With a recent trend of using Large Language Models (LLMs) for different applications within smart cities, there is a need for pushing these models toward the edge of network while still preserving their performance. Edge Computing (EC) as a physically closer computing resource to the end users can help to reduce the communication delay for serving end users' tasks for LLM-dependent services. However, EC servers have limited capacity in terms of communication, computation, and storage capacity. This paper introduces DILEMMA, a novel framework addressing the challenges of deploying LLMs in EC systems by jointly optimizing layer placement and layer quantization in EC systems. DILEMMA formulates an Integer Linear Programming problem to minimize total inference delay while ensuring acceptable LLM performance levels, leveraging layer-wise quantization and knowledge distillation for LLM performance control. Experimental evaluations on OPT-350 model using the SQuAD dataset demonstrate that DILEMMA achieves a quantization ratio of up to 12.75% while preserving model loss, highlighting its effectiveness in resource-constrained environments.
Omnimodal Large Language Models (OLLMs) have shown significant progress in integrating vision and text, but still struggle with integrating vision and audio, often exhibiting suboptimal performance when processing audio queries compared to text queries. This disparity is primarily due to insufficient alignment between vision and audio modalities during training, leading to inadequate attention to visual information when using audio queries. To mitigate this issue, we propose a Self-Knowledge Distillation (Self-KD) training method where the vision-text component of the OLLM serves as the teacher and the vision-audio component as the student. This enables the model to process audio in a manner analogous to its text processing. Our experimental results demonstrate that Self-KD is an effective method for enhancing the vision-audio capabilities of OLLMs by learning from the vision-text components, which subsequently improves the interaction between audio and images and results in improved performance on multimodal tasks.
We introduce SEKI, a novel large language model (LLM)-based neural architecture search (NAS) method. Inspired by the chain-of-thought (CoT) paradigm in modern LLMs, SEKI operates in two key stages: self-evolution and knowledge distillation. In the self-evolution stage, LLMs initially lack sufficient reference examples, so we implement an iterative refinement mechanism that enhances architectures based on performance feedback. Over time, this process accumulates a repository of high-performance architectures. In the knowledge distillation stage, LLMs analyze common patterns among these architectures to generate new, optimized designs. Combining these two stages, SEKI greatly leverages the capacity of LLMs on NAS and without requiring any domain-specific data. Experimental results show that SEKI achieves state-of-the-art (SOTA) performance across various datasets and search spaces while requiring only 0.05 GPU-days, outperforming existing methods in both efficiency and accuracy. Furthermore, SEKI demonstrates strong generalization capabilities, achieving SOTA-competitive results across multiple tasks.
Deep online cross-modal hashing has gained much attention from researchers recently, as its promising applications with low storage requirement, fast retrieval efficiency and cross modality adaptive, etc. However, there still exists some technical hurdles that hinder its applications, e.g., 1) how to extract the coexistent semantic relevance of cross-modal data, 2) how to achieve competitive performance when handling the real time data streams, 3) how to transfer the knowledge learned from offline to online training in a lightweight manner. To address these problems, this paper proposes a lightweight contrastive distilled hashing (LCDH) for cross-modal retrieval, by innovatively bridging the offline and online cross-modal hashing by similarity matrix approximation in a knowledge distillation framework. Specifically, in the teacher network, LCDH first extracts the cross-modal features by the contrastive language-image pre-training (CLIP), which are further fed into an attention module for representation enhancement after feature fusion. Then, the output of the attention module is fed into a FC layer to obtain hash codes for aligning the sizes of similarity matrices for online and offline training. In the student network, LCDH extracts the visual and textual features by lightweight models, and then the features are fed into a FC layer to generate binary codes. Finally, by approximating the similarity matrices, the performance of online hashing in the lightweight student network can be enhanced by the supervision of coexistent semantic relevance that is distilled from the teacher network. Experimental results on three widely used datasets demonstrate that LCDH outperforms some state-of-the-art methods.
We introduce the Granite Embedding models, a family of encoder-based embedding models designed for retrieval tasks, spanning dense-retrieval and sparse retrieval architectures, with both English and Multilingual capabilities. This report provides the technical details of training these highly effective 12 layer embedding models, along with their efficient 6 layer distilled counterparts. Extensive evaluations show that the models, developed with techniques like retrieval oriented pretraining, contrastive finetuning, knowledge distillation, and model merging significantly outperform publicly available models of similar sizes on both internal IBM retrieval and search tasks, and have equivalent performance on widely used information retrieval benchmarks, while being trained on high-quality data suitable for enterprise use. We publicly release all our Granite Embedding models under the Apache 2.0 license, allowing both research and commercial use at https://huggingface.co/collections/ibm-granite.
We introduce XCOMPS in this work, a multilingual conceptual minimal pair dataset covering 17 languages. Using this dataset, we evaluate LLMs' multilingual conceptual understanding through metalinguistic prompting, direct probability measurement, and neurolinguistic probing. By comparing base, instruction-tuned, and knowledge-distilled models, we find that: 1) LLMs exhibit weaker conceptual understanding for low-resource languages, and accuracy varies across languages despite being tested on the same concept sets. 2) LLMs excel at distinguishing concept-property pairs that are visibly different but exhibit a marked performance drop when negative pairs share subtle semantic similarities. 3) Instruction tuning improves performance in concept understanding but does not enhance internal competence; knowledge distillation can enhance internal competence in conceptual understanding for low-resource languages with limited gains in explicit task performance. 4) More morphologically complex languages yield lower concept understanding scores and require deeper layers for conceptual reasoning.
The deployment of Large Language Models (LLMs) in customer support is constrained by hallucination-generating false information-and the high cost of proprietary models. To address these challenges, we propose a retrieval-augmented question-answering (QA) pipeline and explore how to balance human input and automation. Using a dataset of questions about a Samsung Smart TV user manual, we demonstrate that synthetic data generated by LLMs outperforms crowdsourced data in reducing hallucination in finetuned models. We also compare self-training (fine-tuning models on their own outputs) and knowledge distillation (fine-tuning on stronger models' outputs, e.g., GPT-4o), and find that self-training achieves comparable hallucination reduction. We conjecture that this surprising finding can be attributed to increased exposure bias issues in the knowledge distillation case and support this conjecture with post hoc analysis. We also improve robustness to unanswerable questions and retrieval failures with contextualized "I don't know" responses. These findings show that scalable, cost-efficient QA systems can be built using synthetic data and self-training with open-source models, reducing reliance on proprietary tools or costly human annotations.
Small language models (SLMs) have become increasingly prominent in the deployment on edge devices due to their high efficiency and low computational cost. While researchers continue to advance the capabilities of SLMs through innovative training strategies and model compression techniques, the security risks of SLMs have received considerably less attention compared to large language models (LLMs).To fill this gap, we provide a comprehensive empirical study to evaluate the security performance of 13 state-of-the-art SLMs under various jailbreak attacks. Our experiments demonstrate that most SLMs are quite susceptible to existing jailbreak attacks, while some of them are even vulnerable to direct harmful prompts.To address the safety concerns, we evaluate several representative defense methods and demonstrate their effectiveness in enhancing the security of SLMs. We further analyze the potential security degradation caused by different SLM techniques including architecture compression, quantization, knowledge distillation, and so on. We expect that our research can highlight the security challenges of SLMs and provide valuable insights to future work in developing more robust and secure SLMs.
While large models pre-trained on high-quality data exhibit excellent performance across various reasoning tasks, including mathematical reasoning (e.g. GSM8k, MultiArith), specializing smaller models to excel at mathematical reasoning remains a challenging problem. Common approaches to address this challenge include knowledge distillation, where smaller student models learn from large pre-trained teacher models, and data augmentation, such as rephrasing questions. Despite these efforts, smaller models struggle with arithmetic computations, leading to errors in mathematical reasoning. In this work, we focus on leveraging a programmatically generated arithmetic dataset to enhance the reasoning capabilities of smaller models. We investigate two key approaches to incorporate this dataset -- (1) intermediate fine-tuning, where a model is fine-tuned on the arithmetic dataset before being trained on a reasoning dataset, and (2) integrating the arithmetic dataset into the instruction-tuning mixture, allowing the model to learn arithmetic skills alongside general instruction-following abilities. Our experiments on multiple reasoning benchmarks demonstrate that incorporating an arithmetic dataset, whether through targeted fine-tuning or within the instruction-tuning mixture, enhances the models' arithmetic capabilities, which in turn improves their mathematical reasoning performance.
Drone-based remote sensing combined with AI-driven methodologies has shown great potential for accurate mapping and monitoring of coral reef ecosystems. This study presents a novel multi-scale approach to coral reef monitoring, integrating fine-scale underwater imagery with medium-scale aerial imagery. Underwater images are captured using an Autonomous Surface Vehicle (ASV), while aerial images are acquired with an aerial drone. A transformer-based deep-learning model is trained on underwater images to detect the presence of 31 classes covering various coral morphotypes, associated fauna, and habitats. These predictions serve as annotations for training a second model applied to aerial images. The transfer of information across scales is achieved through a weighted footprint method that accounts for partial overlaps between underwater image footprints and aerial image tiles. The results show that the multi-scale methodology successfully extends fine-scale classification to larger reef areas, achieving a high degree of accuracy in predicting coral morphotypes and associated habitats. The method showed a strong alignment between underwater-derived annotations and ground truth data, reflected by an AUC (Area Under the Curve) score of 0.9251. This shows that the integration of underwater and aerial imagery, supported by deep-learning models, can facilitate scalable and accurate reef assessments. This study demonstrates the potential of combining multi-scale imaging and AI to facilitate the monitoring and conservation of coral reefs. Our approach leverages the strengths of underwater and aerial imagery, ensuring the precision of fine-scale analysis while extending it to cover a broader reef area.
Language model compression through knowledge distillation has emerged as a promising approach for deploying large language models in resource-constrained environments. However, existing methods often struggle to maintain performance when distilling multilingual models, especially for low-resource languages. In this paper, we present a novel hybrid distillation approach that combines traditional knowledge distillation with a simplified attention matching mechanism, specifically designed for multilingual contexts. Our method introduces an extremely compact student model architecture, significantly smaller than conventional multilingual models. We evaluate our approach on five African languages: Kinyarwanda, Swahili, Hausa, Igbo, and Yoruba. The distilled student model; AfroXLMR-Comet successfully captures both the output distribution and internal attention patterns of a larger teacher model (AfroXLMR-Large) while reducing the model size by over 85%. Experimental results demonstrate that our hybrid approach achieves competitive performance compared to the teacher model, maintaining an accuracy within 85% of the original model's performance while requiring substantially fewer computational resources. Our work provides a practical framework for deploying efficient multilingual models in resource-constrained environments, particularly benefiting applications involving African languages.
Alignment techniques enable Large Language Models (LLMs) to generate outputs that align with human preferences and play a crucial role in their effectiveness. However, their impact often diminishes when applied to Small Language Models (SLMs), likely due to the limited capacity of these models. Instead of directly applying existing alignment techniques to SLMs, we propose to utilize a well-aligned teacher LLM to guide the alignment process for these models, thereby facilitating the transfer of the teacher's knowledge of human preferences to the student model. To achieve this, we first explore a straightforward approach, Dual-Constrained Knowledge Distillation (DCKD), that employs knowledge distillation with two KL-divergence constraints from the aligned teacher to the unaligned student. To further enhance the student's ability to distinguish between preferred and dispreferred responses, we then propose Advantage-Guided Distillation for Preference Alignment (ADPA), which leverages an advantage function from the aligned teacher to deliver more nuanced, distribution-level reward signals for the student's alignment. Our experimental results show that these two approaches appreciably improve the alignment of SLMs and narrow the performance gap with larger counterparts. Among them, ADPA demonstrates superior performance and achieves even greater effectiveness when integrated with DCKD. Our code is available at https://github.com/SLIT-AI/ADPA.
Knowledge distillation is used, in generative language modeling, to train a smaller student model using the help of a larger teacher model, resulting in improved capabilities for the student model. In this paper, we formulate a more general framework for knowledge distillation where the student learns from the teacher during training, and also learns to ask for the teacher's help at test-time following rules specifying test-time restrictions. Towards this, we first formulate knowledge distillation as an entropy-regularized value optimization problem. Adopting Path Consistency Learning to solve this, leads to a new knowledge distillation algorithm using on-policy and off-policy demonstrations. We extend this using constrained reinforcement learning to a framework that incorporates the use of the teacher model as a test-time reference, within constraints. In this situation, akin to a human learner, the model needs to learn not only the learning material, but also the relative difficulty of different sections to prioritize for seeking teacher help. We examine the efficacy of our method through experiments in translation and summarization tasks, observing trends in accuracy and teacher use, noting that our approach unlocks operating points not available to the popular Speculative Decoding approach.
Scaling large language models (LLMs) has shown great potential for improving retrieval model performance; however, previous studies have mainly focused on dense retrieval trained with contrastive loss (CL), neglecting the scaling behavior of other retrieval paradigms and optimization techniques, such as sparse retrieval and knowledge distillation (KD). In this work, we conduct a systematic comparative study on how different retrieval paradigms (sparse vs. dense) and fine-tuning objectives (CL vs. KD vs. their combination) affect retrieval performance across different model scales. Using MSMARCO passages as the training dataset, decoder-only LLMs (Llama-3 series: 1B, 3B, 8B), and a fixed compute budget, we evaluate various training configurations on both in-domain (MSMARCO, TREC DL) and out-of-domain (BEIR) benchmarks. Our key findings reveal that: (1) Scaling behaviors emerge clearly only with CL, where larger models achieve significant performance gains, whereas KD-trained models show minimal improvement, performing similarly across the 1B, 3B, and 8B scales. (2) Sparse retrieval models consistently outperform dense retrieval across both in-domain (MSMARCO, TREC DL) and out-of-domain (BEIR) benchmarks, and they demonstrate greater robustness to imperfect supervised signals. (3) We successfully scale sparse retrieval models with the combination of CL and KD losses at 8B scale, achieving state-of-the-art (SOTA) results in all evaluation sets.
While 3D instance segmentation (3DIS) has advanced significantly, existing methods typically assume that all object classes are known in advance and are uniformly distributed. However, this assumption is unrealistic in dynamic, real-world environments where new classes emerge gradually and exhibit natural imbalance. Although some approaches have addressed class emergence, they often overlook class imbalance, resulting in suboptimal performance -- particularly on rare categories. To tackle this challenge, we propose CLIMB-3D, a unified framework for \textbf{CL}ass-incremental \textbf{Imb}alance-aware \textbf{3D}IS. Building upon established exemplar replay (ER) strategies, we show that ER alone is insufficient to achieve robust performance under constrained memory conditions. To mitigate this, we introduce a novel pseudo-label generator (PLG) that extends supervision to previously learned categories by leveraging predictions from a frozen prior model. Despite its promise, PLG tends to bias towards frequent classes. Therefore, we propose a class-balanced re-weighting (CBR) scheme, that estimates object frequencies from pseudo-labels and dynamically adjusts training bias -- without requiring access to past data. We design and evaluate three incremental scenarios for 3DIS on the challenging ScanNet200 dataset, and additionally on semantic segmentation on ScanNetV2. Our approach achieves state-of-the-art results, surpassing prior work by up to 16.76\% mAP for instance segmentation and approximately 30\% mIoU for semantic segmentation, demonstrating strong generalization across both frequent and rare classes.
Artistic style transfer is concerned with the generation of imagery that combines the content of an image with the style of an artwork. In the realm of computer games, most work has focused on post-processing video frames. Some recent work has integrated style transfer into the game pipeline, but it is limited to single styles. Integrating an arbitrary style transfer method into the game pipeline is challenging due to the memory and speed requirements of games. We present PQDAST, the first solution to address this. We use a perceptual quality-guided knowledge distillation framework and train a compressed model using the FLIP evaluator, which substantially reduces both memory usage and processing time with limited impact on stylisation quality. For better preservation of depth and fine details, we utilise a synthetic dataset with depth and temporal considerations during training. The developed model is injected into the rendering pipeline to further enforce temporal stability and avoid diminishing post-process effects. Quantitative and qualitative experiments demonstrate that our approach achieves superior performance in temporal consistency, with comparable style transfer quality, to state-of-the-art image, video and in-game methods.
This paper presents a novel knowledge distillation neural architecture leveraging efficient transformer networks for effective image classification. Natural images display intricate arrangements encompassing numerous extraneous elements. Vision transformers utilize localized patches to compute attention. However, exclusive dependence on patch segmentation proves inadequate in sufficiently encompassing the comprehensive nature of the image. To address this issue, we have proposed an inner-outer transformer-based architecture, which gives attention to the global and local aspects of the image. Moreover, The training of transformer models poses significant challenges due to their demanding resource, time, and data requirements. To tackle this, we integrate knowledge distillation into the architecture, enabling efficient learning. Leveraging insights from a larger teacher model, our approach enhances learning efficiency and effectiveness. Significantly, the transformer-in-transformer network acquires lightweight characteristics by means of distillation conducted within the feature extraction layer. Our featured network's robustness is established through substantial experimentation on the MNIST, CIFAR10, and CIFAR100 datasets, demonstrating commendable top-1 and top-5 accuracy. The conducted ablative analysis comprehensively validates the effectiveness of the chosen parameters and settings, showcasing their superiority against contemporary methodologies. Remarkably, the proposed Transformer-in-Transformer Network (TITN) model achieves impressive performance milestones across various datasets: securing the highest top-1 accuracy of 74.71% and a top-5 accuracy of 92.28% for the CIFAR100 dataset, attaining an unparalleled top-1 accuracy of 92.03% and top-5 accuracy of 99.80% for the CIFAR-10 dataset, and registering an exceptional top-1 accuracy of 99.56% for the MNIST dataset.
Word order difference between source and target languages is a major obstacle to cross-lingual transfer, especially in the dependency parsing task. Current works are mostly based on order-agnostic models or word reordering to mitigate this problem. However, such methods either do not leverage grammatical information naturally contained in word order or are computationally expensive as the permutation space grows exponentially with the sentence length. Moreover, the reordered source sentence with an unnatural word order may be a form of noising that harms the model learning. To this end, we propose an Implicit Word Reordering framework with Knowledge Distillation (IWR-KD). This framework is inspired by that deep networks are good at learning feature linearization corresponding to meaningful data transformation, e.g. word reordering. To realize this idea, we introduce a knowledge distillation framework composed of a word-reordering teacher model and a dependency parsing student model. We verify our proposed method on Universal Dependency Treebanks across 31 different languages and show it outperforms a series of competitors, together with experimental analysis to illustrate how our method works towards training a robust parser.
Large Language Models (LLMs) achieve state-of-the-art performance across various NLP tasks but face deployment challenges due to high computational costs and memory constraints. Knowledge distillation (KD) is a promising solution, transferring knowledge from large teacher models to smaller student models. However, existing KD methods often assume shared vocabularies and tokenizers, limiting their flexibility. While approaches like Universal Logit Distillation (ULD) and Dual-Space Knowledge Distillation (DSKD) address vocabulary mismatches, they overlook the critical \textbf{reasoning-aware distillation} aspect. To bridge this gap, we propose CoT2Align a universal KD framework that integrates Chain-of-Thought (CoT) augmentation and introduces Cross-CoT Alignment to enhance reasoning transfer. Additionally, we extend Optimal Transport beyond token-wise alignment to a sequence-level and layer-wise alignment approach that adapts to varying sequence lengths while preserving contextual integrity. Comprehensive experiments demonstrate that CoT2Align outperforms existing KD methods across different vocabulary settings, improving reasoning capabilities and robustness in domain-specific tasks.
When designing circuits, engineers obtain the information of electronic devices by browsing a large number of documents, which is low efficiency and heavy workload. The use of artificial intelligence technology to automatically parse documents can greatly improve the efficiency of engineers. However, the current document layout analysis model is aimed at various types of documents and is not suitable for electronic device documents. This paper proposes to use EDocNet to realize the document layout analysis function for document analysis, and use the electronic device document data set created by myself for training. The training method adopts the focus and global knowledge distillation method, and a model suitable for electronic device documents is obtained, which can divide the contents of electronic device documents into 21 categories. It has better average accuracy and average recall rate. It also greatly improves the speed of model checking.