Knowledge distillation is widely used to transfer behavior from a large teacher model to a smaller student. It is often assumed to be safe when the teacher is clean, because classic backdoor attacks rely on poisoned labels and triggers in supervised training, whereas distillation trains the student to match a teacher's outputs. We show that this assumption can fail when the distillation dataset itself is poisoned. Our attack injects triggered and manipulated images that a clean teacher already predicts as an attacker chosen target label, which causes the student to learn a backdoor even though the teacher remains unaffected. We evaluate this threat across multiple manipulation strategies, including targeted adversarial perturbations and targeted GAN based class transitions, and study how distillation settings influence both accuracy and attack success. We show that the attack remains effective even at a 10% poisoning rates. The results demonstrate that a clean teacher alone is not a sufficient safeguard: poisoned distillation data can produce a strongly backdoored student while maintaining competitive performance on clean images. These findings show that the integrity and provenance of distillation data are part of the security boundary of data intensive KD pipelines, even when the teacher itself is trusted.
Multiple Instance Learning is the predominant method for Whole Slide Image classification in digital pathology, enabling the use of slide-level labels to supervise model training. Although MIL eliminates the tedious fine-grained annotation process for supervised learning, whether it can learn accurate bag- and instance-level classifiers remains a question. To address the issue, instance-level classifiers and instance masks were incorporated to ground the prediction on supporting patches. These methods, while practically improving the performance of MIL methods, may potentially introduce noisy labels. We propose to bridge the gap between commonly used MIL and fully supervised learning by augmenting both the bag- and instance-level learning processes with pseudo-label correction capabilities elicited from weak to strong generalization techniques. The proposed algorithm improves the performance of dual-level MIL algorithms on both bag- and instance-level predictions. Experiments on public pathology datasets showcase the advantage of the proposed methods.
Image Quality Assessment (IQA) is a core task in computer vision. Multimodal methods based on vision-language models, such as CLIP, have demonstrated exceptional generalization capabilities in IQA tasks. To address the issues of excessive parameter burden and insufficient ability to identify local distorted features in CLIP for IQA, this study proposes a visual-language model knowledge distillation method aimed at guiding the training of models with architectural advantages using CLIP's IQA knowledge. First, quality-graded prompt templates were designed to guide CLIP to output quality scores. Then, CLIP is fine-tuned to enhance its capabilities in IQA tasks. Finally, a modality-adaptive knowledge distillation strategy is proposed to achieve guidance from the CLIP teacher model to the student model. Our experiments were conducted on multiple IQA datasets, and the results show that the proposed method significantly reduces model complexity while outperforming existing IQA methods, demonstrating strong potential for practical deployment.
Electroencephalography (EEG) is a fundamental modality for cognitive state monitoring in brain-computer interfaces (BCIs). However, it is highly susceptible to intrinsic signal errors and human-induced labeling errors, which lead to label noise and ultimately degrade model performance. To enhance EEG learning, multimodal knowledge distillation (KD) has been explored to transfer knowledge from visual models with rich representations to EEG-based models. Nevertheless, KD faces two key challenges: modality gap and soft label misalignment. The former arises from the heterogeneous nature of EEG and visual feature spaces, while the latter stems from label inconsistencies that create discrepancies between ground truth labels and distillation targets. This paper addresses semantic uncertainty caused by ambiguous features and weakly defined labels. We propose a novel cross-modal knowledge distillation framework that mitigates both modality and label inconsistencies. It aligns feature semantics through a prototype-based similarity module and introduces a task-specific distillation head to resolve label-induced inconsistency in supervision. Experimental results demonstrate that our approach improves EEG-based emotion regression and classification performance, outperforming both unimodal and multimodal baselines on a public multimodal dataset. These findings highlight the potential of our framework for BCI applications.
The landscape of video recognition has evolved significantly, shifting from traditional Convolutional Neural Networks (CNNs) to Transformer-based architectures for improved accuracy. While 3D CNNs have been effective at capturing spatiotemporal dynamics, recent Transformer models leverage self-attention to model long-range spatial and temporal dependencies. Despite achieving state-of-the-art performance on major benchmarks, Transformers remain computationally expensive, particularly with dense video data. To address this, we propose a lightweight Video Focal Modulation Network, DVFL-Net, which distills spatiotemporal knowledge from a large pre-trained teacher into a compact nano student model, enabling efficient on-device deployment. DVFL-Net utilizes knowledge distillation and spatial-temporal feature modulation to significantly reduce computation while preserving high recognition performance. We employ forward Kullback-Leibler (KL) divergence alongside spatio-temporal focal modulation to effectively transfer both local and global context from the Video-FocalNet Base (teacher) to the proposed VFL-Net (student). We evaluate DVFL-Net on UCF50, UCF101, HMDB51, SSV2, and Kinetics-400, benchmarking it against recent state-of-the-art methods in Human Action Recognition (HAR). Additionally, we conduct a detailed ablation study analyzing the impact of forward KL divergence. The results confirm the superiority of DVFL-Net in achieving an optimal balance between performance and efficiency, demonstrating lower memory usage, reduced GFLOPs, and strong accuracy, making it a practical solution for real-time HAR applications.
Large language models (LLMs) often show poor performance in low-resource languages like Korean, partly due to unique linguistic challenges such as homophonous Sino-Korean words that are indistinguishable in Hangul script. To address this semantic ambiguity, we propose HanjaBridge, a novel meaning-injection technique integrated into a continual pre-training (CPT) framework. Instead of deterministically mapping a word to a single Hanja (Chinese character), HanjaBridge presents the model with all possible Hanja candidates for a given homograph, encouraging the model to learn contextual disambiguation. This process is paired with token-level knowledge distillation to prevent catastrophic forgetting. Experimental results show that HanjaBridge significantly improves Korean language understanding, achieving a 21\% relative improvement on the KoBALT benchmark. Notably, by reinforcing semantic alignment between Korean and Chinese through shared Hanja, we observe a strong positive cross-lingual transfer. Furthermore, these gains persist even when Hanja augmentation is omitted at inference time, ensuring practical efficiency with no additional run-time cost.
Model-Heterogeneous Federated Learning (Hetero-FL) has attracted growing attention for its ability to aggregate knowledge from heterogeneous models while keeping private data locally. To better aggregate knowledge from clients, ensemble distillation, as a widely used and effective technique, is often employed after global aggregation to enhance the performance of the global model. However, simply combining Hetero-FL and ensemble distillation does not always yield promising results and can make the training process unstable. The reason is that existing methods primarily focus on logit distillation, which, while being model-agnostic with softmax predictions, fails to compensate for the knowledge bias arising from heterogeneous models. To tackle this challenge, we propose a stable and efficient Feature Distillation for model-heterogeneous Federated learning, dubbed FedFD, that can incorporate aligned feature information via orthogonal projection to integrate knowledge from heterogeneous models better. Specifically, a new feature-based ensemble federated knowledge distillation paradigm is proposed. The global model on the server needs to maintain a projection layer for each client-side model architecture to align the features separately. Orthogonal techniques are employed to re-parameterize the projection layer to mitigate knowledge bias from heterogeneous models and thus maximize the distilled knowledge. Extensive experiments show that FedFD achieves superior performance compared to state-of-the-art methods.
Collaborative fairness is a crucial challenge in federated learning. However, existing approaches often overlook a practical yet complex form of heterogeneity: imbalanced covariate shift. We provide a theoretical analysis of this setting, which motivates the design of FedAKD (Federated Asynchronous Knowledge Distillation)- simple yet effective approach that balances accurate prediction with collaborative fairness. FedAKD consists of client and server updates. In the client update, we introduce a novel asynchronous knowledge distillation strategy based on our preliminary analysis, which reveals that while correctly predicted samples exhibit similar feature distributions across clients, incorrectly predicted samples show significant variability. This suggests that imbalanced covariate shift primarily arises from misclassified samples. Leveraging this insight, our approach first applies traditional knowledge distillation to update client models while keeping the global model fixed. Next, we select correctly predicted high-confidence samples and update the global model using these samples while keeping client models fixed. The server update simply aggregates all client models. We further provide a theoretical proof of FedAKD's convergence. Experimental results on public datasets (FashionMNIST and CIFAR10) and a real-world Electronic Health Records (EHR) dataset demonstrate that FedAKD significantly improves collaborative fairness, enhances predictive accuracy, and fosters client participation even under highly heterogeneous data distributions.
Federated Learning (FL) is a distributed machine learning paradigm which coordinates multiple clients to collaboratively train a global model via a central server. Sequential Federated Learning (SFL) is a newly-emerging FL training framework where the global model is trained in a sequential manner across clients. Since SFL can provide strong convergence guarantees under data heterogeneity, it has attracted significant research attention in recent years. However, experiments show that SFL suffers from severe catastrophic forgetting in heterogeneous environments, meaning that the model tends to forget knowledge learned from previous clients. To address this issue, we propose an SFL framework with discrepancy-aware multi-teacher knowledge distillation, called SFedKD, which selects multiple models from the previous round to guide the current round of training. In SFedKD, we extend the single-teacher Decoupled Knowledge Distillation approach to our multi-teacher setting and assign distinct weights to teachers' target-class and non-target-class knowledge based on the class distributional discrepancy between teacher and student data. Through this fine-grained weighting strategy, SFedKD can enhance model training efficacy while mitigating catastrophic forgetting. Additionally, to prevent knowledge dilution, we eliminate redundant teachers for the knowledge distillation and formalize it as a variant of the maximum coverage problem. Based on the greedy strategy, we design a complementary-based teacher selection mechanism to ensure that the selected teachers achieve comprehensive knowledge space coverage while reducing communication and computational costs. Extensive experiments show that SFedKD effectively overcomes catastrophic forgetting in SFL and outperforms state-of-the-art FL methods.
This paper introduces a novel framework for designing efficient neural network architectures specifically tailored to tiny machine learning (TinyML) platforms. By leveraging large language models (LLMs) for neural architecture search (NAS), a vision transformer (ViT)-based knowledge distillation (KD) strategy, and an explainability module, the approach strikes an optimal balance between accuracy, computational efficiency, and memory usage. The LLM-guided search explores a hierarchical search space, refining candidate architectures through Pareto optimization based on accuracy, multiply-accumulate operations (MACs), and memory metrics. The best-performing architectures are further fine-tuned using logits-based KD with a pre-trained ViT-B/16 model, which enhances generalization without increasing model size. Evaluated on the CIFAR-100 dataset and deployed on an STM32H7 microcontroller (MCU), the three proposed models, LMaNet-Elite, LMaNet-Core, and QwNet-Core, achieve accuracy scores of 74.50%, 74.20% and 73.00%, respectively. All three models surpass current state-of-the-art (SOTA) models, such as MCUNet-in3/in4 (69.62% / 72.86%) and XiNet (72.27%), while maintaining a low computational cost of less than 100 million MACs and adhering to the stringent 320 KB static random-access memory (SRAM) constraint. These results demonstrate the efficiency and performance of the proposed framework for TinyML platforms, underscoring the potential of combining LLM-driven search, Pareto optimization, KD, and explainability to develop accurate, efficient, and interpretable models. This approach opens new possibilities in NAS, enabling the design of efficient architectures specifically suited for TinyML.
We present Kwaipilot-AutoThink (KAT), an open-source 40B large language model developed to address the overthinking problem in reasoning-intensive tasks, where an automatic thinking training paradigm is proposed to dynamically switch between reasoning and non-reasoning modes based on task complexity. Specifically, first, we construct the dual-regime dataset based on a novel tagging pipeline and a multi-agent synthesis strategy, and then we apply Multi-Token Prediction (MTP)-enhanced knowledge distillation, enabling efficient and fine-grained reasoning transfer with minimal pretraining cost. Besides, we implement a cold-start initialization strategy that introduces mode-selection priors using majority-vote signals and intent-aware prompting. Finally, we propose Step-SRPO, a reinforcement learning algorithm that incorporates intermediate supervision into the GRPO framework, offering structured guidance over both reasoning-mode selection and response accuracy. Extensive experiments across multiple benchmarks demonstrate that KAT consistently matches or even outperforms current state-of-the-art models, including DeepSeek-R1-0528 and Qwen3-235B-A22B, across a wide range of reasoning-intensive tasks while reducing token usage by up to approximately 30\%. Beyond academic evaluation, KAT has been successfully deployed in Kwaipilot (i.e., Kuaishou's internal coding assistant), and improves real-world development workflows with high accuracy, efficiency, and controllable reasoning behaviors. Moreover, we are actively training a 200B Mixture-of-Experts (MoE) with 40B activation parameters, where the early-stage results already demonstrate promising improvements in performance and efficiency, further showing the scalability of the AutoThink paradigm.
Large Language Models (LLMs) continue to set new standards in knowledge-intensive and complex reasoning tasks, yet their high computational demands limit widespread adoption. While distilling large models into smaller ones offers a sustainable solution, current techniques--such as static knowledge distillation, resource-intensive reinforcement learning from human feedback, or limited self-reflection--struggle to yield substantial and lasting performance gains. In this paper, we present a novel Debate and Reflect (D&R) framework that orchestrates multi-turn debates between smaller models and stronger teacher models, eliciting actionable feedback (e.g., error analysis, corrective strategies) to guide student models. Further, we introduce Tree-structured Direct Preference Optimization (T-DPO) to efficiently leverage these debate logs, organizing interactions into a hierarchical format for effective training. Empirical evaluations across diverse NLP benchmarks demonstrate that our approach significantly improves smaller-model accuracy, robustness, and generalization, outperforming conventional baselines by a large margin.
In this paper, we propose a novel approach for solving the Visual Question Answering (VQA) task in autonomous driving by integrating Vision-Language Models (VLMs) with continual learning. In autonomous driving, VQA plays a vital role in enabling the system to understand and reason about its surroundings. However, traditional models often struggle with catastrophic forgetting when sequentially exposed to new driving tasks, such as perception, prediction, and planning, each requiring different forms of knowledge. To address this challenge, we present a novel continual learning framework that combines VLMs with selective memory replay and knowledge distillation, reinforced by task-specific projection layer regularization. The knowledge distillation allows a previously trained model to act as a "teacher" to guide the model through subsequent tasks, minimizing forgetting. Meanwhile, task-specific projection layers calculate the loss based on the divergence of feature representations, ensuring continuity in learning and reducing the shift between tasks. Evaluated on the DriveLM dataset, our framework shows substantial performance improvements, with gains ranging from 21.40% to 32.28% across various metrics. These results highlight the effectiveness of combining continual learning with VLMs in enhancing the resilience and reliability of VQA systems in autonomous driving. We will release our source code.
We often attribute human characteristics to large language models (LLMs) and claim that they "know" certain things. LLMs have an internal probabilistic knowledge that represents information retained during training. How can we assess the veracity of this knowledge? We examine two common methods for probing the veracity of LLMs and discover several assumptions that are flawed. To address these flawed assumptions, we introduce sAwMIL (short for Sparse Aware Multiple-Instance Learning), a probing method that utilizes the internal activations of LLMs to separate statements into true, false, and neither. sAwMIL is based on multiple-instance learning and conformal prediction. We evaluate sAwMIL on 5 validity criteria across 16 open-source LLMs, including both default and chat-based variants, as well as on 3 new datasets. Among the insights we provide are: (1) the veracity signal is often concentrated in the third quarter of an LLM's depth; (2) truth and falsehood signals are not always symmetric; (3) linear probes perform better on chat models than on default models; (4) nonlinear probes may be required to capture veracity signals for some LLMs with reinforcement learning from human feedback or knowledge distillation; and (5) LLMs capture a third type of signal that is distinct from true and false and is neither true nor false. These findings provide a reliable method for verifying what LLMs "know" and how certain they are of their probabilistic internal knowledge.
Large language models can exhibit improved mathematical reasoning capabilities following post-training with instruction tuning, reinforcement learning, or knowledge distillation. However, it remains unclear whether these improvements are driven by major changes in transformer layers or from minor adjustments that leave the relative layer importance structures of the base model largely unchanged. We investigate this question through systematic layer-wise ablation experiments, examining base, instruction-tuned, knowledge-distilled, and reinforcement learning variants on mathematical reasoning benchmarks. Our findings show that mathematical reasoning gives rise to a specific layer importance structure, and this structure persists across all post-training paradigms. Removal of such layers causes accuracy drops of up to 80%. In contrast, non-mathematical tasks like factual recall exhibit no critical layers. This distinction suggests that mathematical reasoning requires specialized layers that emerge during pre-training, while other non-reasoning tasks do not. From an information-theoretic perspective, we also observe that these critical layers are the same layers where major representational transformation occurs.
Real-time, on-device segmentation is critical for latency-sensitive and privacy-aware applications like smart glasses and IoT devices. We introduce PicoSAM2, a lightweight (1.3M parameters, 336M MACs) promptable segmentation model optimized for edge and in-sensor execution, including the Sony IMX500. It builds on a depthwise separable U-Net, with knowledge distillation and fixed-point prompt encoding to learn from the Segment Anything Model 2 (SAM2). On COCO and LVIS, it achieves 51.9% and 44.9% mIoU, respectively. The quantized model (1.22MB) runs at 14.3 ms on the IMX500-achieving 86 MACs/cycle, making it the only model meeting both memory and compute constraints for in-sensor deployment. Distillation boosts LVIS performance by +3.5% mIoU and +5.1% mAP. These results demonstrate that efficient, promptable segmentation is feasible directly on-camera, enabling privacy-preserving vision without cloud or host processing.
Visual SLAM is particularly challenging in environments affected by noise, varying lighting conditions, and darkness. Learning-based optical flow algorithms can leverage multiple modalities to address these challenges, but traditional optical flow-based visual SLAM approaches often require significant computational resources.To overcome this limitation, we propose FMF-SLAM, an efficient multimodal fusion SLAM method that utilizes fast Fourier transform (FFT) to enhance the algorithm efficiency. Specifically, we introduce a novel Fourier-based self-attention and cross-attention mechanism to extract features from RGB and depth signals. We further enhance the interaction of multimodal features by incorporating multi-scale knowledge distillation across modalities. We also demonstrate the practical feasibility of FMF-SLAM in real-world scenarios with real time performance by integrating it with a security robot by fusing with a global positioning module GNSS-RTK and global Bundle Adjustment. Our approach is validated using video sequences from TUM, TartanAir, and our real-world datasets, showcasing state-of-the-art performance under noisy, varying lighting, and dark conditions.Our code and datasets are available at https://github.com/youjie-zhou/FMF-SLAM.git.
Explicit density learners are becoming an increasingly popular technique for generative models because of their ability to better model probability distributions. They have advantages over Generative Adversarial Networks due to their ability to perform density estimation and having exact latent-variable inference. This has many advantages, including: being able to simply interpolate, calculate sample likelihood, and analyze the probability distribution. The downside of these models is that they are often more difficult to train and have lower sampling quality. Normalizing flows are explicit density models, that use composable bijective functions to turn an intractable probability function into a tractable one. In this work, we present novel knowledge distillation techniques to increase sampling quality and density estimation of smaller student normalizing flows. We seek to study the capacity of knowledge distillation in Compositional Normalizing Flows to understand the benefits and weaknesses provided by these architectures. Normalizing flows have unique properties that allow for a non-traditional forms of knowledge transfer, where we can transfer that knowledge within intermediate layers. We find that through this distillation, we can make students significantly smaller while making substantial performance gains over a non-distilled student. With smaller models there is a proportionally increased throughput as this is dependent upon the number of bijectors, and thus parameters, in the network.
Multimodal learning aims to leverage information from diverse data modalities to achieve more comprehensive performance. However, conventional multimodal models often suffer from modality imbalance, where one or a few modalities dominate model optimization, leading to suboptimal feature representation and underutilization of weak modalities. To address this challenge, we introduce Gradient-Guided Distillation (G$^{2}$D), a knowledge distillation framework that optimizes the multimodal model with a custom-built loss function that fuses both unimodal and multimodal objectives. G$^{2}$D further incorporates a dynamic sequential modality prioritization (SMP) technique in the learning process to ensure each modality leads the learning process, avoiding the pitfall of stronger modalities overshadowing weaker ones. We validate G$^{2}$D on multiple real-world datasets and show that G$^{2}$D amplifies the significance of weak modalities while training and outperforms state-of-the-art methods in classification and regression tasks. Our code is available at https://github.com/rAIson-Lab/G2D.
The development of continual learning (CL) methods, which aim to learn new tasks in a sequential manner from the training data acquired continuously, has gained great attention in remote sensing (RS). The existing CL methods in RS, while learning new tasks, enhance robustness towards catastrophic forgetting. This is achieved by using a large number of labeled training samples, which is costly and not always feasible to gather in RS. To address this problem, we propose a novel continual self-supervised learning method in the context of masked autoencoders (denoted as CoSMAE). The proposed CoSMAE consists of two components: i) data mixup; and ii) model mixup knowledge distillation. Data mixup is associated with retaining information on previous data distributions by interpolating images from the current task with those from the previous tasks. Model mixup knowledge distillation is associated with distilling knowledge from past models and the current model simultaneously by interpolating their model weights to form a teacher for the knowledge distillation. The two components complement each other to regularize the MAE at the data and model levels to facilitate better generalization across tasks and reduce the risk of catastrophic forgetting. Experimental results show that CoSMAE achieves significant improvements of up to 4.94% over state-of-the-art CL methods applied to MAE. Our code is publicly available at: https://git.tu-berlin.de/rsim/CoSMAE.