The dominance of large decoder-only language models has overshadowed encoder-decoder architectures, despite their fundamental efficiency advantages in sequence processing. For small language models (SLMs) - those with 1 billion parameters or fewer - our systematic analysis across GPU, CPU, and NPU platforms reveals that encoder-decoder architectures achieve 47% lower first-token latency and 4.7x higher throughput compared to decoder-only models on edge devices. These gains may be attributed to encoder-decoder's one-time input processing and efficient separation of understanding and generation phases. We introduce a novel knowledge distillation framework that enables encoder-decoder models to leverage capabilities from large scalable decoder-only teachers while preserving their architectural advantages, achieving up to 6 average performance points improvement across diverse tasks, with significant gains in asymmetric sequence tasks where input and output distributions can benefit from different processing approaches. When combined with modern advances like Rotary Positional Embeddings (RoPE) and Vision encoders, our systematic investigation demonstrates that encoder-decoder architectures provide a more practical path toward deploying capable language models in resource-constrained environments. Our findings challenge the prevailing trend toward decoder-only scaling, showing that architectural choices become increasingly crucial as parameter budgets decrease, particularly for on-device and edge deployments where computational efficiency is paramount.
Graph neural networks (GNNs) have gained considerable attention in recent years for traffic flow prediction due to their ability to learn spatio-temporal pattern representations through a graph-based message-passing framework. Although GNNs have shown great promise in handling traffic datasets, their deployment in real-life applications has been hindered by scalability constraints arising from high-order message passing. Additionally, the over-smoothing problem of GNNs may lead to indistinguishable region representations as the number of layers increases, resulting in performance degradation. To address these challenges, we propose a new knowledge distillation paradigm termed LightST that transfers spatial and temporal knowledge from a high-capacity teacher to a lightweight student. Specifically, we introduce a spatio-temporal knowledge distillation framework that helps student MLPs capture graph-structured global spatio-temporal patterns while alleviating the over-smoothing effect with adaptive knowledge distillation. Extensive experiments verify that LightST significantly speeds up traffic flow predictions by 5X to 40X compared to state-of-the-art spatio-temporal GNNs, all while maintaining superior accuracy.
Recent advancements in deep learning have significantly improved performance on computer vision tasks. Previous image classification methods primarily modify model architectures or add features, and they optimize models using cross-entropy loss on class logits. Since they focus on classifying images with considering class labels, these methods may struggle to learn various \emph{aspects} of classes (e.g., natural positions and shape changes). Rethinking the previous approach from a novel view, we propose a multi-aspect knowledge distillation method using Multimodal Large Language Models (MLLMs). Our approach involves: 1) querying Large Language Model with multi-aspect questions relevant to the knowledge we want to transfer to the model, 2) extracting corresponding logits from MLLM, and 3) expanding the model's output dimensions to distill these multi-aspect logits. We then apply cross-entropy loss to class logits and binary cross-entropy loss to multi-aspect logits. Through our method, the model can learn not only the knowledge about visual aspects but also the abstract and complex aspects that require a deeper understanding. We primarily apply our method to image classification, and to explore the potential for extending our model, we expand it to other tasks, such as object detection. In all experimental results, our method improves the performance of the baselines. Additionally, we analyze the effect of multi-aspect knowledge distillation. These results demonstrate that our method can transfer knowledge about various aspects to the model and the aspect knowledge can enhance model performance in computer vision tasks. This paper demonstrates the great potential of multi-aspect knowledge distillation, and we believe it offers a promising direction for future research in computer vision and beyond.
With the advent of the COVID-19 pandemic, ultrasound imaging has emerged as a promising technique for COVID-19 detection, due to its non-invasive nature, affordability, and portability. In response, researchers have focused on developing AI-based scoring systems to provide real-time diagnostic support. However, the limited size and lack of proper annotation in publicly available ultrasound datasets pose significant challenges for training a robust AI model. This paper proposes MeDiVLAD, a novel pipeline to address the above issue for multi-level lung-ultrasound (LUS) severity scoring. In particular, we leverage self-knowledge distillation to pretrain a vision transformer (ViT) without label and aggregate frame-level features via dual-level VLAD aggregation. We show that with minimal finetuning, MeDiVLAD outperforms conventional fully-supervised methods in both frame- and video-level scoring, while offering classification reasoning with exceptional quality. This superior performance enables key applications such as the automatic identification of critical lung pathology areas and provides a robust solution for broader medical video classification tasks.
Class-incremental fault diagnosis requires a model to adapt to new fault classes while retaining previous knowledge. However, limited research exists for imbalanced and long-tailed data. Extracting discriminative features from few-shot fault data is challenging, and adding new fault classes often demands costly model retraining. Moreover, incremental training of existing methods risks catastrophic forgetting, and severe class imbalance can bias the model's decisions toward normal classes. To tackle these issues, we introduce a Supervised Contrastive knowledge distiLlation for class Incremental Fault Diagnosis (SCLIFD) framework proposing supervised contrastive knowledge distillation for improved representation learning capability and less forgetting, a novel prioritized exemplar selection method for sample replay to alleviate catastrophic forgetting, and the Random Forest Classifier to address the class imbalance. Extensive experimentation on simulated and real-world industrial datasets across various imbalance ratios demonstrates the superiority of SCLIFD over existing approaches. Our code can be found at https://github.com/Zhang-Henry/SCLIFD_TII.
Data scarcity in pharmaceutical research has led to reliance on labour-intensive trial-and-error approaches for development rather than data-driven methods. While Machine Learning offers a solution, existing datasets are often small and noisy, limiting their utility. To address this, we developed a Variationally Encoded Conditional Tabular Generative Adversarial Network (VECT-GAN), a novel generative model specifically designed for augmenting small, noisy datasets. We introduce a pipeline where data is augmented before regression model development and demonstrate that this consistently and significantly improves performance over other state-of-the-art tabular generative models. We apply this pipeline across six pharmaceutical datasets, and highlight its real-world applicability by developing novel polymers with medically desirable mucoadhesive properties, which we made and experimentally characterised. Additionally, we pre-train the model on the ChEMBL database of drug-like molecules, leveraging knowledge distillation to enhance its generalisability, making it readily available for use on pharmaceutical datasets containing small molecules, an extremely common pharmaceutical task. We demonstrate the power of synthetic data for regularising small tabular datasets, highlighting its potential to become standard practice in pharmaceutical model development, and make our method, including VECT-GAN pre-trained on ChEMBL available as a pip package.
The foundation model (FM) paradigm is transforming Machine Learning Force Fields (MLFFs), leveraging general-purpose representations and scalable training to perform a variety of computational chemistry tasks. Although MLFF FMs have begun to close the accuracy gap relative to first-principles methods, there is still a strong need for faster inference speed. Additionally, while research is increasingly focused on general-purpose models which transfer across chemical space, practitioners typically only study a small subset of systems at a given time. This underscores the need for fast, specialized MLFFs relevant to specific downstream applications, which preserve test-time physical soundness while maintaining train-time scalability. In this work, we introduce a method for transferring general-purpose representations from MLFF foundation models to smaller, faster MLFFs specialized to specific regions of chemical space. We formulate our approach as a knowledge distillation procedure, where the smaller "student" MLFF is trained to match the Hessians of the energy predictions of the "teacher" foundation model. Our specialized MLFFs can be up to 20 $\times$ faster than the original foundation model, while retaining, and in some cases exceeding, its performance and that of undistilled models. We also show that distilling from a teacher model with a direct force parameterization into a student model trained with conservative forces (i.e., computed as derivatives of the potential energy) successfully leverages the representations from the large-scale teacher for improved accuracy, while maintaining energy conservation during test-time molecular dynamics simulations. More broadly, our work suggests a new paradigm for MLFF development, in which foundation models are released along with smaller, specialized simulation "engines" for common chemical subsets.
Spiking Neural Networks (SNNs) are promising for low-power computation due to their event-driven mechanism but often suffer from lower accuracy compared to Artificial Neural Networks (ANNs). ANN-to-SNN knowledge distillation can improve SNN performance, but previous methods either focus solely on label information, missing valuable intermediate layer features, or use a layer-wise approach that neglects spatial and temporal semantic inconsistencies, leading to performance degradation.To address these limitations, we propose a novel method called self-attentive spatio-temporal calibration (SASTC). SASTC uses self-attention to identify semantically aligned layer pairs between ANN and SNN, both spatially and temporally. This enables the autonomous transfer of relevant semantic information. Extensive experiments show that SASTC outperforms existing methods, effectively solving the mismatching problem. Superior accuracy results include 95.12% on CIFAR-10, 79.40% on CIFAR-100 with 2 time steps, and 68.69% on ImageNet with 4 time steps for static datasets, and 97.92% on DVS-Gesture and 83.60% on DVS-CIFAR10 for neuromorphic datasets. This marks the first time SNNs have outperformed ANNs on both CIFAR-10 and CIFAR-100, shedding the new light on the potential applications of SNNs.
Large Vision Language Models (LVLMs) have demonstrated impressive capabilities in video understanding, yet their adoption for Activities of Daily Living (ADL) remains limited by their inability to capture fine-grained interactions and spatial relationships. To address this, we aim to leverage the complementary nature of egocentric views to enhance LVLM's understanding of exocentric ADL videos. Consequently, we propose ego2exo knowledge distillation to learn ego-augmented exp representations. While effective, this approach requires paired ego-exo videos, which are impractical to collect at scale. To address this, we propose Skeleton-guided Synthetic Ego Generation (SK-EGO), which leverages human skeleton motion to generate synthetic ego views from exocentric videos. To enhance the ego representation of LVLMs trained on synthetic data, we develop a domain-agnostic bootstrapped ego2exo strategy that effectively transfers knowledge from real ego-exo pairs to synthetic ego-exo pairs, while mitigating domain misalignment. We find that the exo representations of our ego-augmented LVLMs successfully learn to extract ego-perspective cues, demonstrated through comprehensive evaluation on six ADL benchmarks and our proposed Ego-in-Exo PerceptionMCQ benchmark designed specifically to assess egocentric understanding from exocentric videos. Code, models, and data will be open-sourced at https://github.com/dominickrei/EgoExo4ADL.
We introduce LLMQuoter, a lightweight, distillation-based model designed to enhance Retrieval Augmented Generation (RAG) by extracting the most relevant textual evidence for downstream reasoning tasks. Built on the LLaMA-3B architecture and fine-tuned with Low-Rank Adaptation (LoRA) on a 15,000-sample subset of HotpotQA, LLMQuoter adopts a "quote-first-then-answer" strategy, efficiently identifying key quotes before passing curated snippets to reasoning models. This workflow reduces cognitive overhead and outperforms full-context approaches like Retrieval-Augmented Fine-Tuning (RAFT), achieving over 20-point accuracy gains across both small and large language models. By leveraging knowledge distillation from a high-performing teacher model, LLMQuoter achieves competitive results in a resource-efficient fine-tuning setup. It democratizes advanced RAG capabilities, delivering significant performance improvements without requiring extensive model retraining. Our results highlight the potential of distilled quote-based reasoning to streamline complex workflows, offering a scalable and practical solution for researchers and practitioners alike.
Federated Learning (FL) provides novel solutions for machine learning (ML)-based lithography hotspot detection (LHD) under distributed privacy-preserving settings. Currently, two research pipelines have been investigated to aggregate local models and achieve global consensus, including parameter/nonparameter based (also known as knowledge distillation, namely KD). While these two kinds of methods show effectiveness in specific scenarios, we note they have not fully utilized and transferred the information learned, leaving the potential of FL-based LDH remains unexplored. Thus, we propose FedKDhybrid in this study to mitigate the research gap. Specifically, FedKD-hybrid clients agree on several identical layers across all participants and a public dataset for achieving global consensus. During training, the trained local model will be evaluated on the public dataset, and the generated logits will be uploaded along with the identical layer parameters. The aggregated information is consequently used to update local models via the public dataset as a medium. We compare our proposed FedKD-hybrid with several state-of-the-art (SOTA) FL methods under ICCAD-2012 and FAB (real-world collected) datasets with different settings; the experimental results demonstrate the superior performance of the FedKD-hybrid algorithm. Our code is available at https://github.com/itsnotacie/NN-FedKD-hybrid
With the rapid development of 3D reconstruction technology, the widespread distribution of 3D data has become a future trend. While traditional visual data (such as images and videos) and NeRF-based formats already have mature techniques for copyright protection, steganographic techniques for the emerging 3D Gaussian Splatting (3D-GS) format have yet to be fully explored. To address this, we propose ConcealGS, an innovative method for embedding implicit information into 3D-GS. By introducing the knowledge distillation and gradient optimization strategy based on 3D-GS, ConcealGS overcomes the limitations of NeRF-based models and enhances the robustness of implicit information and the quality of 3D reconstruction. We evaluate ConcealGS in various potential application scenarios, and experimental results have demonstrated that ConcealGS not only successfully recovers implicit information but also has almost no impact on rendering quality, providing a new approach for embedding invisible and recoverable information into 3D models in the future.
Math Word Problem (MWP) solving is a critical task in natural language processing, has garnered significant research interest in recent years. Various recent studies heavily rely on Seq2Seq models and their extensions (e.g., Seq2Tree and Graph2Tree) to generate mathematical equations. While effective, these models struggle to generate diverse but counterpart solution equations, limiting their generalization across various math problem scenarios. In this paper, we introduce a novel Diversity-enhanced Knowledge Distillation (DivKD) model for practical MWP solving. Our approach proposes an adaptive diversity distillation method, in which a student model learns diverse equations by selectively transferring high-quality knowledge from a teacher model. Additionally, we design a diversity prior-enhanced student model to better capture the diversity distribution of equations by incorporating a conditional variational auto-encoder. Extensive experiments on {four} MWP benchmark datasets demonstrate that our approach achieves higher answer accuracy than strong baselines while maintaining high efficiency for practical applications.
The emergence of 5G and edge computing hardware has brought about a significant shift in artificial intelligence, with edge AI becoming a crucial technology for enabling intelligent applications. With the growing amount of data generated and stored on edge devices, deploying AI models for local processing and inference has become increasingly necessary. However, deploying state-of-the-art AI models on resource-constrained edge devices faces significant challenges that must be addressed. This paper presents an optimization triad for efficient and reliable edge AI deployment, including data, model, and system optimization. First, we discuss optimizing data through data cleaning, compression, and augmentation to make it more suitable for edge deployment. Second, we explore model design and compression methods at the model level, such as pruning, quantization, and knowledge distillation. Finally, we introduce system optimization techniques like framework support and hardware acceleration to accelerate edge AI workflows. Based on an in-depth analysis of various application scenarios and deployment challenges of edge AI, this paper proposes an optimization paradigm based on the data-model-system triad to enable a whole set of solutions to effectively transfer ML models, which are initially trained in the cloud, to various edge devices for supporting multiple scenarios.
Graph neural networks (GNNs) have demonstrated superior performance in collaborative recommendation through their ability to conduct high-order representation smoothing, effectively capturing structural information within users' interaction patterns. However, existing GNN paradigms face significant challenges in scalability and robustness when handling large-scale, noisy, and real-world datasets. To address these challenges, we present LightGNN, a lightweight and distillation-based GNN pruning framework designed to substantially reduce model complexity while preserving essential collaboration modeling capabilities. Our LightGNN framework introduces a computationally efficient pruning module that adaptively identifies and removes redundant edges and embedding entries for model compression. The framework is guided by a resource-friendly hierarchical knowledge distillation objective, whose intermediate layer augments the observed graph to maintain performance, particularly in high-rate compression scenarios. Extensive experiments on public datasets demonstrate LightGNN's effectiveness, significantly improving both computational efficiency and recommendation accuracy. Notably, LightGNN achieves an 80% reduction in edge count and 90% reduction in embedding entries while maintaining performance comparable to more complex state-of-the-art baselines. The implementation of our LightGNN framework is available at the github repository: https://github.com/HKUDS/LightGNN.
In V2X collaborative perception, the domain gaps between heterogeneous nodes pose a significant challenge for effective information fusion. Pose errors arising from latency and GPS localization noise further exacerbate the issue by leading to feature misalignment. To overcome these challenges, we propose V2X-DGPE, a high-accuracy and robust V2X feature-level collaborative perception framework. V2X-DGPE employs a Knowledge Distillation Framework and a Feature Compensation Module to learn domain-invariant representations from multi-source data, effectively reducing the feature distribution gap between vehicles and roadside infrastructure. Historical information is utilized to provide the model with a more comprehensive understanding of the current scene. Furthermore, a Collaborative Fusion Module leverages a heterogeneous self-attention mechanism to extract and integrate heterogeneous representations from vehicles and infrastructure. To address pose errors, V2X-DGPE introduces a deformable attention mechanism, enabling the model to adaptively focus on critical parts of the input features by dynamically offsetting sampling points. Extensive experiments on the real-world DAIR-V2X dataset demonstrate that the proposed method outperforms existing approaches, achieving state-of-the-art detection performance. The code is available at https://github.com/wangsch10/V2X-DGPE.
Previous debiasing studies utilize unbiased data to make supervision of model training. They suffer from the high trial risks and experimental costs to obtain unbiased data. Recent research attempts to use invariant learning to detach the invariant preference of users for unbiased recommendations in an unsupervised way. However, it faces the drawbacks of low model accuracy and unstable prediction performance due to the losing cooperation with variant preference. In this paper, we experimentally demonstrate that invariant learning causes information loss by directly discarding the variant information, which reduces the generalization ability and results in the degradation of model performance in unbiased recommendations. Based on this consideration, we propose a novel lightweight knowledge distillation framework (KDDebias) to automatically learn the unbiased preference of users from both invariant and variant information. Specifically, the variant information is imputed to the invariant user preference in the distance-aware knowledge distillation process. Extensive experiments on three public datasets, i.e., Yahoo!R3, Coat, and MIND, show that with the biased imputation from the variant preference of users, our proposed method achieves significant improvements with less than 50% learning parameters compared to the SOTA unsupervised debiasing model in recommender systems. Our code are publicly available at https://github.com/BAI-LAB/KD-Debias.
Visual tracking has made significant strides due to the adoption of transformer-based models. Most state-of-the-art trackers struggle to meet real-time processing demands on mobile platforms with constrained computing resources, particularly for real-time unmanned aerial vehicle (UAV) tracking. To achieve a better balance between performance and efficiency, we introduce AVTrack, an adaptive computation framework designed to selectively activate transformer blocks for real-time UAV tracking. The proposed Activation Module (AM) dynamically optimizes the ViT architecture by selectively engaging relevant components, thereby enhancing inference efficiency without significant compromise to tracking performance. Furthermore, to tackle the challenges posed by extreme changes in viewing angles often encountered in UAV tracking, the proposed method enhances ViTs' effectiveness by learning view-invariant representations through mutual information (MI) maximization. Two effective design principles are proposed in the AVTrack. Building on it, we propose an improved tracker, dubbed AVTrack-MD, which introduces the novel MI maximization-based multi-teacher knowledge distillation (MD) framework. It harnesses the benefits of multiple teachers, specifically the off-the-shelf tracking models from the AVTrack, by integrating and refining their outputs, thereby guiding the learning process of the compact student network. Specifically, we maximize the MI between the softened feature representations from the multi-teacher models and the student model, leading to improved generalization and performance of the student model, particularly in noisy conditions. Extensive experiments on multiple UAV tracking benchmarks demonstrate that AVTrack-MD not only achieves performance comparable to the AVTrack baseline but also reduces model complexity, resulting in a significant 17\% increase in average tracking speed.
Knowledge Distillation (KD) has achieved widespread success in compressing large Vision Transformers (ViTs), but a unified theoretical framework for both ViTs and KD is still lacking. In this paper, we propose SpectralKD, a novel unified analytical framework that offers deeper insights into ViTs and optimizes KD via spectral analysis. Our model-wise analysis reveals that CaiT concentrates information in their first and last few layers, informing optimal layer selection for KD. Surprisingly, our layer-wise analysis discovers that Swin Transformer and CaiT exhibit similar spectral encoding patterns despite their architectural differences, leading to feature map alignment guideline. Building on these insights, we propose a simple yet effective spectral alignment method for KD. Benefiting from the deeper understanding by above analysis results, even such a simple strategy achieves state-of-the-art performance on ImageNet-1K without introducing any trainable parameters, improving DeiT-Tiny by $+5.2\%$ and Swin-Tiny by $+1.4\%$ in top-1 accuracy. Furthermore, our post-training analysis reveals that distilled students can reproduce spectral patterns similar to their teachers, opening a new area we term ``distillation dynamics". Code and experimental logs are available in https://github.com/thy960112/SpectralKD.
Geographic health disparities pose a pressing global challenge, particularly in underserved regions of low- and middle-income nations. Addressing this issue requires a collaborative approach to enhance healthcare quality, leveraging support from medically more developed areas. Federated learning emerges as a promising tool for this purpose. However, the scarcity of medical data and limited computation resources in underserved regions make collaborative training of powerful machine learning models challenging. Furthermore, there exists an asymmetrical reciprocity between underserved and developed regions. To overcome these challenges, we propose a novel cross-silo federated learning framework, named FedHelp, aimed at alleviating geographic health disparities and fortifying the diagnostic capabilities of underserved regions. Specifically, FedHelp leverages foundational model knowledge via one-time API access to guide the learning process of underserved small clients, addressing the challenge of insufficient data. Additionally, we introduce a novel asymmetric dual knowledge distillation module to manage the issue of asymmetric reciprocity, facilitating the exchange of necessary knowledge between developed large clients and underserved small clients. We validate the effectiveness and utility of FedHelp through extensive experiments on both medical image classification and segmentation tasks. The experimental results demonstrate significant performance improvement compared to state-of-the-art baselines, particularly benefiting clients in underserved regions.