minimal sufficient perception for each reasoning step. Evaluations across vision-centric benchmarks demonstrate consistent improvements, validating that active query generation outperforms passive approaches in both perceptual grounding and reasoning accuracy.

Ci Zhang, Huayu Li, Changdi Yang et al.

Recent studies show that using diffusion models for time series signal reconstruction holds great promise. However, such approaches remain largely unexplored in the domain of medical time series. The unique characteristics of the physiological time series signals, such as multivariate, high temporal variability, highly noisy, and artifact-prone, make deep learning-based approaches still challenging for tasks such as imputation. Hence, we propose a novel Mixture of Experts (MoE)-based noise estimator within a score-based diffusion framework. Specifically, the Receptive Field Adaptive MoE (RFAMoE) module is designed to enable each channel to adaptively select desired receptive fields throughout the diffusion process. Moreover, recent literature has found that when generating a physiological signal, performing multiple inferences and averaging the reconstructed signals can effectively reduce reconstruction errors, but at the cost of significant computational and latency overhead. We design a Fusion MoE module and innovatively leverage the nature of MoE module to generate K noise signals in parallel, fuse them using a routing mechanism, and complete signal reconstruction in a single inference step. This design not only improves performance over previous methods but also eliminates the substantial computational cost and latency associated with multiple inference processes. Extensive results demonstrate that our proposed framework consistently outperforms diffusion-based SOTA works on different tasks and datasets.

Zhendong Mi, Qitao Tan, Grace Li Zhang et al.

Fine-tuning large language models (LLMs) using zeroth-order (ZO) optimization has emerged as a promising alternative to traditional gradient-based methods due to its reduced memory footprint requirement. However, existing ZO methods suffer from high variance in gradient estimation, leading to slow convergence and suboptimal performance on large-scale models. In this work, we propose P-GAP, a fast LLM fine-tuning approach through zeroth-order optimization with Projected Gradient-Aligned Perturbations. Specifically, we first estimate a low-dimensional gradient space and then align perturbations in projected gradients' direction within the space. This approach enables reduced the number of perturbed parameters and decreased variance, therefore accelerated convergence for LLM fine-tuning. Experiments on LLMs show that P-GAP consistently surpasses the baselines, achieving up to 6% increase in accuracy on classification tasks and up to 12% higher accuracy on generation tasks, with up to about 81% less training iterations and 70% less GPU hours. These results demonstrate that P-GAP enables fast, scalable, and resource-efficient ZO LLM fine-tuning.

Qitao Tan, Xiaoying Song, Jin Lu et al.

Quantization is an effective technique to reduce the deployment cost of large language models (LLMs), and post-training quantization (PTQ) has been widely studied due to its efficiency. However, existing PTQ methods are limited by their inability to fine-tune model parameters and often suffer significant accuracy loss in low-bit scenarios. Quantization-aware training (QAT) provides a more principled solution, but its reliance on backpropagation incurs prohibitive memory costs, limiting its practicality for LLM deployment. To address these challenges, we propose ZeroQAT, a zeroth-order optimization-based QAT framework that supports both weight and activation quantization. ZeroQAT leverages forward-only gradient estimation to eliminate backpropagation, substantially reducing computational and memory overhead while retaining the benefits of end-to-end optimization. We further introduce a lightweight variant of ZeroQAT for quantized fine-tuning, which freezes and pre-quantizes most parameters to further cut memory usage. Experiments show that ZeroQAT consistently outperforms representative PTQ and QAT baselines while requiring significantly less memory. For example, ZeroQAT enables fine-tuning of a 13B model at extremely low bit-widths (e.g., 2-4 bits) on a single 8GB GPU, and even allows fine-tuning a 6.7B model on a OnePlus 12 smartphone, demonstrating its practicality for end-to-end QAT on resource-limited edge devices.

Chence Yang, Ci Zhang, Lei Lu et al.

With the growing size of deep neural networks and datasets, the computational costs of training have significantly increased. The layer-freezing technique has recently attracted great attention as a promising method to effectively reduce the cost of network training. However, in traditional layer-freezing methods, frozen layers are still required for forward propagation to generate feature maps for unfrozen layers, limiting the reduction of computation costs. To overcome this, prior works proposed a hypothetical solution, which caches feature maps from frozen layers as a new dataset, allowing later layers to train directly on stored feature maps. While this approach appears to be straightforward, it presents several major challenges that are severely overlooked by prior literature, such as how to effectively apply augmentations to feature maps and the substantial storage overhead introduced. If these overlooked challenges are not addressed, the performance of the caching method will be severely impacted and even make it infeasible. This paper is the first to comprehensively explore these challenges and provides a systematic solution. To improve training accuracy, we propose \textit{similarity-aware channel augmentation}, which caches channels with high augmentation sensitivity with a minimum additional storage cost. To mitigate storage overhead, we incorporate lossy data compression into layer freezing and design a \textit{progressive compression} strategy, which increases compression rates as more layers are frozen, effectively reducing storage costs. Finally, our solution achieves significant reductions in training cost while maintaining model accuracy, with a minor time overhead. Additionally, we conduct a comprehensive evaluation of freezing and compression strategies, providing insights into optimizing their application for efficient DNN training.

Zhenglun Kong, Peiyan Dong, Xiaolong Ma et al.

Recently, Vision Transformer (ViT) has continuously established new milestones in the computer vision field, while the high computation and memory cost makes its propagation in industrial production difficult. Pruning, a traditional model compression paradigm for hardware efficiency, has been widely applied in various DNN structures. Nevertheless, it stays ambiguous on how to perform exclusive pruning on the ViT structure. Considering three key points: the structural characteristics, the internal data pattern of ViTs, and the related edge device deployment, we leverage the input token sparsity and propose a computation-aware soft pruning framework, which can be set up on vanilla Transformers of both flatten and CNN-type structures, such as Pooling-based ViT (PiT). More concretely, we design a dynamic attention-based multi-head token selector, which is a lightweight module for adaptive instance-wise token selection. We further introduce a soft pruning technique, which integrates the less informative tokens generated by the selector module into a package token that will participate in subsequent calculations rather than being completely discarded. Our framework is bound to the trade-off between accuracy and computation constraints of specific edge devices through our proposed computation-aware training strategy. Experimental results show that our framework significantly reduces the computation cost of ViTs while maintaining comparable performance on image classification. Moreover, our framework can guarantee the identified model to meet resource specifications of mobile devices and FPGA, and even achieve the real-time execution of DeiT-T on mobile platforms. For example, our method reduces the latency of DeiT-T to 26 ms (26%$\sim $41% superior to existing works) on the mobile device with 0.25%$\sim $4% higher top-1 accuracy on ImageNet.

Yifan Gong, Geng Yuan, Zheng Zhan et al.

Weight pruning is an effective model compression technique to tackle the challenges of achieving real-time deep neural network (DNN) inference on mobile devices. However, prior pruning schemes have limited application scenarios due to accuracy degradation, difficulty in leveraging hardware acceleration, and/or restriction on certain types of DNN layers. In this paper, we propose a general, fine-grained structured pruning scheme and corresponding compiler optimizations that are applicable to any type of DNN layer while achieving high accuracy and hardware inference performance. With the flexibility of applying different pruning schemes to different layers enabled by our compiler optimizations, we further probe into the new problem of determining the best-suited pruning scheme considering the different acceleration and accuracy performance of various pruning schemes. Two pruning scheme mapping methods, one is search-based and the other is rule-based, are proposed to automatically derive the best-suited pruning regularity and block size for each layer of any given DNN. Experimental results demonstrate that our pruning scheme mapping methods, together with the general fine-grained structured pruning scheme, outperform the state-of-the-art DNN optimization framework with up to 2.48$\times$ and 1.73$\times$ DNN inference acceleration on CIFAR-10 and ImageNet dataset without accuracy loss.