Reasoning models have demonstrated remarkable progress in solving complex and logic-intensive tasks by generating extended Chain-of-Thoughts (CoTs) prior to arriving at a final answer. Yet, the emergence of this "slow-thinking" paradigm, with numerous tokens generated in sequence, inevitably introduces substantial computational overhead. To this end, it highlights an urgent need for effective acceleration. This survey aims to provide a comprehensive overview of recent advances in efficient reasoning. It categorizes existing works into three key directions: (1) shorter - compressing lengthy CoTs into concise yet effective reasoning chains; (2) smaller - developing compact language models with strong reasoning capabilities through techniques such as knowledge distillation, other model compression techniques, and reinforcement learning; and (3) faster - designing efficient decoding strategies to accelerate inference. A curated collection of papers discussed in this survey is available in our GitHub repository.
Estimating the Kullback--Leibler (KL) divergence between language models has many applications, e.g., reinforcement learning from human feedback (RLHF), interpretability, and knowledge distillation. However, computing the exact KL divergence between two arbitrary language models is intractable. Thus, practitioners often resort to the use of sampling-based estimators. While it is easy to fashion a simple Monte Carlo (MC) estimator that provides an unbiased estimate of the KL divergence between language models, this estimator notoriously suffers from high variance, and can even result in a negative estimate of the KL divergence, a non-negative quantity. In this paper, we introduce a Rao--Blackwellized estimator that is also unbiased and provably has variance less than or equal to that of the standard Monte Carlo estimator. In an empirical study on sentiment-controlled fine-tuning, we show that our estimator provides more stable KL estimates and reduces variance substantially in practice. Additionally, we derive an analogous Rao--Blackwellized estimator of the gradient of the KL divergence, which leads to more stable training and produces models that more frequently appear on the Pareto frontier of reward vs. KL compared to the ones trained with the MC estimator of the gradient.
Staining is essential in cell imaging and medical diagnostics but poses significant challenges, including high cost, time consumption, labor intensity, and irreversible tissue alterations. Recent advances in deep learning have enabled digital staining through supervised model training. However, collecting large-scale, perfectly aligned pairs of stained and unstained images remains difficult. In this work, we propose a novel unsupervised deep learning framework for digital cell staining that reduces the need for extensive paired data using knowledge distillation. We explore two training schemes: (1) unpaired and (2) paired-but-misaligned settings. For the unpaired case, we introduce a two-stage pipeline, comprising light enhancement followed by colorization, as a teacher model. Subsequently, we obtain a student staining generator through knowledge distillation with hybrid non-reference losses. To leverage the pixel-wise information between adjacent sections, we further extend to the paired-but-misaligned setting, adding the Learning to Align module to utilize pixel-level information. Experiment results on our dataset demonstrate that our proposed unsupervised deep staining method can generate stained images with more accurate positions and shapes of the cell targets in both settings. Compared with competing methods, our method achieves improved results both qualitatively and quantitatively (e.g., NIQE and PSNR).We applied our digital staining method to the White Blood Cell (WBC) dataset, investigating its potential for medical applications.
Action recognition is an essential task in egocentric vision due to its wide range of applications across many fields. While deep learning methods have been proposed to address this task, most rely on a single modality, typically video. However, including additional modalities may improve the robustness of the approaches to common issues in egocentric videos, such as blurriness and occlusions. Recent efforts in multimodal egocentric action recognition often assume the availability of all modalities, leading to failures or performance drops when any modality is missing. To address this, we introduce an efficient multimodal knowledge distillation approach for egocentric action recognition that is robust to missing modalities (KARMMA) while still benefiting when multiple modalities are available. Our method focuses on resource-efficient development by leveraging pre-trained models as unimodal feature extractors in our teacher model, which distills knowledge into a much smaller and faster student model. Experiments on the Epic-Kitchens and Something-Something datasets demonstrate that our student model effectively handles missing modalities while reducing its accuracy drop in this scenario.
Autonomous Underwater Vehicles (AUVs) play a crucial role in underwater exploration. Vision-based methods offer cost-effective solutions for localization and mapping in the absence of conventional sensors like GPS and LIDAR. However, underwater environments present significant challenges for feature extraction and matching due to image blurring and noise caused by attenuation, scattering, and the interference of \textit{marine snow}. In this paper, we aim to improve the robustness of the feature extraction and matching in the turbid underwater environment using the cross-modal knowledge distillation method that transfers the in-air feature extraction models to underwater settings using synthetic underwater images as the medium. We first propose a novel adaptive GAN-synthesis method to estimate water parameters and underwater noise distribution, to generate environment-specific synthetic underwater images. We then introduce a general knowledge distillation framework compatible with different teacher models. The evaluation of GAN-based synthesis highlights the significance of the new components, i.e. GAN-synthesized noise and forward scattering, in the proposed model. Additionally, the downstream application of feature extraction and matching (VSLAM) on real underwater sequences validates the effectiveness of the transferred model.
We present SpineTrack, the first comprehensive dataset for 2D spine pose estimation in unconstrained settings, addressing a crucial need in sports analytics, healthcare, and realistic animation. Existing pose datasets often simplify the spine to a single rigid segment, overlooking the nuanced articulation required for accurate motion analysis. In contrast, SpineTrack annotates nine detailed spinal keypoints across two complementary subsets: a synthetic set comprising 25k annotations created using Unreal Engine with biomechanical alignment through OpenSim, and a real-world set comprising over 33k annotations curated via an active learning pipeline that iteratively refines automated annotations with human feedback. This integrated approach ensures anatomically consistent labels at scale, even for challenging, in-the-wild images. We further introduce SpinePose, extending state-of-the-art body pose estimators using knowledge distillation and an anatomical regularization strategy to jointly predict body and spine keypoints. Our experiments in both general and sports-specific contexts validate the effectiveness of SpineTrack for precise spine pose estimation, establishing a robust foundation for future research in advanced biomechanical analysis and 3D spine reconstruction in the wild.
Continual generalized category discovery has been introduced and studied in the literature as a method that aims to continuously discover and learn novel categories in incoming data batches while avoiding catastrophic forgetting of previously learned categories. A key component in addressing this challenge is the model's ability to separate novel samples, where Extreme Value Theory (EVT) has been effectively employed. In this work, we propose a novel method that integrates EVT with proxy anchors to define boundaries around proxies using a probability of inclusion function, enabling the rejection of unknown samples. Additionally, we introduce a novel EVT-based loss function to enhance the learned representation, achieving superior performance compared to other deep-metric learning methods in similar settings. Using the derived probability functions, novel samples are effectively separated from previously known categories. However, category discovery within these novel samples can sometimes overestimate the number of new categories. To mitigate this issue, we propose a novel EVT-based approach to reduce the model size and discard redundant proxies. We also incorporate experience replay and knowledge distillation mechanisms during the continual learning stage to prevent catastrophic forgetting. Experimental results demonstrate that our proposed approach outperforms state-of-the-art methods in continual generalized category discovery scenarios.
Gene expression profiling provides critical insights into cellular heterogeneity, biological processes and disease mechanisms. There has been an increasing interest in computational approaches that can predict gene expression directly from digitalized histopathology images. While image foundation models have shown promise in a variety of pathology downstream analysis, their performances on gene-expression prediction are still limited. Explicitly incorporating information from the transcriptomic models can help image models to address domain shift, yet the fine-tuning and alignment of foundation models can be expensive. In the work, we propose Parameter Efficient Knowledge trAnsfer (PEKA), a novel framework that leverages Block-Affine Adaptation and integrates knowledge distillation and structure alignment losses for cross-modal knowledge transfer. We evaluated PEKA for gene expression prediction using multiple spatial transcriptomics datasets (comprising 206,123 image tiles with matched gene expression profiles) that encompassed various types of tissue. PEKA achieved at least 5\% performance improvement over baseline foundation models while also outperforming alternative parameter-efficient fine-tuning strategies. We will release the code, datasets and aligned models after peer-review to facilitate broader adoption and further development for parameter efficient model alignment.
Graphs are growing rapidly, along with the number of distinct label categories associated with them. Applications like e-commerce, healthcare, recommendation systems, and various social media platforms are rapidly moving towards graph representation of data due to their ability to capture both structural and attribute information. One crucial task in graph analysis is node classification, where unlabeled nodes are categorized into predefined classes. In practice, novel classes appear incrementally sometimes with just a few labels (seen classes) or even without any labels (unseen classes), either because they are new or haven't been explored much. Traditional methods assume abundant labeled data for training, which isn't always feasible. We investigate a broader objective: \emph{Graph Class Incremental Learning under Weak Supervision (GCL)}, addressing this challenge by meta-training on base classes with limited labeled instances. During the incremental streams, novel classes can have few-shot or zero-shot representation. Our proposed framework GOTHAM efficiently accommodates these unlabeled nodes by finding the closest prototype representation, serving as class representatives in the attribute space. For Text-Attributed Graphs (TAGs), our framework additionally incorporates semantic information to enhance the representation. By employing teacher-student knowledge distillation to mitigate forgetting, GOTHAM achieves promising results across various tasks. Experiments on datasets such as Cora-ML, Amazon, and OBGN-Arxiv showcase the effectiveness of our approach in handling evolving graph data under limited supervision. The repository is available here: \href{https://github.com/adityashahane10/GOTHAM--Graph-based-Class-Incremental-Learning-Framework-under-Weak-Supervision}{\small \textcolor{blue}{Code}}
Federated learning (FL) offers a privacy-preserving framework for distributed machine learning, enabling collaborative model training across diverse clients without centralizing sensitive data. However, statistical heterogeneity, characterized by non-independent and identically distributed (non-IID) client data, poses significant challenges, leading to model drift and poor generalization. This paper proposes a novel algorithm, pFedKD-WCL (Personalized Federated Knowledge Distillation with Weighted Combination Loss), which integrates knowledge distillation with bi-level optimization to address non-IID challenges. pFedKD-WCL leverages the current global model as a teacher to guide local models, optimizing both global convergence and local personalization efficiently. We evaluate pFedKD-WCL on the MNIST dataset and a synthetic dataset with non-IID partitioning, using multinomial logistic regression and multilayer perceptron models. Experimental results demonstrate that pFedKD-WCL outperforms state-of-the-art algorithms, including FedAvg, FedProx, Per-FedAvg, and pFedMe, in terms of accuracy and convergence speed.
In this paper, we present an effective method to enhance visual reasoning with significantly fewer training samples, relying purely on self-improvement with no knowledge distillation. Our key insight is that the difficulty of training data during reinforcement fine-tuning (RFT) is critical. Appropriately challenging samples can substantially boost reasoning capabilities even when the dataset is small. Despite being intuitive, the main challenge remains in accurately quantifying sample difficulty to enable effective data filtering. To this end, we propose a novel way of repurposing Monte Carlo Tree Search (MCTS) to achieve that. Starting from our curated 70k open-source training samples, we introduce an MCTS-based selection method that quantifies sample difficulty based on the number of iterations required by the VLMs to solve each problem. This explicit step-by-step reasoning in MCTS enforces the model to think longer and better identifies samples that are genuinely challenging. We filter and retain 11k samples to perform RFT on Qwen2.5-VL-7B-Instruct, resulting in our final model, ThinkLite-VL. Evaluation results on eight benchmarks show that ThinkLite-VL improves the average performance of Qwen2.5-VL-7B-Instruct by 7%, using only 11k training samples with no knowledge distillation. This significantly outperforms all existing 7B-level reasoning VLMs, and our fairly comparable baselines that use classic selection methods such as accuracy-based filtering. Notably, on MathVista, ThinkLite-VL-7B achieves the SoTA accuracy of 75.1, surpassing Qwen2.5-VL-72B, GPT-4o, and O1. Our code, data, and model are available at https://github.com/si0wang/ThinkLite-VL.
Current knowledge distillation (KD) methods for semantic segmentation focus on guiding the student to imitate the teacher's knowledge within homogeneous architectures. However, these methods overlook the diverse knowledge contained in architectures with different inductive biases, which is crucial for enabling the student to acquire a more precise and comprehensive understanding of the data during distillation. To this end, we propose for the first time a generic knowledge distillation method for semantic segmentation from a heterogeneous perspective, named HeteroAKD. Due to the substantial disparities between heterogeneous architectures, such as CNN and Transformer, directly transferring cross-architecture knowledge presents significant challenges. To eliminate the influence of architecture-specific information, the intermediate features of both the teacher and student are skillfully projected into an aligned logits space. Furthermore, to utilize diverse knowledge from heterogeneous architectures and deliver customized knowledge required by the student, a teacher-student knowledge mixing mechanism (KMM) and a teacher-student knowledge evaluation mechanism (KEM) are introduced. These mechanisms are performed by assessing the reliability and its discrepancy between heterogeneous teacher-student knowledge. Extensive experiments conducted on three main-stream benchmarks using various teacher-student pairs demonstrate that our HeteroAKD outperforms state-of-the-art KD methods in facilitating distillation between heterogeneous architectures.
We introduce ThermoStereoRT, a real-time thermal stereo matching method designed for all-weather conditions that recovers disparity from two rectified thermal stereo images, envisioning applications such as night-time drone surveillance or under-bed cleaning robots. Leveraging a lightweight yet powerful backbone, ThermoStereoRT constructs a 3D cost volume from thermal images and employs multi-scale attention mechanisms to produce an initial disparity map. To refine this map, we design a novel channel and spatial attention module. Addressing the challenge of sparse ground truth data in thermal imagery, we utilize knowledge distillation to boost performance without increasing computational demands. Comprehensive evaluations on multiple datasets demonstrate that ThermoStereoRT delivers both real-time capacity and robust accuracy, making it a promising solution for real-world deployment in various challenging environments. Our code will be released on https://github.com/SJTU-ViSYS-team/ThermoStereoRT
Accurate and efficient positioning in complex environments is critical for applications where traditional satellite-based systems face limitations, such as indoors or urban canyons. This paper introduces WK-Pnet, an FM-based indoor positioning framework that combines wavelet packet decomposition (WPD) and knowledge distillation. WK-Pnet leverages WPD to extract rich time-frequency features from FM signals, which are then processed by a deep learning model for precise position estimation. To address computational demands, we employ knowledge distillation, transferring insights from a high-capacity model to a streamlined student model, achieving substantial reductions in complexity without sacrificing accuracy. Experimental results across diverse environments validate WK-Pnet's superior positioning accuracy and lower computational requirements, making it a viable solution for positioning in real-time resource-constraint applications.
Beamforming is a key technology in millimeter-wave (mmWave) communications that improves signal transmission by optimizing directionality and intensity. However, conventional channel estimation methods, such as pilot signals or beam sweeping, often fail to adapt to rapidly changing communication environments. To address this limitation, multimodal sensing-aided beam prediction has gained significant attention, using various sensing data from devices such as LiDAR, radar, GPS, and RGB images to predict user locations or network conditions. Despite its promising potential, the adoption of multimodal sensing-aided beam prediction is hindered by high computational complexity, high costs, and limited datasets. Thus, in this paper, a resource-efficient learning approach is proposed to transfer knowledge from a multimodal network to a monomodal (radar-only) network based on cross-modal relational knowledge distillation (CRKD), while reducing computational overhead and preserving predictive accuracy. To enable multimodal learning with realistic data, a novel multimodal simulation framework is developed while integrating sensor data generated from the autonomous driving simulator CARLA with MATLAB-based mmWave channel modeling, and reflecting real-world conditions. The proposed CRKD achieves its objective by distilling relational information across different feature spaces, which enhances beam prediction performance without relying on expensive sensor data. Simulation results demonstrate that CRKD efficiently distills multimodal knowledge, allowing a radar-only model to achieve $94.62\%$ of the teacher performance. In particular, this is achieved with just $10\%$ of the teacher network's parameters, thereby significantly reducing computational complexity and dependence on multimodal sensor data.
As an emerging paradigm of federated learning, asynchronous federated learning offers significant speed advantages over traditional synchronous federated learning. Unlike synchronous federated learning, which requires waiting for all clients to complete updates before aggregation, asynchronous federated learning aggregates the models that have arrived in realtime, greatly improving training speed. However, this mechanism also introduces the issue of client model version inconsistency. When the differences between models of different versions during aggregation become too large, it may lead to conflicts, thereby reducing the models accuracy. To address this issue, this paper proposes an asynchronous federated learning version correction algorithm based on knowledge distillation, named FedADT. FedADT applies knowledge distillation before aggregating gradients, using the latest global model to correct outdated information, thus effectively reducing the negative impact of outdated gradients on the training process. Additionally, FedADT introduces an adaptive weighting function that adjusts the knowledge distillation weight according to different stages of training, helps mitigate the misleading effects caused by the poorer performance of the global model in the early stages of training. This method significantly improves the overall performance of asynchronous federated learning without adding excessive computational overhead. We conducted experimental comparisons with several classical algorithms, and the results demonstrate that FedADT achieves significant improvements over other asynchronous methods and outperforms all methods in terms of convergence speed.
We present a novel approach for training small language models for reasoning-intensive document ranking that combines knowledge distillation with reinforcement learning optimization. While existing methods often rely on expensive human annotations or large black-box language models, our methodology leverages web data and a teacher LLM to automatically generate high-quality training examples with relevance explanations. By framing document ranking as a reinforcement learning problem and incentivizing explicit reasoning capabilities, we train a compact 3B parameter language model that achieves state-of-the-art performance on the BRIGHT benchmark. Our model ranks third on the leaderboard while using substantially fewer parameters than other approaches, outperforming models that are over 20 times larger. Through extensive experiments, we demonstrate that generating explanations during inference, rather than directly predicting relevance scores, enables more effective reasoning with smaller language models. The self-supervised nature of our method offers a scalable and interpretable solution for modern information retrieval systems.
Knowledge distillation has emerged as an effective strategy for compressing large language models' (LLMs) knowledge into smaller, more efficient student models. However, standard one-shot distillation methods often produce suboptimal results due to a mismatch between teacher-generated rationales and the student's specific learning requirements. In this paper, we introduce the UNDO: UNderstanding Distillation as Optimization framework, designed to bridge this gap by iteratively identifying the student's errors and prompting the teacher to refine its explanations accordingly. Each iteration directly targets the student's learning deficiencies, motivating the teacher to provide tailored and enhanced rationales that specifically address these weaknesses. Empirical evaluations on various challenging mathematical and commonsense reasoning tasks demonstrate that our iterative distillation method, UNDO, significantly outperforms standard one-step distillation methods, achieving performance gains of up to 20%. Additionally, we show that teacher-generated data refined through our iterative process remains effective even when applied to different student models, underscoring the broad applicability of our approach. Our work fundamentally reframes knowledge distillation as an iterative teacher-student interaction, effectively leveraging dynamic refinement by the teacher for better knowledge distillation.
Attribute skew in federated learning leads local models to focus on learning non-causal associations, guiding them towards inconsistent optimization directions, which inevitably results in performance degradation and unstable convergence. Existing methods typically leverage data augmentation to enhance sample diversity or employ knowledge distillation to learn invariant representations. However, the instability in the quality of generated data and the lack of domain information limit their performance on unseen samples. To address these issues, this paper presents a global intervention and distillation method, termed FedGID, which utilizes diverse attribute features for backdoor adjustment to break the spurious association between background and label. It includes two main modules, where the global intervention module adaptively decouples objects and backgrounds in images, injects background information into random samples to intervene in the sample distribution, which links backgrounds to all categories to prevent the model from treating background-label associations as causal. The global distillation module leverages a unified knowledge base to guide the representation learning of client models, preventing local models from overfitting to client-specific attributes. Experimental results on three datasets demonstrate that FedGID enhances the model's ability to focus on the main subjects in unseen data and outperforms existing methods in collaborative modeling.
Accurate traffic flow prediction is vital for optimizing urban mobility, yet it remains difficult in many cities due to complex spatio-temporal dependencies and limited high-quality data. While deep graph-based models demonstrate strong predictive power, their performance often comes at the cost of high computational overhead and substantial training data requirements, making them impractical for deployment in resource-constrained or data-scarce environments. We propose the FlowDistill, a lightweight and scalable traffic prediction framework based on knowledge distillation from large language models (LLMs). In this teacher-student setup, a fine-tuned LLM guides a compact multi-layer perceptron (MLP) student model using a novel combination of the information bottleneck principle and teacher-bounded regression loss, ensuring the distilled model retains only essential and transferable knowledge. Spatial and temporal correlations are explicitly encoded to enhance the model's generalization across diverse urban settings. Despite its simplicity, FlowDistill consistently outperforms state-of-the-art models in prediction accuracy while requiring significantly less training data, and achieving lower memory usage and inference latency, highlighting its efficiency and suitability for real-world, scalable deployment.