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.
This work explores transfer learning from several synthetic languages to English. We investigate the structure of the embeddings in the fine-tuned models, the information they contain, and the capabilities of the fine-tuned models on simple linguistic tasks. We also introduce a new synthetic language that leads to better transfer to English than the languages used in previous research. Finally, we introduce Tiny-Cloze Benchmark - a new synthetic benchmark for natural language understanding that is more informative for less powerful models. We use Tiny-Cloze Benchmark to evaluate fine-tuned models in several domains demonstrating that fine-tuning on a new synthetic language allows for better performance on a variety of tasks.
In electromyogram (EMG)-based motion recognition, a subject-specific classifier is typically trained with sufficient labeled data. However, this process demands extensive data collection over extended periods, burdening the subject. To address this, utilizing information from pre-training on multiple subjects for the training of the target subject could be beneficial. This paper proposes an inter-subject variance transfer learning method based on a Bayesian approach. This method is founded on the simple hypothesis that while the means of EMG features vary greatly across subjects, their variances may exhibit similar patterns. Our approach transfers variance information, acquired through pre-training on multiple source subjects, to a target subject within a Bayesian updating framework, thereby allowing accurate classification using limited target calibration data. A coefficient was also introduced to adjust the amount of information transferred for efficient transfer learning. Experimental evaluations using two EMG datasets demonstrated the effectiveness of our variance transfer strategy and its superiority compared to existing methods.
Transfer learning under domain shift remains a fundamental challenge due to the divergence between source and target data manifolds. In this paper, we propose MAADA (Manifold-Aware Adversarial Data Augmentation), a novel framework that decomposes adversarial perturbations into on-manifold and off-manifold components to simultaneously capture semantic variation and model brittleness. We theoretically demonstrate that enforcing on-manifold consistency reduces hypothesis complexity and improves generalization, while off-manifold regularization smooths decision boundaries in low-density regions. Moreover, we introduce a geometry-aware alignment loss that minimizes geodesic discrepancy between source and target manifolds. Experiments on DomainNet, VisDA, and Office-Home show that MAADA consistently outperforms existing adversarial and adaptation methods in both unsupervised and few-shot settings, demonstrating superior structural robustness and cross-domain generalization.
Multi-modal data in Earth Observation (EO) presents a huge opportunity for improving transfer learning capabilities when pre-training deep learning models. Unlike prior work that often overlooks multi-modal EO data, recent methods have started to include it, resulting in more effective pre-training strategies. However, existing approaches commonly face challenges in effectively transferring learning to downstream tasks where the structure of available data differs from that used during pre-training. This paper addresses this limitation by exploring a more flexible multi-modal, multi-task pre-training strategy for EO data. Specifically, we adopt a Multi-modal Multi-task Masked Autoencoder (MultiMAE) that we pre-train by reconstructing diverse input modalities, including spectral, elevation, and segmentation data. The pre-trained model demonstrates robust transfer learning capabilities, outperforming state-of-the-art methods on various EO datasets for classification and segmentation tasks. Our approach exhibits significant flexibility, handling diverse input configurations without requiring modality-specific pre-trained models. Code will be available at: https://github.com/josesosajs/multimae-meets-eo.
Three-dimensional reconstruction of buildings, particularly at Level of Detail 1 (LOD1), plays a crucial role in various applications such as urban planning, urban environmental studies, and designing optimized transportation networks. This study focuses on assessing the potential of LiDAR data for accurate 3D building reconstruction at LOD1 and extracting morphological features from these models. Four deep semantic segmentation models, U-Net, Attention U-Net, U-Net3+, and DeepLabV3+, were used, applying transfer learning to extract building footprints from LiDAR data. The results showed that U-Net3+ and Attention U-Net outperformed the others, achieving IoU scores of 0.833 and 0.814, respectively. Various statistical measures, including maximum, range, mode, median, and the 90th percentile, were used to estimate building heights, resulting in the generation of 3D models at LOD1. As the main contribution of the research, the impact of segmentation accuracy on the quality of 3D building modeling and the accuracy of morphological features like building area and external wall surface area was investigated. The results showed that the accuracy of building identification (segmentation performance) significantly affects the 3D model quality and the estimation of morphological features, depending on the height calculation method. Overall, the UNet3+ method, utilizing the 90th percentile and median measures, leads to accurate height estimation of buildings and the extraction of morphological features.
Despite progress in geometry-aware domain adaptation, current methods such as GAMA still suffer from two unresolved issues: (1) insufficient disentanglement of task-relevant and task-irrelevant manifold dimensions, and (2) rigid perturbation schemes that ignore per-class alignment asymmetries. To address this, we propose GAMA++, a novel framework that introduces (i) latent space disentanglement to isolate label-consistent manifold directions from nuisance factors, and (ii) an adaptive contrastive perturbation strategy that tailors both on- and off-manifold exploration to class-specific manifold curvature and alignment discrepancy. We further propose a cross-domain contrastive consistency loss that encourages local semantic clusters to align while preserving intra-domain diversity. Our method achieves state-of-the-art results on DomainNet, Office-Home, and VisDA benchmarks under both standard and few-shot settings, with notable improvements in class-level alignment fidelity and boundary robustness. GAMA++ sets a new standard for semantic geometry alignment in transfer learning.
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.
Large Language Models (LLMs) offer a transparent brain with accessible parameters that encode extensive knowledge, which can be analyzed, located and transferred. Consequently, a key research challenge is to transcend traditional knowledge transfer paradigms rooted in symbolic language and achieve genuine Parametric Knowledge Transfer (PKT). Significantly, exploring effective methods for transferring knowledge across LLMs of different scales through parameters presents an intriguing and valuable research direction. In this paper, we first demonstrate $\textbf{Alignment}$ in parametric space is the fundamental prerequisite to achieve successful cross-scale PKT. We redefine the previously explored knowledge transfer as Post-Align PKT (PostPKT), which utilizes extracted parameters for LoRA initialization and requires subsequent fine-tune for alignment. Hence, to reduce cost for further fine-tuning, we introduce a novel Pre-Align PKT (PrePKT) paradigm and propose a solution called $\textbf{LaTen}$ ($\textbf{L}$oc$\textbf{a}$te-$\textbf{T}$h$\textbf{e}$n-Alig$\textbf{n}$) that aligns the parametric spaces of LLMs across scales only using several training steps without following training. Comprehensive experiments on four benchmarks demonstrate that both PostPKT and PrePKT face challenges in achieving consistently stable transfer. Through in-depth analysis, we identify $\textbf{Neural Incompatibility}$ as the ethological and parametric structural differences between LLMs of varying scales, presenting fundamental challenges to achieving effective PKT. These findings provide fresh insights into the parametric architectures of LLMs and highlight promising directions for future research on efficient PKT. Our code is available at https://github.com/Trae1ounG/Neural_Incompatibility.
Parameter-Efficient Fine-Tuning (PEFT) has emerged as a critical paradigm for adapting Large Language Models (LLMs) to downstream tasks, among which Low-rank Adaptation (LoRA) represents one of the most widely adopted methodologies. However, existing LoRA-based approaches exhibit two fundamental limitations: unstable training dynamics and inefficient knowledge transfer from pre-trained models, both stemming from random initialization of adapter parameters. To overcome these challenges, we propose DuDe, a novel approach that decomposes weight matrices into magnitude and direction components, employing Singular Value Decomposition (SVD) for principled initialization. Our comprehensive evaluation demonstrates DuDe's superior performance and robustness, achieving up to 48.35\% accuracy on MMLU and 62.53\% ($\pm$ 1.59) accuracy on GSM8K. Our theoretical analysis and empirical validation collectively demonstrate that DuDe's decomposition strategy enhances optimization stability and better preserves pre-trained representations, particularly for domain-specific tasks requiring specialized knowledge. The combination of robust empirical performance and rigorous theoretical foundations establishes DuDe as a significant contribution to PEFT methodologies for LLMs.
Knowledge tracing (KT) aims to predict learners' future performance based on historical learning interactions. However, existing KT models predominantly focus on data from a single course, limiting their ability to capture a comprehensive understanding of learners' knowledge states. In this paper, we propose TransKT, a contrastive cross-course knowledge tracing method that leverages concept graph guided knowledge transfer to model the relationships between learning behaviors across different courses, thereby enhancing knowledge state estimation. Specifically, TransKT constructs a cross-course concept graph by leveraging zero-shot Large Language Model (LLM) prompts to establish implicit links between related concepts across different courses. This graph serves as the foundation for knowledge transfer, enabling the model to integrate and enhance the semantic features of learners' interactions across courses. Furthermore, TransKT includes an LLM-to-LM pipeline for incorporating summarized semantic features, which significantly improves the performance of Graph Convolutional Networks (GCNs) used for knowledge transfer. Additionally, TransKT employs a contrastive objective that aligns single-course and cross-course knowledge states, thereby refining the model's ability to provide a more robust and accurate representation of learners' overall knowledge states.
While Large Reasoning Models (LRMs) have demonstrated success in complex reasoning tasks through long chain-of-thought (CoT) reasoning, their inference often involves excessively verbose reasoning traces, resulting in substantial inefficiency. To address this, we propose Distilled Reasoning Pruning (DRP), a hybrid framework that combines inference-time pruning with tuning-based distillation, two widely used strategies for efficient reasoning. DRP uses a teacher model to perform skill-aware step decomposition and content pruning, and then distills the pruned reasoning paths into a student model, enabling it to reason both efficiently and accurately. Across several challenging mathematical reasoning datasets, we find that models trained with DRP achieve substantial improvements in token efficiency without sacrificing accuracy. Specifically, DRP reduces average token usage on GSM8K from 917 to 328 while improving accuracy from 91.7% to 94.1%, and achieves a 43% token reduction on AIME with no performance drop. Further analysis shows that aligning the reasoning structure of training CoTs with the student's reasoning capacity is critical for effective knowledge transfer and performance gains.
Vision Language Models (VLMs) have demonstrated strong performance in multi-modal tasks by effectively aligning visual and textual representations. However, most video understanding VLM research has been domain-agnostic, leaving the understanding of their transfer learning capability to specialized domains under-explored. In this work, we address this by exploring the adaptability of open-source VLMs to specific domains, and focusing on soccer as an initial case study. Our approach uses large-scale soccer datasets and LLM to create instruction-following data, and use them to iteratively fine-tune the general-domain VLM in a curriculum learning fashion (first teaching the model key soccer concepts to then question answering tasks). The final adapted model, trained using a curated dataset of 20k video clips, exhibits significant improvement in soccer-specific tasks compared to the base model, with a 37.5% relative improvement for the visual question-answering task and an accuracy improvement from 11.8% to 63.5% for the downstream soccer action classification task.
Sign Language Recognition (SLR) systems primarily focus on manual gestures, but non-manual features such as mouth movements, specifically mouthing, provide valuable linguistic information. This work directly classifies mouthing instances to their corresponding words in the spoken language while exploring the potential of transfer learning from Visual Speech Recognition (VSR) to mouthing recognition in German Sign Language. We leverage three VSR datasets: one in English, one in German with unrelated words and one in German containing the same target words as the mouthing dataset, to investigate the impact of task similarity in this setting. Our results demonstrate that multi-task learning improves both mouthing recognition and VSR accuracy as well as model robustness, suggesting that mouthing recognition should be treated as a distinct but related task to VSR. This research contributes to the field of SLR by proposing knowledge transfer from VSR to SLR datasets with limited mouthing annotations.
Biological brains learn continually from a stream of unlabeled data, while integrating specialized information from sparsely labeled examples without compromising their ability to generalize. Meanwhile, machine learning methods are susceptible to catastrophic forgetting in this natural learning setting, as supervised specialist fine-tuning degrades performance on the original task. We introduce task-modulated contrastive learning (TMCL), which takes inspiration from the biophysical machinery in the neocortex, using predictive coding principles to integrate top-down information continually and without supervision. We follow the idea that these principles build a view-invariant representation space, and that this can be implemented using a contrastive loss. Then, whenever labeled samples of a new class occur, new affine modulations are learned that improve separation of the new class from all others, without affecting feedforward weights. By co-opting the view-invariance learning mechanism, we then train feedforward weights to match the unmodulated representation of a data sample to its modulated counterparts. This introduces modulation invariance into the representation space, and, by also using past modulations, stabilizes it. Our experiments show improvements in both class-incremental and transfer learning over state-of-the-art unsupervised approaches, as well as over comparable supervised approaches, using as few as 1% of available labels. Taken together, our work suggests that top-down modulations play a crucial role in balancing stability and plasticity.
Photonic neural networks (PNNs), which share the inherent benefits of photonic systems, such as high parallelism and low power consumption, could challenge traditional digital neural networks in terms of energy efficiency, latency, and throughput. However, producing scalable photonic artificial intelligence (AI) solutions remains challenging. To make photonic AI models viable, the scalability problem needs to be solved. Large optical AI models implemented on PNNs are only commercially feasible if the advantages of optical computation outweigh the cost of their input-output overhead. In this Perspective, we discuss how field-programmable metasurface technology may become a key hardware ingredient in achieving scalable photonic AI accelerators and how it can compete with current digital electronic technologies. Programmability or reconfigurability is a pivotal component for PNN hardware, enabling in situ training and accommodating non-stationary use cases that require fine-tuning or transfer learning. Co-integration with electronics, 3D stacking, and large-scale manufacturing of metasurfaces would significantly improve PNN scalability and functionalities. Programmable metasurfaces could address some of the current challenges that PNNs face and enable next-generation photonic AI technology.
Considering the importance of detecting hateful language, labeled hate speech data is expensive and time-consuming to collect, particularly for low-resource languages. Prior work has demonstrated the effectiveness of cross-lingual transfer learning and data augmentation in improving performance on tasks with limited labeled data. To develop an efficient and scalable cross-lingual transfer learning approach, we leverage nearest-neighbor retrieval to augment minimal labeled data in the target language, thereby enhancing detection performance. Specifically, we assume access to a small set of labeled training instances in the target language and use these to retrieve the most relevant labeled examples from a large multilingual hate speech detection pool. We evaluate our approach on eight languages and demonstrate that it consistently outperforms models trained solely on the target language data. Furthermore, in most cases, our method surpasses the current state-of-the-art. Notably, our approach is highly data-efficient, retrieving as small as 200 instances in some cases while maintaining superior performance. Moreover, it is scalable, as the retrieval pool can be easily expanded, and the method can be readily adapted to new languages and tasks. We also apply maximum marginal relevance to mitigate redundancy and filter out highly similar retrieved instances, resulting in improvements in some languages.
Code embedding models attract increasing attention due to the widespread popularity of retrieval-augmented generation (RAG) in software development. These models are expected to capture the rich semantic relationships inherent to code, which differ significantly from those found in text. However, existing models remain severely limited due to the scarcity of high-quality training data. In this work, we introduce \textbf{CodeR} (\underline{Code} \underline{R}etrieval), a state-of-the-art embedding model for general-purpose code retrieval. The superior performance of CodeR is built upon CodeR-Pile, a large-scale synthetic dataset constructed under the DRU (Diversity, Reliability, Usability) principle via a novel data synthesis pipeline. To optimize training effectiveness, we propose Annealing, a curriculum learning strategy that enables effective knowledge transfer across heterogeneous sources of data. We evaluate CodeR based on 16 diverse code retrieval tasks, where it significantly outperforms existing baselines and exhibits strong out-of-domain generalization performance. We have publicly released our code and the well-trained model to facilitate further research in this critical area. https://github.com/FlagOpen/FlagEmbedding/tree/master/research/BGE_Coder.
Transfer learning typically leverages representations learned from a source domain to improve performance on a target task. A common approach is to extract features from a pre-trained model and directly apply them for target prediction. However, this strategy is prone to negative transfer where the source representation fails to align with the target distribution. In this article, we propose Residual Feature Integration (REFINE), a simple yet effective method designed to mitigate negative transfer. Our approach combines a fixed source-side representation with a trainable target-side encoder and fits a shallow neural network on the resulting joint representation, which adapts to the target domain while preserving transferable knowledge from the source domain. Theoretically, we prove that REFINE is sufficient to prevent negative transfer under mild conditions, and derive the generalization bound demonstrating its theoretical benefit. Empirically, we show that REFINE consistently enhances performance across diverse application and data modalities including vision, text, and tabular data, and outperforms numerous alternative solutions. Our method is lightweight, architecture-agnostic, and robust, making it a valuable addition to the existing transfer learning toolbox.
Sequential Monte Carlo (SMC) methods offer a principled approach to Bayesian uncertainty quantification but are traditionally limited by the need for full-batch gradient evaluations. We introduce a scalable variant by incorporating Stochastic Gradient Hamiltonian Monte Carlo (SGHMC) proposals into SMC, enabling efficient mini-batch based sampling. Our resulting SMCSGHMC algorithm outperforms standard stochastic gradient descent (SGD) and deep ensembles across image classification, out-of-distribution (OOD) detection, and transfer learning tasks. We further show that SMCSGHMC mitigates overfitting and improves calibration, providing a flexible, scalable pathway for converting pretrained neural networks into well-calibrated Bayesian models.