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.
Advances in medical imaging and deep learning have propelled progress in whole slide image (WSI) analysis, with multiple instance learning (MIL) showing promise for efficient and accurate diagnostics. However, conventional MIL models often lack adaptability to evolving datasets, as they rely on static training that cannot incorporate new information without extensive retraining. Applying continual learning (CL) to MIL models is a possible solution, but often sees limited improvements. In this paper, we analyze CL in the context of attention MIL models and find that the model forgetting is mainly concentrated in the attention layers of the MIL model. Using the results of this analysis we propose two components for improving CL on MIL: Attention Knowledge Distillation (AKD) and the Pseudo-Bag Memory Pool (PMP). AKD mitigates catastrophic forgetting by focusing on retaining attention layer knowledge between learning sessions, while PMP reduces the memory footprint by selectively storing only the most informative patches, or ``pseudo-bags'' from WSIs. Experimental evaluations demonstrate that our method significantly improves both accuracy and memory efficiency on diverse WSI datasets, outperforming current state-of-the-art CL methods. This work provides a foundation for CL in large-scale, weakly annotated clinical datasets, paving the way for more adaptable and resilient diagnostic models.
Lung cancer is a leading cause of cancer-related deaths globally, where early detection and accurate diagnosis are critical for improving survival rates. While deep learning, particularly convolutional neural networks (CNNs), has revolutionized medical image analysis by detecting subtle patterns indicative of early-stage lung cancer, its adoption faces challenges. These models are often computationally expensive and require significant resources, making them unsuitable for resource constrained environments. Additionally, their lack of transparency hinders trust and broader adoption in sensitive fields like healthcare. Knowledge distillation addresses these challenges by transferring knowledge from large, complex models (teachers) to smaller, lightweight models (students). We propose a knowledge distillation-based approach for lung cancer detection, incorporating explainable AI (XAI) techniques to enhance model transparency. Eight CNNs, including ResNet50, EfficientNetB0, EfficientNetB3, and VGG16, are evaluated as teacher models. We developed and trained a lightweight student model, Distilled Custom Student Network (DCSNet) using ResNet50 as the teacher. This approach not only ensures high diagnostic performance in resource-constrained settings but also addresses transparency concerns, facilitating the adoption of AI-driven diagnostic tools in healthcare.
Large language models (LLMs) have transformed natural-language processing, yet their scale makes real-world deployment costly. Post-training quantization reduces memory and computation but often degrades accuracy, while quantization-aware training can recover performance at the cost of extra training. Pushing quantization to the ternary (2-bit) regime yields even larger savings but is notoriously unstable. Building on recent work showing that a bias-free, RMS-normalized Transformer with straight-through estimation can reach 1.58-bit precision, we demonstrate that simply inserting RMS normalization before every linear projection and applying a gradual, layer-wise quantization schedule stably fine-tunes full-precision checkpoints into ternary LLMs. Our approach matches or surpasses more elaborate knowledge-distillation pipelines on standard language-modeling benchmarks without adding model complexity. These results indicate that careful normalization alone can close much of the accuracy gap between ternary and full-precision LLMs, making ultra-low-bit inference practical.
The widespread adoption of Large Language Models (LLMs) for code generation, exemplified by GitHub Copilot\footnote{A coding extension powered by a Code-LLM to assist in code completion tasks} surpassing a million users, highlights the transformative potential of these tools in improving developer productivity. However, this rapid growth also underscores critical concerns regarding the quality, safety, and reliability of the code they generate. As Code-LLMs evolve, they face significant challenges, including the diminishing returns of model scaling and the scarcity of new, high-quality training data. To address these issues, this paper introduces Adversarial Knowledge Distillation (AKD), a novel approach that leverages adversarially generated synthetic datasets to distill the capabilities of larger models into smaller, more efficient ones. By systematically stress-testing and refining the reasoning capabilities of Code-LLMs, AKD provides a framework for enhancing model robustness, reliability, and security while improving their parameter-efficiency. We believe this work represents a critical step toward ensuring dependable automated code generation within the constraints of existing data and the cost-efficiency of model execution.
Accurate channel state information (CSI) acquisition for massive multiple-input multiple-output (MIMO) systems is essential for future mobile communication networks. Channel fingerprint (CF), also referred to as channel knowledge map, is a key enabler for intelligent environment-aware communication and can facilitate CSI acquisition. However, due to the cost limitations of practical sensing nodes and test vehicles, the resulting CF is typically coarse-grained, making it insufficient for wireless transceiver design. In this work, we introduce the concept of CF twins and design a conditional generative diffusion model (CGDM) with strong implicit prior learning capabilities as the computational core of the CF twin to establish the connection between coarse- and fine-grained CFs. Specifically, we employ a variational inference technique to derive the evidence lower bound (ELBO) for the log-marginal distribution of the observed fine-grained CF conditioned on the coarse-grained CF, enabling the CGDM to learn the complicated distribution of the target data. During the denoising neural network optimization, the coarse-grained CF is introduced as side information to accurately guide the conditioned generation of the CGDM. To make the proposed CGDM lightweight, we further leverage the additivity of network layers and introduce a one-shot pruning approach along with a multi-objective knowledge distillation technique. Experimental results show that the proposed approach exhibits significant improvement in reconstruction performance compared to the baselines. Additionally, zero-shot testing on reconstruction tasks with different magnification factors further demonstrates the scalability and generalization ability of the proposed approach.
Compact models can be effectively trained through Knowledge Distillation (KD), a technique that transfers knowledge from larger, high-performing teacher models. Two key challenges in Knowledge Distillation (KD) are: 1) balancing learning from the teacher's guidance and the task objective, and 2) handling the disparity in knowledge representation between teacher and student models. To address these, we propose Multi-Task Optimization for Knowledge Distillation (MoKD). MoKD tackles two main gradient issues: a) Gradient Conflicts, where task-specific and distillation gradients are misaligned, and b) Gradient Dominance, where one objective's gradient dominates, causing imbalance. MoKD reformulates KD as a multi-objective optimization problem, enabling better balance between objectives. Additionally, it introduces a subspace learning framework to project feature representations into a high-dimensional space, improving knowledge transfer. Our MoKD is demonstrated to outperform existing methods through extensive experiments on image classification using the ImageNet-1K dataset and object detection using the COCO dataset, achieving state-of-the-art performance with greater efficiency. To the best of our knowledge, MoKD models also achieve state-of-the-art performance compared to models trained from scratch.
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
Continual learning (CL) aims to train models that can learn a sequence of tasks without forgetting previously acquired knowledge. A core challenge in CL is balancing stability -- preserving performance on old tasks -- and plasticity -- adapting to new ones. Recently, large pre-trained models have been widely adopted in CL for their ability to support both, offering strong generalization for new tasks and resilience against forgetting. However, their high computational cost at inference time limits their practicality in real-world applications, especially those requiring low latency or energy efficiency. To address this issue, we explore model compression techniques, including pruning and knowledge distillation (KD), and propose two efficient frameworks tailored for class-incremental learning (CIL), a challenging CL setting where task identities are unavailable during inference. The pruning-based framework includes pre- and post-pruning strategies that apply compression at different training stages. The KD-based framework adopts a teacher-student architecture, where a large pre-trained teacher transfers downstream-relevant knowledge to a compact student. Extensive experiments on multiple CIL benchmarks demonstrate that the proposed frameworks achieve a better trade-off between accuracy and inference complexity, consistently outperforming strong baselines. We further analyze the trade-offs between the two frameworks in terms of accuracy and efficiency, offering insights into their use across different scenarios.
The rapid development of large language models (LLMs) has provided significant support and opportunities for the advancement of domain-specific LLMs. However, fine-tuning these large models using Intangible Cultural Heritage (ICH) data inevitably faces challenges such as bias, incorrect knowledge inheritance, and catastrophic forgetting. To address these issues, we propose a novel training method that integrates a bidirectional chains of thought and a reward mechanism. This method is built upon ICH-Qwen, a large language model specifically designed for the field of intangible cultural heritage. The proposed method enables the model to not only perform forward reasoning but also enhances the accuracy of the generated answers by utilizing reverse questioning and reverse reasoning to activate the model's latent knowledge. Additionally, a reward mechanism is introduced during training to optimize the decision-making process. This mechanism improves the quality of the model's outputs through structural and content evaluations with different weighting schemes. We conduct comparative experiments on ICH-Qwen, with results demonstrating that our method outperforms 0-shot, step-by-step reasoning, knowledge distillation, and question augmentation methods in terms of accuracy, Bleu-4, and Rouge-L scores on the question-answering task. Furthermore, the paper highlights the effectiveness of combining the bidirectional chains of thought and reward mechanism through ablation experiments. In addition, a series of generalizability experiments are conducted, with results showing that the proposed method yields improvements on various domain-specific datasets and advanced models in areas such as Finance, Wikidata, and StrategyQA. This demonstrates that the method is adaptable to multiple domains and provides a valuable approach for model training in future applications across diverse fields.
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.
The increasing volume of healthcare textual data requires computationally efficient, yet highly accurate classification approaches able to handle the nuanced and complex nature of medical terminology. This research presents Knowledge Distillation for Healthcare Multi-Label Text Classification (KDH-MLTC), a framework leveraging model compression and Large Language Models (LLMs). The proposed approach addresses conventional healthcare Multi-Label Text Classification (MLTC) challenges by integrating knowledge distillation and sequential fine-tuning, subsequently optimized through Particle Swarm Optimization (PSO) for hyperparameter tuning. KDH-MLTC transfers knowledge from a more complex teacher LLM (i.e., BERT) to a lighter student LLM (i.e., DistilBERT) through sequential training adapted to MLTC that preserves the teacher's learned information while significantly reducing computational requirements. As a result, the classification is enabled to be conducted locally, making it suitable for healthcare textual data characterized by sensitivity and, therefore, ensuring HIPAA compliance. The experiments conducted on three medical literature datasets of different sizes, sampled from the Hallmark of Cancer (HoC) dataset, demonstrate that KDH-MLTC achieves superior performance compared to existing approaches, particularly for the largest dataset, reaching an F1 score of 82.70%. Additionally, statistical validation and an ablation study are carried out, proving the robustness of KDH-MLTC. Furthermore, the PSO-based hyperparameter optimization process allowed the identification of optimal configurations. The proposed approach contributes to healthcare text classification research, balancing efficiency requirements in resource-constrained healthcare settings with satisfactory accuracy demands.
Ensuring the products displayed in e-commerce search results are relevant to users queries is crucial for improving the user experience. With their advanced semantic understanding, deep learning models have been widely used for relevance matching in search tasks. While large language models (LLMs) offer superior ranking capabilities, it is challenging to deploy LLMs in real-time systems due to the high-latency requirements. To leverage the ranking power of LLMs while meeting the low-latency demands of production systems, we propose a novel framework that distills a high performing LLM into a more efficient, low-latency student model. To help the student model learn more effectively from the teacher model, we first train the teacher LLM as a classification model with soft targets. Then, we train the student model to capture the relevance margin between pairs of products for a given query using mean squared error loss. Instead of using the same training data as the teacher model, we significantly expand the student model dataset by generating unlabeled data and labeling it with the teacher model predictions. Experimental results show that the student model performance continues to improve as the size of the augmented training data increases. In fact, with enough augmented data, the student model can outperform the teacher model. The student model has been successfully deployed in production at Walmart.com with significantly positive metrics.
Medical disease prediction, particularly through imaging, remains a challenging task due to the complexity and variability of medical data, including noise, ambiguity, and differing image quality. Recent deep learning models, including Knowledge Distillation (KD) methods, have shown promising results in brain tumor image identification but still face limitations in handling uncertainty and generalizing across diverse medical conditions. Traditional KD methods often rely on a context-unaware temperature parameter to soften teacher model predictions, which does not adapt effectively to varying uncertainty levels present in medical images. To address this issue, we propose a novel framework that integrates Ant Colony Optimization (ACO) for optimal teacher-student model selection and a novel context-aware predictor approach for temperature scaling. The proposed context-aware framework adjusts the temperature based on factors such as image quality, disease complexity, and teacher model confidence, allowing for more robust knowledge transfer. Additionally, ACO efficiently selects the most appropriate teacher-student model pair from a set of pre-trained models, outperforming current optimization methods by exploring a broader solution space and better handling complex, non-linear relationships within the data. The proposed framework is evaluated using three publicly available benchmark datasets, each corresponding to a distinct medical imaging task. The results demonstrate that the proposed framework significantly outperforms current state-of-the-art methods, achieving top accuracy rates: 98.01% on the MRI brain tumor (Kaggle) dataset, 92.81% on the Figshare MRI dataset, and 96.20% on the GastroNet dataset. This enhanced performance is further evidenced by the improved results, surpassing existing benchmarks of 97.24% (Kaggle), 91.43% (Figshare), and 95.00% (GastroNet).
Place recognition plays a significant role in SLAM, robot navigation, and autonomous driving applications. Benefiting from deep learning, the performance of LiDAR place recognition (LPR) has been greatly improved. However, many existing learning-based LPR methods suffer from catastrophic forgetting, which severely harms the performance of LPR on previously trained places after training on a new environment. In this paper, we introduce a continual learning framework for LPR via Knowledge Distillation and Fusion (KDF) to alleviate forgetting. Inspired by the ranking process of place recognition retrieval, we present a ranking-aware knowledge distillation loss that encourages the network to preserve the high-level place recognition knowledge. We also introduce a knowledge fusion module to integrate the knowledge of old and new models for LiDAR place recognition. Our extensive experiments demonstrate that KDF can be applied to different networks to overcome catastrophic forgetting, surpassing the state-of-the-art methods in terms of mean Recall@1 and forgetting score.
Latent space representations are critical for understanding and improving the behavior of machine learning models, yet they often remain obscure and intricate. Understanding and exploring the latent space has the potential to contribute valuable human intuition and expertise about respective domains. In this work, we present HILL, an interactive framework allowing users to incorporate human intuition into the model training by interactively reshaping latent space representations. The modifications are infused into the model training loop via a novel approach inspired by knowledge distillation, treating the user's modifications as a teacher to guide the model in reshaping its intrinsic latent representation. The process allows the model to converge more effectively and overcome inefficiencies, as well as provide beneficial insights to the user. We evaluated HILL in a user study tasking participants to train an optimal model, closely observing the employed strategies. The results demonstrated that human-guided latent space modifications enhance model performance while maintaining generalization, yet also revealing the risks of including user biases. Our work introduces a novel human-AI interaction paradigm that infuses human intuition into model training and critically examines the impact of human intervention on training strategies and potential biases.
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.
Large language models (LLMs) have achieved unprecedented performance by leveraging vast pretraining corpora, yet their performance remains suboptimal in knowledge-intensive domains such as medicine and scientific research, where high factual precision is required. While synthetic data provides a promising avenue for augmenting domain knowledge, existing methods frequently generate redundant samples that do not align with the model's true knowledge gaps. To overcome this limitation, we propose a novel Structural Entropy-guided Knowledge Navigator (SENATOR) framework that addresses the intrinsic knowledge deficiencies of LLMs. Our approach employs the Structure Entropy (SE) metric to quantify uncertainty along knowledge graph paths and leverages Monte Carlo Tree Search (MCTS) to selectively explore regions where the model lacks domain-specific knowledge. Guided by these insights, the framework generates targeted synthetic data for supervised fine-tuning, enabling continuous self-improvement. Experimental results on LLaMA-3 and Qwen2 across multiple domain-specific benchmarks show that SENATOR effectively detects and repairs knowledge deficiencies, achieving notable performance improvements. The code and data for our methods and experiments are available at https://github.com/weiyifan1023/senator.
Watermarking has emerged as a critical technique for combating misinformation and protecting intellectual property in large language models (LLMs). A recent discovery, termed watermark radioactivity, reveals that watermarks embedded in teacher models can be inherited by student models through knowledge distillation. On the positive side, this inheritance allows for the detection of unauthorized knowledge distillation by identifying watermark traces in student models. However, the robustness of watermarks against scrubbing attacks and their unforgeability in the face of spoofing attacks under unauthorized knowledge distillation remain largely unexplored. Existing watermark attack methods either assume access to model internals or fail to simultaneously support both scrubbing and spoofing attacks. In this work, we propose Contrastive Decoding-Guided Knowledge Distillation (CDG-KD), a unified framework that enables bidirectional attacks under unauthorized knowledge distillation. Our approach employs contrastive decoding to extract corrupted or amplified watermark texts via comparing outputs from the student model and weakly watermarked references, followed by bidirectional distillation to train new student models capable of watermark removal and watermark forgery, respectively. Extensive experiments show that CDG-KD effectively performs attacks while preserving the general performance of the distilled model. Our findings underscore critical need for developing watermarking schemes that are robust and unforgeable.
Attribute bias in federated learning (FL) typically leads local models to optimize inconsistently due to the learning of non-causal associations, resulting degraded performance. Existing methods either use data augmentation for increasing sample diversity or knowledge distillation for learning invariant representations to address this problem. However, they lack a comprehensive analysis of the inference paths, and the interference from confounding factors limits their performance. To address these limitations, we propose the \underline{Fed}erated \underline{D}econfounding and \underline{D}ebiasing \underline{L}earning (FedDDL) method. It constructs a structured causal graph to analyze the model inference process, and performs backdoor adjustment to eliminate confounding paths. Specifically, we design an intra-client deconfounding learning module for computer vision tasks to decouple background and objects, generating counterfactual samples that establish a connection between the background and any label, which stops the model from using the background to infer the label. Moreover, we design an inter-client debiasing learning module to construct causal prototypes to reduce the proportion of the background in prototype components. Notably, it bridges the gap between heterogeneous representations via causal prototypical regularization. Extensive experiments on 2 benchmarking datasets demonstrate that \methodname{} significantly enhances the model capability to focus on main objects in unseen data, leading to 4.5\% higher Top-1 Accuracy on average over 9 state-of-the-art existing methods.