Large Language Models (LLMs) represent substantial intellectual and economic investments, yet their effectiveness can inadvertently facilitate model imitation via knowledge distillation (KD).In practical scenarios, competitors can distill proprietary LLM capabilities by simply observing publicly accessible outputs, akin to reverse-engineering a complex performance by observation alone. Existing protective methods like watermarking only identify imitation post-hoc, while other defenses assume the student model mimics the teacher's internal logits, rendering them ineffective against distillation purely from observed output text. This paper confronts the challenge of actively protecting LLMs within the realistic constraints of API-based access. We introduce an effective and efficient Defensive Output Generation (DOGe) strategy that subtly modifies the output behavior of an LLM. Its outputs remain accurate and useful for legitimate users, yet are designed to be misleading for distillation, significantly undermining imitation attempts. We achieve this by fine-tuning only the final linear layer of the teacher LLM with an adversarial loss. This targeted training approach anticipates and disrupts distillation attempts during inference time. Our experiments show that, while preserving or even improving the original performance of the teacher model, student models distilled from the defensively generated teacher outputs demonstrate catastrophically reduced performance, demonstrating our method's effectiveness as a practical safeguard against KD-based model imitation.
Cross-capability transfer is a key challenge in large language model (LLM) research, with applications in multi-task integration, model compression, and continual learning. Recent works like FuseLLM and FuseChat have demonstrated the potential of transferring multiple model capabilities to lightweight models, enhancing adaptability and efficiency, which motivates our investigation into more efficient cross-capability transfer methods. However, existing approaches primarily focus on small, homogeneous models, limiting their applicability. For large, heterogeneous models, knowledge distillation with full-parameter fine-tuning often overlooks the student model's intrinsic capacity and risks catastrophic forgetting, while PEFT methods struggle to effectively absorb knowledge from source LLMs. To address these issues, we introduce GraftLLM, a novel method that stores source model capabilities in a target model with SkillPack format. This approach preserves general capabilities, reduces parameter conflicts, and supports forget-free continual learning and model fusion. We employ a module-aware adaptive compression strategy to compress parameter updates, ensuring efficient storage while maintaining task-specific knowledge. The resulting SkillPack serves as a compact and transferable knowledge carrier, ideal for heterogeneous model fusion and continual learning. Experiments across various scenarios demonstrate that GraftLLM outperforms existing techniques in knowledge transfer, knowledge fusion, and forget-free learning, providing a scalable and efficient solution for cross-capability transfer. The code is publicly available at: https://github.com/duguodong7/GraftLLM.
White Light Imaging (WLI) and Narrow Band Imaging (NBI) are the two main colonoscopic modalities for polyp classification. While NBI, as optical chromoendoscopy, offers valuable vascular details, WLI remains the most common and often the only available modality in resource-limited settings. However, WLI-based methods typically underperform, limiting their clinical applicability. Existing approaches transfer knowledge from NBI to WLI through global feature alignment but often rely on cropped lesion regions, which are susceptible to detection errors and neglect contextual and subtle diagnostic cues. To address this, this paper proposes a novel holistic classification framework that leverages full-image diagnosis without requiring polyp localization. The key innovation lies in the Alignment-free Dense Distillation (ADD) module, which enables fine-grained cross-domain knowledge distillation regardless of misalignment between WLI and NBI images. Without resorting to explicit image alignment, ADD learns pixel-wise cross-domain affinities to establish correspondences between feature maps, guiding the distillation along the most relevant pixel connections. To further enhance distillation reliability, ADD incorporates Class Activation Mapping (CAM) to filter cross-domain affinities, ensuring the distillation path connects only those semantically consistent regions with equal contributions to polyp diagnosis. Extensive results on public and in-house datasets show that our method achieves state-of-the-art performance, relatively outperforming the other approaches by at least 2.5% and 16.2% in AUC, respectively. Code is available at: https://github.com/Huster-Hq/ADD.
Artificial intelligence and machine learning models deployed on edge devices, e.g., for quality control in Additive Manufacturing (AM), are frequently small in size. Such models usually have to deliver highly accurate results within a short time frame. Methods that are commonly employed in literature start out with larger trained models and try to reduce their memory and latency footprint by structural pruning, knowledge distillation, or quantization. It is, however, also possible to leverage hardware-aware Neural Architecture Search (NAS), an approach that seeks to systematically explore the architecture space to find optimized configurations. In this study, a hardware-aware NAS workflow is introduced that couples an edge device located in Belgium with a powerful High-Performance Computing system in Germany, to train possible architecture candidates as fast as possible while performing real-time latency measurements on the target hardware. The approach is verified on a use case in the AM domain, based on the open RAISE-LPBF dataset, achieving ~8.8 times faster inference speed while simultaneously enhancing model quality by a factor of ~1.35, compared to a human-designed baseline.
Knowledge graph-based retrieval-augmented generation seeks to mitigate hallucinations in Large Language Models (LLMs) caused by insufficient or outdated knowledge. However, existing methods often fail to fully exploit the prior knowledge embedded in knowledge graphs (KGs), particularly their structural information and explicit or implicit constraints. The former can enhance the faithfulness of LLMs' reasoning, while the latter can improve the reliability of response generation. Motivated by these, we propose a trustworthy reasoning framework, termed Deliberation over Priors (DP), which sufficiently utilizes the priors contained in KGs. Specifically, DP adopts a progressive knowledge distillation strategy that integrates structural priors into LLMs through a combination of supervised fine-tuning and Kahneman-Tversky optimization, thereby improving the faithfulness of relation path generation. Furthermore, our framework employs a reasoning-introspection strategy, which guides LLMs to perform refined reasoning verification based on extracted constraint priors, ensuring the reliability of response generation. Extensive experiments on three benchmark datasets demonstrate that DP achieves new state-of-the-art performance, especially a Hit@1 improvement of 13% on the ComplexWebQuestions dataset, and generates highly trustworthy responses. We also conduct various analyses to verify its flexibility and practicality. The code is available at https://github.com/reml-group/Deliberation-on-Priors.
Knowledge distillation (KD) is a key technique for compressing large language models into smaller ones while preserving performance. Despite the recent traction of KD research, its effectiveness for smaller language models (LMs) and the mechanisms driving knowledge transfer remain underexplored. In this work, we present the first large-scale empirical and statistical analysis of KD across models ranging from 0.5B to 7B parameters on 14 complex reasoning tasks in a zero-shot setting. Our findings reveal that KD can improve the average performance of smaller models by up to $10\%$, with a peak task specific gain of $22\%$, while providing only marginal benefits ($\sim 1.3\%$) for larger models. Surprisingly, teacher performance has a minimal impact on student outcomes, while teacher task expertise impacts KD effectiveness. A correlation study indicates that smaller LMs benefit more from KD, whereas larger LMs show diminished gains. Additionally, we uncover a misalignment between improvements in student performance and reasoning fidelity, suggesting that while KD enhances accuracy, it does not always maintain the structured decision-making processes of the teacher. Our ablation study further highlights the importance of teacher signals and logit smoothing in influencing students' performance after distillation. Overall, our study offers a comprehensive empirical and statistical assessment of KD, highlighting both its benefits and trade-offs when distilling knowledge from larger to smaller LMs.
Recent advances in knowledge distillation have emphasized the importance of decoupling different knowledge components. While existing methods utilize momentum mechanisms to separate task-oriented and distillation gradients, they overlook the inherent conflict between target-class and non-target-class knowledge flows. Furthermore, low-confidence dark knowledge in non-target classes introduces noisy signals that hinder effective knowledge transfer. To address these limitations, we propose DeepKD, a novel training framework that integrates dual-level decoupling with adaptive denoising. First, through theoretical analysis of gradient signal-to-noise ratio (GSNR) characteristics in task-oriented and non-task-oriented knowledge distillation, we design independent momentum updaters for each component to prevent mutual interference. We observe that the optimal momentum coefficients for task-oriented gradient (TOG), target-class gradient (TCG), and non-target-class gradient (NCG) should be positively related to their GSNR. Second, we introduce a dynamic top-k mask (DTM) mechanism that gradually increases K from a small initial value to incorporate more non-target classes as training progresses, following curriculum learning principles. The DTM jointly filters low-confidence logits from both teacher and student models, effectively purifying dark knowledge during early training. Extensive experiments on CIFAR-100, ImageNet, and MS-COCO demonstrate DeepKD's effectiveness. Our code is available at https://github.com/haiduo/DeepKD.
With recent breakthroughs in large-scale modeling, the Segment Anything Model (SAM) has demonstrated significant potential in a variety of visual applications. However, due to the lack of underwater domain expertise, SAM and its variants face performance limitations in end-to-end underwater instance segmentation tasks, while their higher computational requirements further hinder their application in underwater scenarios. To address this challenge, we propose a large-scale underwater instance segmentation dataset, UIIS10K, which includes 10,048 images with pixel-level annotations for 10 categories. Then, we introduce UWSAM, an efficient model designed for automatic and accurate segmentation of underwater instances. UWSAM efficiently distills knowledge from the SAM ViT-Huge image encoder into the smaller ViT-Small image encoder via the Mask GAT-based Underwater Knowledge Distillation (MG-UKD) method for effective visual representation learning. Furthermore, we design an End-to-end Underwater Prompt Generator (EUPG) for UWSAM, which automatically generates underwater prompts instead of explicitly providing foreground points or boxes as prompts, thus enabling the network to locate underwater instances accurately for efficient segmentation. Comprehensive experimental results show that our model is effective, achieving significant performance improvements over state-of-the-art methods on multiple underwater instance datasets. Datasets and codes are available at https://github.com/LiamLian0727/UIIS10K.
Knowledge distillation (KD) is a valuable technique for compressing large deep learning models into smaller, edge-suitable networks. However, conventional KD frameworks rely on pre-trained high-capacity teacher networks, which introduce significant challenges such as increased memory/storage requirements, additional training costs, and ambiguity in selecting an appropriate teacher for a given student model. Although a teacher-free distillation (self-distillation) has emerged as a promising alternative, many existing approaches still rely on architectural modifications or complex training procedures, which limit their generality and efficiency. To address these limitations, we propose a novel framework based on teacher-free distillation that operates using a single student network without any auxiliary components, architectural modifications, or additional learnable parameters. Our approach is built on a simple yet highly effective augmentation, called intra-class patch swap augmentation. This augmentation simulates a teacher-student dynamic within a single model by generating pairs of intra-class samples with varying confidence levels, and then applying instance-to-instance distillation to align their predictive distributions. Our method is conceptually simple, model-agnostic, and easy to implement, requiring only a single augmentation function. Extensive experiments across image classification, semantic segmentation, and object detection show that our method consistently outperforms both existing self-distillation baselines and conventional teacher-based KD approaches. These results suggest that the success of self-distillation could hinge on the design of the augmentation itself. Our codes are available at https://github.com/hchoi71/Intra-class-Patch-Swap.
Nowadays, Large Language Models (LLMs) are trained on huge datasets, some including sensitive information. This poses a serious privacy concern because privacy attacks such as Membership Inference Attacks (MIAs) may detect this sensitive information. While knowledge distillation compresses LLMs into efficient, smaller student models, its impact on privacy remains underexplored. In this paper, we investigate how knowledge distillation affects model robustness against MIA. We focus on two questions. First, how is private data protected in teacher and student models? Second, how can we strengthen privacy preservation against MIAs in knowledge distillation? Through comprehensive experiments, we show that while teacher and student models achieve similar overall MIA accuracy, teacher models better protect member data, the primary target of MIA, whereas student models better protect non-member data. To address this vulnerability in student models, we propose 5 privacy-preserving distillation methods and demonstrate that they successfully reduce student models' vulnerability to MIA, with ensembling further stabilizing the robustness, offering a reliable approach for distilling more secure and efficient student models. Our implementation source code is available at https://github.com/richardcui18/MIA_in_KD.
Multi-modal semantic segmentation (MMSS) faces significant challenges in real-world scenarios due to dynamic environments, sensor failures, and noise interference, creating a gap between theoretical models and practical performance. To address this, we propose a two-stage framework called RobustSeg, which enhances multi-modal robustness through two key components: the Hybrid Prototype Distillation Module (HPDM) and the Representation Regularization Module (RRM). In the first stage, RobustSeg pre-trains a multi-modal teacher model using complete modalities. In the second stage, a student model is trained with random modality dropout while learning from the teacher via HPDM and RRM. HPDM transforms features into compact prototypes, enabling cross-modal hybrid knowledge distillation and mitigating bias from missing modalities. RRM reduces representation discrepancies between the teacher and student by optimizing functional entropy through the log-Sobolev inequality. Extensive experiments on three public benchmarks demonstrate that RobustSeg outperforms previous state-of-the-art methods, achieving improvements of +2.76%, +4.56%, and +0.98%, respectively. Code is available at: https://github.com/RobustSeg/RobustSeg.
Monocular depth estimation is critical for applications such as autonomous driving and scene reconstruction. While existing methods perform well under normal scenarios, their performance declines in adverse weather, due to challenging domain shifts and difficulties in extracting scene information. To address this issue, we present a robust monocular depth estimation method called \textbf{ACDepth} from the perspective of high-quality training data generation and domain adaptation. Specifically, we introduce a one-step diffusion model for generating samples that simulate adverse weather conditions, constructing a multi-tuple degradation dataset during training. To ensure the quality of the generated degradation samples, we employ LoRA adapters to fine-tune the generation weights of diffusion model. Additionally, we integrate circular consistency loss and adversarial training to guarantee the fidelity and naturalness of the scene contents. Furthermore, we elaborate on a multi-granularity knowledge distillation strategy (MKD) that encourages the student network to absorb knowledge from both the teacher model and pretrained Depth Anything V2. This strategy guides the student model in learning degradation-agnostic scene information from various degradation inputs. In particular, we introduce an ordinal guidance distillation mechanism (OGD) that encourages the network to focus on uncertain regions through differential ranking, leading to a more precise depth estimation. Experimental results demonstrate that our ACDepth surpasses md4all-DD by 2.50\% for night scene and 2.61\% for rainy scene on the nuScenes dataset in terms of the absRel metric.
Existing tracking algorithms typically rely on low-frame-rate RGB cameras coupled with computationally intensive deep neural network architectures to achieve effective tracking. However, such frame-based methods inherently face challenges in achieving low-latency performance and often fail in resource-constrained environments. Visual object tracking using bio-inspired event cameras has emerged as a promising research direction in recent years, offering distinct advantages for low-latency applications. In this paper, we propose a novel Slow-Fast Tracking paradigm that flexibly adapts to different operational requirements, termed SFTrack. The proposed framework supports two complementary modes, i.e., a high-precision slow tracker for scenarios with sufficient computational resources, and an efficient fast tracker tailored for latency-aware, resource-constrained environments. Specifically, our framework first performs graph-based representation learning from high-temporal-resolution event streams, and then integrates the learned graph-structured information into two FlashAttention-based vision backbones, yielding the slow and fast trackers, respectively. The fast tracker achieves low latency through a lightweight network design and by producing multiple bounding box outputs in a single forward pass. Finally, we seamlessly combine both trackers via supervised fine-tuning and further enhance the fast tracker's performance through a knowledge distillation strategy. Extensive experiments on public benchmarks, including FE240, COESOT, and EventVOT, demonstrate the effectiveness and efficiency of our proposed method across different real-world scenarios. The source code has been released on https://github.com/Event-AHU/SlowFast_Event_Track.
Training high-performing Small Language Models (SLMs) remains costly, even with knowledge distillation and pruning from larger teacher models. Existing work often faces three key challenges: (1) information loss from hard pruning, (2) inefficient alignment of representations, and (3) underutilization of informative activations, particularly from Feed-Forward Networks (FFNs). To address these challenges, we introduce Low-Rank Clone (LRC), an efficient pre-training method that constructs SLMs aspiring to behavioral equivalence with strong teacher models. LRC trains a set of low-rank projection matrices that jointly enable soft pruning by compressing teacher weights, and activation clone by aligning student activations, including FFN signals, with those of the teacher. This unified design maximizes knowledge transfer while removing the need for explicit alignment modules. Extensive experiments with open-source teachers (e.g., Llama-3.2-3B-Instruct, Qwen2.5-3B/7B-Instruct) show that LRC matches or surpasses state-of-the-art models trained on trillions of tokens--while using only 20B tokens, achieving over 1,000x training efficiency. Our codes and model checkpoints are available at https://github.com/CURRENTF/LowRankClone and https://huggingface.co/collections/JitaiHao/low-rank-clone-lrc-6828389e96a93f1d4219dfaf.
Knowledge distillation (KD) is a powerful strategy for training deep neural networks (DNNs). Although it was originally proposed to train a more compact "student" model from a large "teacher" model, many recent efforts have focused on adapting it to promote generalization of the model itself, such as online KD and self KD. Here, we propose an accessible and compatible strategy named Spaced KD to improve the effectiveness of both online KD and self KD, in which the student model distills knowledge from a teacher model trained with a space interval ahead. This strategy is inspired by a prominent theory named spacing effect in biological learning and memory, positing that appropriate intervals between learning trials can significantly enhance learning performance. With both theoretical and empirical analyses, we demonstrate that the benefits of the proposed Spaced KD stem from convergence to a flatter loss landscape during stochastic gradient descent (SGD). We perform extensive experiments to validate the effectiveness of Spaced KD in improving the learning performance of DNNs (e.g., the performance gain is up to 2.31% and 3.34% on Tiny-ImageNet over online KD and self KD, respectively). Our codes have been released on github https://github.com/SunGL001/Spaced-KD.
Pre-trained vision-language models such as contrastive language-image pre-training (CLIP) have demonstrated a remarkable generalizability, which has enabled a wide range of applications represented by zero-shot classification. However, vision-language models still suffer when they face datasets with large gaps from training ones, i.e., distribution shifts. We found that CLIP is especially vulnerable to sensor degradation, a type of realistic distribution shift caused by sensor conditions such as weather, light, or noise. Collecting a new dataset from a test distribution for fine-tuning highly costs since sensor degradation occurs unexpectedly and has a range of variety. Thus, we investigate test-time adaptation (TTA) of zero-shot classification, which enables on-the-fly adaptation to the test distribution with unlabeled test data. Existing TTA methods for CLIP mainly focus on modifying image and text embeddings or predictions to address distribution shifts. Although these methods can adapt to domain shifts, such as fine-grained labels spaces or different renditions in input images, they fail to adapt to distribution shifts caused by sensor degradation. We found that this is because image embeddings are "corrupted" in terms of uniformity, a measure related to the amount of information. To make models robust to sensor degradation, we propose a novel method called uniformity-aware information-balanced TTA (UnInfo). To address the corruption of image embeddings, we introduce uniformity-aware confidence maximization, information-aware loss balancing, and knowledge distillation from the exponential moving average (EMA) teacher. Through experiments, we demonstrate that our UnInfo improves accuracy under sensor degradation by retaining information in terms of uniformity.
Knowledge distillation compresses a larger neural model (teacher) into smaller, faster student models by training them to match teacher outputs. However, the internal computational transformations that occur during this process remain poorly understood. We apply techniques from mechanistic interpretability to analyze how internal circuits, representations, and activation patterns differ between teacher and student. Focusing on GPT2-small and its distilled counterpart DistilGPT2, we find that student models reorganize, compress, and discard teacher components, often resulting in stronger reliance on fewer individual components. To quantify functional alignment beyond output similarity, we introduce an alignment metric based on influence-weighted component similarity, validated across multiple tasks. Our findings reveal that while knowledge distillation preserves broad functional behaviors, it also causes significant shifts in internal computation, with important implications for the robustness and generalization capacity of distilled models.
Point cloud processing has gained significant attention due to its critical role in applications such as autonomous driving and 3D object recognition. However, deploying high-performance models like Point Transformer V3 in resource-constrained environments remains challenging due to their high computational and memory demands. This work introduces a novel distillation framework that leverages topology-aware representations and gradient-guided knowledge distillation to effectively transfer knowledge from a high-capacity teacher to a lightweight student model. Our approach captures the underlying geometric structures of point clouds while selectively guiding the student model's learning process through gradient-based feature alignment. Experimental results in the Nuscenes, SemanticKITTI, and Waymo datasets demonstrate that the proposed method achieves competitive performance, with an approximately 16x reduction in model size and a nearly 1.9x decrease in inference time compared to its teacher model. Notably, on NuScenes, our method achieves state-of-the-art performance among knowledge distillation techniques trained solely on LiDAR data, surpassing prior knowledge distillation baselines in segmentation performance. Our implementation is available publicly at: https://github.com/HySonLab/PointDistill
Accurate estimation of lithium-ion battery capacity degradation is critical for enhancing the reliability and safety of battery operations. Traditional expert models, tailored to specific scenarios, provide isolated estimations. With the rapid advancement of data-driven techniques, a series of general-purpose time-series foundation models have been developed. However, foundation models specifically designed for battery capacity degradation remain largely unexplored. To enable zero-shot generalization in battery degradation prediction using large model technology, this study proposes a degradation-aware fine-tuning strategy for time-series foundation models. We apply this strategy to fine-tune the Timer model on approximately 10 GB of open-source battery charge discharge data. Validation on our released CycleLife-SJTUIE dataset demonstrates that the fine-tuned Battery-Timer possesses strong zero-shot generalization capability in capacity degradation forecasting. To address the computational challenges of deploying large models, we further propose a knowledge distillation framework that transfers the knowledge of pre-trained foundation models into compact expert models. Distillation results across several state-of-the-art time-series expert models confirm that foundation model knowledge significantly improves the multi-condition generalization of expert models.
Vision-language models (VLMs) have achieved remarkable success across diverse tasks by leveraging rich textual information with minimal labeled data. However, deploying such large models remains challenging, particularly in resource-constrained environments. Knowledge distillation (KD) offers a well-established solution to this problem; however, recent KD approaches from VLMs often involve multi-stage training or additional tuning, increasing computational overhead and optimization complexity. In this paper, we propose $\mathbf{\texttt{D}}$ual-$\mathbf{\texttt{H}}$ead $\mathbf{\texttt{O}}$ptimization ($\mathbf{\texttt{DHO}}$) -- a simple yet effective KD framework that transfers knowledge from VLMs to compact, task-specific models in semi-supervised settings. Specifically, we introduce dual prediction heads that independently learn from labeled data and teacher predictions, and propose to linearly combine their outputs during inference. We observe that $\texttt{DHO}$ mitigates gradient conflicts between supervised and distillation signals, enabling more effective feature learning than single-head KD baselines. As a result, extensive experiments show that $\texttt{DHO}$ consistently outperforms baselines across multiple domains and fine-grained datasets. Notably, on ImageNet, it achieves state-of-the-art performance, improving accuracy by 3% and 0.1% with 1% and 10% labeled data, respectively, while using fewer parameters.