Knowledge Distillation

Follow this topic Ask about this topic

Knowledge distillation studies how a student model can learn from a teacher’s outputs, internal representations, examples, or behavior, often to reduce inference cost or transfer capabilities. Common approaches match softened class probabilities, hidden features, attention or relational structure, and use teacher-generated data; newer language-model methods also train on student-generated rollouts. The field spans model compression, cross-architecture and multimodal transfer, and settings such as federated learning, robotics, and low-resource tasks. A central challenge is choosing supervision that fits the student and remains reliable: stronger or denser teacher signals can help, but can also cause negative transfer, forgetting, or loss of capabilities not captured by the benchmark.

Latest papers

829 papers164 papers in the last four weeks, up 95% on the four weeks before. 0.9% of all new papers.

Apr 29, 2026

Edge AI for Automotive Vulnerable Road User Safety: Deployable Detection via Knowledge Distillation

Deploying accurate object detection for Vulnerable Road User (VRU) safety on edge hardware requires balancing model capacity against computational constraints. Large models achieve high accuracy but fail under INT8 quantization required for edge deployment, while small models sacrifice detection performance. This paper presents a knowledge distillation (KD) framework that trains a compact YOLOv8-S student (11.2M parameters) to mimic a YOLOv8-L teacher (43.7M parameters), achieving 3.9x compression while preserving quantization robustness. We evaluate on full-scale BDD100K (70K training images) with Post-Training Quantization to INT8. The teacher suffers catastrophic degradation under INT8 (-23% mAP), while the KD student retains accuracy (-5.6% mAP). Analysis reveals that KD transfers precision calibration rather than raw detection capacity: the KD student achieves 0.748 precision versus 0.653 for direct training at INT8, a 14.5% gain at equivalent recall, reducing false alarms by 44% versus the collapsed teacher. At INT8, the KD student exceeds the teacher's FP32 precision (0.748 vs. 0.718) in a model 3.9x smaller. These findings establish knowledge distillation as a requirement for deploying accurate, safety-critical VRU detection on edge hardware.

Akshay Karjol, Darrin M. Hanna

Department of Electrical and Computer Engineering, Oakland University, Rochester, MI 48309 USA

Edge Inference Autonomous Driving Perception Knowledge Distillation

Apr 29, 2026

GaitKD: A Universal Decoupled Distillation Framework for Efficient Gait Recognition

Gait recognition is an attractive biometric modality for long-range and contact-free identification, but high-performing gait models often rely on deep and computationally expensive architectures that are difficult to deploy in practice. Knowledge distillation (KD) offers a natural way to transfer knowledge from a powerful teacher to an efficient student; however, standard KD is often less effective for part-structured gait models, where supervision is formed from both part-wise classification logits and part-wise retrieval embeddings. In this paper, we propose GaitKD, a distillation framework that decouples gait knowledge transfer into two complementary components: decision-level distillation and boundary-level distillation. Specifically, GaitKD aligns the teacher and student through part-calibrated logit distillation to transfer inter-class decision relations, while preserving the teacher-induced partitioning of the embedding space through an activation-boundary objective instead of direct feature regression. With a simple aligned part-wise design, GaitKD supports heterogeneous teacher-student gait models without introducing additional inference cost. Experimental results across multiple gait recognition benchmarks and teacher-student configurations show consistent improvements over strong gait baselines. Our study demonstrates that the two transfer components are complementary, and boundary-preserving distillation provides more stable performance than direct feature regression. Source code is available at https://github.com/liyiersan/GaitKD/

Yuqi Li, Qian Zhou, Huiran Duan +5

The City University of New York · Wuhan University · Beijing Jiaotong University +2

Gait Analysis Gait Recognition Teacher-Student Learning Knowledge Distillation

Apr 28, 2026

Privileged Foresight Distillation: Zero-Cost Future Correction for World Action Models

World action models jointly predict future video and action during training, raising an open question about what role the future-prediction branch actually plays. A recent finding shows that this branch can be removed at inference with little to no loss on common manipulation benchmarks, suggesting that future information may act merely as a regularizer on the shared visual backbone. We propose instead that joint training induces an action-conditioned correction that privileged future observations impose on action denoising, and that current-only policies capture this correction only partially. Making the account precise, we formulate privileged foresight as a residual in the action-denoising direction -- the difference between what a model predicts given the true future and what it predicts given only the current frame -- and introduce Privileged Foresight Distillation (PFD), which transfers this residual from a training-time teacher into a small adapter on a current-only student. The teacher and student share the same backbone and differ only in the attention mask over video tokens; future video is never generated at inference. Controlled experiments verify that this gain reflects a genuine future-conditioned correction rather than a side effect of capacity or regularization. Empirically, PFD achieves consistent improvements on LIBERO and RoboTwin manipulation benchmarks while preserving the current-only inference interface at negligible added latency. This view reframes the role of future information in world action models: not as a target to predict, nor as a regularizer to absorb, but as a compressible correction to be distilled.

Pengcheng Fang, Hongli Chen, Xiaohao Cai

The University of Southampton · The University of Queensland

World Model Learning Action-Conditioned World Models Robotic Manipulation Learning Using Privileged Information Knowledge Distillation

Apr 28, 2026

Improving Diversity in Black-box Few-shot Knowledge Distillation

Knowledge distillation (KD) is a well-known technique to effectively compress a large network (teacher) to a smaller network (student) with little sacrifice in performance. However, most KD methods require a large training set and internal access to the teacher, which are rarely available due to various restrictions. These challenges have originated a more practical setting known as black-box few-shot KD, where the student is trained with few images and a black-box teacher. Recent approaches typically generate additional synthetic images but lack an active strategy to promote their diversity, a crucial factor for student learning. To address these problems, we propose a novel training scheme for generative adversarial networks, where we adaptively select high-confidence images under the teacher's supervision and introduce them to the adversarial learning on-the-fly. Our approach helps expand and improve the diversity of the distillation set, significantly boosting student accuracy. Through extensive experiments, we achieve state-of-the-art results among other few-shot KD methods on seven image datasets. The code is available at https://github.com/votrinhan88/divbfkd.

Tri-Nhan Vo, Dang Nguyen, Kien Do +1

Applied Artificial Intelligence Institute (A2I2), Deakin University, Australia

Synthetic Data Augmentation Few-Shot Learning Knowledge Distillation Generative Adversarial Networks

Apr 28, 2026

Diverse Image Priors for Black-box Data-free Knowledge Distillation

Knowledge distillation (KD) represents a vital mechanism to transfer expertise from complex teacher networks to efficient student models. However, in decentralized or secure AI ecosystems, privacy regulations and proprietary interests often restrict access to the teacher's interface and original datasets. These constraints define a challenging black-box data-free KD scenario where only top-1 predictions and no training data are available. While recent approaches utilize synthetic data, they still face limitations in data diversity and distillation signals. We propose Diverse Image Priors Knowledge Distillation (DIP-KD), a framework that addresses these challenges through a three-phase collaborative pipeline: (1) Synthesis of image priors to capture diverse visual patterns and semantics; (2) Contrast to enhance the collective distinction between synthetic samples via contrastive learning; and (3) Distillation via a novel primer student that enables soft-probability KD. Our evaluation across 12 benchmarks shows that DIP-KD achieves state-of-the-art performance, with ablations confirming data diversity as critical for knowledge acquisition in restricted AI environments.

Tri-Nhan Vo, Dang Nguyen, Trung Le +2

Applied Artificial Intelligence Initiative, Deakin University, Geelong, Australia · Department of Data Science & AI, Monash University, Melbourne, Australia

Synthetic Data Augmentation Data-Free Knowledge Distillation Knowledge Distillation

Apr 28, 2026

Sustained Gradient Alignment Mediates Subliminal Learning in a Multi-Step Setting: Evidence from MNIST Auxiliary Logit Distillation Experiment

In the MNIST auxiliary logit distillation experiment, a student can acquire an unintended teacher trait despite distilling only on no-class logits through a phenomenon called subliminal learning. Under a single-step gradient descent assumption, subliminal learning theory attributes this effect to alignment between the trait and distillation gradients, but does not guarantee that this alignment persists in a multi-step setting. We empirically show that gradient alignment remains weakly but consistently positive throughout training and causally contributes to trait acquisition. We show that a mitigation method called liminal training works by attenuating the alignment and fails to stop trait acquisition in this setup. These results suggest that mitigation methods that operate in this regime may not reliably suppress trait acquisition when the first-order drive dominates.

Chayanon Kitkana, Shivam Arora

Independent · Equivariant labs

Subliminal Learning Teacher-Student Learning Knowledge Distillation Gradient Descent Dynamics Gradient Alignment

Apr 28, 2026

The Surprising Effectiveness of Canonical Knowledge Distillation for Semantic Segmentation

Recent knowledge distillation (KD) methods for semantic segmentation introduce increasingly complex hand-crafted objectives, yet are typically evaluated under fixed iteration schedules. These objectives substantially increase per-iteration cost, meaning equal iteration counts do not correspond to equal training budgets. It is therefore unclear whether reported gains reflect stronger distillation signals or simply greater compute. We show that iteration-based comparisons are misleading: when wall-clock compute is matched, canonical logit- and feature-based KD outperform recent segmentation-specific methods. Under extended training, feature-based distillation achieves state-of-the-art ResNet-18 performance on Cityscapes and ADE20K. A PSPNet ResNet-18 student closely approaches its ResNet-101 teacher despite using only one quarter of the parameters, reaching 99% of the teacher's mIoU on Cityscapes (79.0 vs 79.8) and 92% on ADE20K. Our results challenge the prevailing assumption that KD for segmentation requires task-specific mechanisms and suggest that scaling, rather than complex hand-crafted objectives, should guide future method design.

Muhammad Ali, Kevin Alexander Laube, Madan Ravi Ganesh +3

University of Freiburg · Bosch Center for Artificial Intelligence · Aleph Alpha Research +1

Image Segmentation Feature-Based Knowledge Distillation Knowledge Distillation Semantic Segmentation

Apr 28, 2026

Knowledge Distillation Must Account for What It Loses

This position paper argues that knowledge distillation must account for what it loses: student models should be judged not only by retained task scores, but by whether they preserve the teacher capabilities that make those scores reliable. This matters because distillation is increasingly used to turn large teacher models into deployable students, yet headline metrics can obscure losses in the capabilities that make teacher behavior reliable. Conceptually, we show that current evaluation often assumes retained task scores imply retained teacher capabilities. Reframing distillation as a lossy projection exposes this flaw: students may match selected teacher observables without preserving the capabilities that make them reliable. We then synthesize existing evidence into a taxonomy of off-metric distillation losses, showing that such losses are concrete, recurring, and measurable, yet often unaccounted for when studies report what students retain rather than what they lose. To make the position actionable, we propose scenario-specific preservation targets and a Distillation Loss Statement that reports what was preserved, what was lost, and why the remaining losses are acceptable. The goal is not lossless distillation, but accountable distillation.

Wenshuo Wang

School of Future Technology, South China University of Technology, China

Teacher-Student Learning Knowledge Distillation

Apr 27, 2026

S-SONDO: Self-Supervised Knowledge Distillation for General Audio Foundation Models

General audio foundation models have recently achieved remarkable progress, enabling strong performance across diverse tasks. However, state-of-the-art models remain extremely large, often with hundreds of millions of parameters, leading to high inference costs and limited deployability on edge devices. Knowledge distillation is a proven strategy for model compression, but prior work in audio has mostly focused on supervised settings, relying on class logits, intermediate features, or architecture-specific techniques. Such assumptions exclude models that output only embeddings, such as self-supervised or metric-learning models. We introduce S-SONDO (Self-Supervised KnOwledge DistillatioN for General AuDio FOundation Models), the first framework to distill general audio models using only their output embeddings. By avoiding the need for logits or layer-level alignment, S-SONDO is architecture-agnostic and broadly applicable to embedding-based teachers. We demonstrate its effectiveness by distilling two audio foundation models into three efficient students that are up to 61 times smaller while retaining up to 96% of teacher performance. We also provide practical insights on loss choice and clustering-based balanced data sampling. Code is available here: https://github.com/MedAliAdlouni/ssondo.

Mohammed Ali El Adlouni, Aurian Quelennec, Pierre Chouteau +2

LTCI, T´el´ecom Paris, Institut Polytechnique de Paris, Palaiseau, France

Model Compression Audio Representation Learning Self-Supervised Knowledge Distillation Knowledge Distillation

Apr 27, 2026

Selective Correlation Based Knowledge Distillation for Ground Reaction Force Estimation

Wearable sensor-based human gait analysis holds great promise in healthcare, rehabilitation, clinical diagnosis and monitoring, and sports activities. Specifically, ground reaction force (GRF) provides essential insights into the body's interaction with the ground during movement and is typically measured using instrumented treadmills equipped with force plates. However, such equipment is expensive and restricted to laboratory environments. To enable a more portable solution, wearable insole sensors have been used to measure GRF. These sensors, however, are prone to noise and external interference, which reduces measurement accuracy. Deep learning methodologies could be adopted to address these issues, but they often require significant computing resources to achieve high accuracy, limiting their applicability for real-time analysis on portable devices. To overcome these limitations, we propose Selective Correlation Based Knowledge Distillation (SCKD) for estimating GRF from data collected by insole sensors. Our proposed method utilizes selected features considering temporal characteristics in the process of extracting correlation maps for knowledge transfer, enhancing interpretability and mitigating issues in high dimensional data processing. We demonstrate the effectiveness of the compact models generated by our distillation framework through comparison with existing methods. Various configurations of teacher-student architectures and training approaches are examined based on multiple evaluation criteria, utilizing data collected at different walking speeds and with different window sizes. Experimental results confirm that our approach outperforms existing methods in estimating GRF from wearable insole sensor data. Therefore, our approach offers a reliable and resource-efficient solution for human gait analysis.

Eun Som Jeon, Jisoo Lee, Huisu Lim +3

Department of Computer Science and Engineering, Seoul National University of Science and Technology, Seoul, 01811, Republic of Korea · Geometric Media Lab, School of Computing and Augmented Intelligence, Arizona State University, Tempe, 85281, AZ, USA · School for Engineering of Matter, Transport and Energy, Arizona State University, Tempe, 82587, AZ, USA +1

Gait Analysis Wearable Sensing Knowledge Distillation

Apr 25, 2026

Hiding in Plain Sight: Detectability-Aware Antidistillation of Reasoning Models

Distillation via reasoning traces exposes frontier models to adversarial third parties, who can bypass their guardrails and misappropriate their capabilities. Existing antidistillation defenses perturb the teacher's sampling distribution to hinder downstream student learning, but these perturbations also degrade the trace quality even when task accuracy is preserved, often leaving visible artifacts such as syntactic incoherencies and repetition chains. Beyond eroding user trust, such traces are statistically distinguishable, allowing adversaries to filter or resample them and weaken the defense. An effective defense must therefore limit how visibly it degrades the teacher. To address this gap, we formulate antidistillation as a max-min game whose constraint set explicitly encodes defense detectability. Satisfying both axes--- degrading the student while remaining undetectable---may require more clever perturbations, akin to sparse but targeted adversarial attacks in robustness literature. We apply the same analogy to reasoning steps, targeting thought anchors: sentences identified in prior interpretability work as strongly influencing a model's output. We propose Thought Anchor RemovalS (TARS), which identifies and removes thought anchors, whose content is semantically redundant, so that the reasoning trace remains legible to the user. Modeling these removals as burst-deletions, we then provide formal guarantees on its detectability. Finally, we empirically demonstrate the efficacy of TARS in consistently hindering downstream distillation across teacher model scales (DeepSeek R1 Distill Qwen 7B, Gemma 3 12B), reasoning benchmarks (MMLU Pro, MATH, DeepMath), and student architectures (Llama 3.2 1B, Llama 3.2 3B, Gemma 3 1B, Qwen 3 1.7B) while preserving teacher performance.

Max Hartman, Vidhata Jayaraman, Moulik Choraria +1

Stony Brook University

CoT Distillation Teacher-Student Learning Knowledge Distillation Reasoning Trace Analysis

Apr 24, 2026

Distilling Vision Transformers for Distortion-Robust Representation Learning

Self-supervised learning has achieved remarkable success in learning visual representations from clean data, yet remains challenging when clean observations are sparse or not available at all. In this paper, we demonstrate that pretrained vision models can be leveraged to learn distortion-robust representations, which can then be effectively applied to downstream tasks operating on distorted observations. In particular, we propose an asymmetric knowledge distillation framework in which both teacher and student are initialized from the same pretrained Vision Transformer but receive different views of each image: the teacher processes clean images, while the student sees their distorted versions. We introduce multi-level distillation that aligns global embeddings, patch-level features, and attention maps and show that the student is able to approximate clean-image representations despite never directly accessing clean data. We evaluate our approach on image classification tasks across several datasets and under various distortions, consistently outperforming existing alternatives for the same amount of human supervision.

Konstantinos Alexis, Giorgos Giannopoulos, Dimitrios Gunopulos

Department of Informatics and Telecommunications, National and Kapodistrian University of Athens, Greece · National Observatory of Athens, Greece

Representation Learning Visual Representation Learning Vision Transformer Knowledge Distillation

Apr 23, 2026

Pre-trained LLMs Meet Sequential Recommenders: Efficient User-Centric Knowledge Distillation

Sequential recommender systems have achieved significant success in modeling temporal user behavior but remain limited in capturing rich user semantics beyond interaction patterns. Large Language Models (LLMs) present opportunities to enhance user understanding with their reasoning capabilities, yet existing integration approaches create prohibitive inference costs in real time. To address these limitations, we present a novel knowledge distillation method that utilizes textual user profile generated by pre-trained LLMs into sequential recommenders without requiring LLM inference at serving time. The resulting approach maintains the inference efficiency of traditional sequential models while requiring neither architectural modifications nor LLM fine-tuning.

Nikita Severin, Danil Kartushov, Vladislav Urzhumov +8

Independent researcher, Belgrade, Serbia · ITMO University, Saint Petersburg, Russia · Innopolis University, Innopolis, Russia +3

Sequential Recommendation Language Model Distillation Large Language Model-Based Recommendation User Preference Modeling Knowledge Distillation

Apr 23, 2026

VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution

Recent advancements in visual autoregressive models (VAR) have demonstrated their effectiveness in image generation, highlighting their potential for real-world image super-resolution (Real-ISR). However, adapting VAR for ISR presents critical challenges. The next-scale prediction mechanism, constrained by causal attention, fails to fully exploit global low-quality (LQ) context, resulting in blurry and inconsistent high-quality (HQ) outputs. Additionally, error accumulation in the iterative prediction severely degrades coherence in ISR task. To address these issues, we propose VARestorer, a simple yet effective distillation framework that transforms a pre-trained text-to-image VAR model into a one-step ISR model. By leveraging distribution matching, our method eliminates the need for iterative refinement, significantly reducing error propagation and inference time. Furthermore, we introduce pyramid image conditioning with cross-scale attention, which enables bidirectional scale-wise interactions and fully utilizes the input image information while adapting to the autoregressive mechanism. This prevents later LQ tokens from being overlooked in the transformer. By fine-tuning only 1.2% of the model parameters through parameter-efficient adapters, our method maintains the expressive power of the original VAR model while significantly enhancing efficiency. Extensive experiments show that VARestorer achieves state-of-the-art performance with 72.32 MUSIQ and 0.7669 CLIPIQA on DIV2K dataset, while accelerating inference by 10 times compared to conventional VAR inference.

Yixuan Zhu, Shilin Ma, Haolin Wang +6

Tsinghua University

Autoregressive Image Super-Resolution Autoregressive Image Generation Knowledge Distillation Image Super-Resolution

Apr 23, 2026

When Agents Look the Same: Quantifying Distillation-Induced Similarity in Tool-Use Behaviors

Model distillation is a primary driver behind the rapid progress of LLM agents, yet it often leads to behavioral homogenization. Many emerging agents share nearly identical reasoning steps and failure modes, suggesting they may be distilled echoes of a few dominant teachers. Existing metrics, however, fail to distinguish mandatory behaviors required for task success from non-mandatory patterns that reflect a model's autonomous preferences. We propose two complementary metrics to isolate non-mandatory behavioral patterns: Response Pattern Similarity (RPS) for verbal alignment and Action Graph Similarity (AGS) for tool-use habits modeled as directed graphs. Evaluating 18 models from 8 providers on

τ

-Bench and

τ^2

-Bench against Claude Sonnet 4.5 (thinking), we find that within-family model pairs score 5.9 pp higher in AGS than cross-family pairs, and that Kimi-K2 (thinking) reaches 82.6%

S_{\text{node}}

and 94.7%

S_{\text{dep}}

, exceeding Anthropic's own Opus 4.1. A controlled distillation experiment further confirms that AGS distinguishes teacher-specific convergence from general improvement. RPS and AGS capture distinct behavioral dimensions (Pearson

r

= 0.491), providing complementary diagnostic signals for behavioral convergence in the agent ecosystem. Our code is available at https://github.com/Syuchin/AgentEcho.

Chenghao Yang, Yuning Zhang, Zhoufutu Wen +4

School of Cyber Science and Technology, USTC · Anhui Province Key Laboratory of Digital Security · 3M-A-P +1

Tool-Augmented Language Model Agents LLM Agent Evaluation AI Agent Benchmarks Tool-Using Agents Knowledge Distillation Behavioral Consistency of LLM Agents

Apr 22, 2026

Hybrid Policy Distillation for LLMs

Knowledge distillation (KD) is a powerful paradigm for compressing large language models (LLMs), whose effectiveness depends on intertwined choices of divergence direction, optimization strategy, and data regime. We break down the design of existing KD methods and present a unified view that establishes connections between them, reformulating KD as a reweighted log-likelihood objective at the token level. We further propose Hybrid Policy Distillation (HPD), which integrates the complementary advantages of forward and reverse KL to balance mode coverage and mode-seeking, and combines off-policy data with lightweight, approximate on-policy sampling. We validate HPD on long-generation math reasoning as well as short-generation dialogue and code tasks, demonstrating improved optimization stability, computational efficiency, and final performance across diverse model families and scales. The code related to this work is available at https://github.com/zwhong714/Hybrid-Policy-Distillation.

Wenhong Zhu, Ruobing Xie, Rui Wang +1

Department of Computer Science and Engineering, Shanghai Jiao Tong University · Shanghai Innovation Institute · Large Language Department, Tencent

Language Model Distillation On-Policy Distillation Policy Distillation Knowledge Distillation

Apr 22, 2026

Weighted Knowledge Distillation for Semi-Supervised Segmentation of Maxillary Sinus in Panoramic X-ray Images

Accurate segmentation of maxillary sinus in panoramic X-ray images is essential for dental diagnosis and surgical planning; however, this task remains relatively underexplored in dental imaging research. Structural overlap, ambiguous anatomical boundaries inherent to two-dimensional panoramic projections, and the limited availability of large scale clinical datasets with reliable pixel-level annotations make the development and evaluation of segmentation models challenging. To address these challenges, we propose a semi-supervised segmentation framework that effectively leverages both labeled and unlabeled panoramic radiographs, where knowledge distillation is utilized to train a student model with reliable structural information distilled from a teacher model. Specifically, we introduce a weighted knowledge distillation loss to suppress unreliable distillation signals caused by structural discrepancies between teacher and student predictions. To further enhance the quality of pseudo labels generated by the teacher network, we introduce SinusCycle-GAN which is a refinement network based on unpaired image-to-image translation. This refinement process improves the precision of boundaries and reduces noise propagation when learning from unlabeled data during semi-supervised training. To evaluate the proposed method, we collected clinical panoramic X-ray images from 2,511 patients, and experimental results demonstrate that the proposed method outperforms state-of-the-art segmentation models, achieving the Dice score of 96.35% while reducing boundary error. The results indicate that the proposed semi-supervised framework provides robust and anatomically consistent segmentation performance under limited labeled data conditions, highlighting its potential for broader dental image analysis applications.

Juha Park, Jiho Choi, Jong Pil Yun +4

Division of Electronics and Information Engineering, College of Engineering, Jeonbuk National University, 567, Baekje-daero, Deokjin-gu, 54896, Jeonju, Republic of Korea · Manufacturing AI Research Center, Korea Institute of Industrial Technology (KITECH), Incheon 21999, Republic of Korea · KITECH School, University of Science and Technology, Cheonan 31056, Republic of Korea +2

Medical Imaging Dental Imaging Medical Image Segmentation Semi-Supervised Learning Knowledge Distillation

Apr 21, 2026

From Skeletons to Pixels: Few-Shot Precise Event Spotting via Representation and Prediction Distillation

Precise Event Spotting (PES) is essential in fast-paced sports such as tennis, where fine-grained events occur within very short temporal windows. Accurate frame-level localization is challenging because of motion blur, subtle action differences, and limited annotated data. We study two complementary distillation strategies for few-shot PES: Adaptive Weight Distillation (AWD), a prediction-level method that adaptively weights teacher supervision on unlabeled data, and Annealed Multimodal Distillation for Few-Shot Event Detection (AMD-FED), a representation-level framework that transfers robust skeleton knowledge into visual modalities through annealed pseudo-labeling. Both methods use multimodal distillation to improve generalization under limited supervision. We evaluate them on F3Set-Tennis(sub) under few-shot k-clip settings, where they consistently outperform single-modality baselines and prior PES approaches. After observing the stronger performance of representation-level distillation on tennis, we further validate AMD-FED on a second sports dataset, Figure Skating, where it also shows robust performance in the k-clip scenario. These results highlight the effectiveness of multimodal distillation, especially representation-level transfer, for few-shot precise event spotting.

Zhong Han Ervin Yeoh, Jiang Kan

Department of Business Analytics School of Computing National University of Singapore · Department of Computer Science School of Computing National University of Singapore

Cross-Modal Knowledge Distillation Few-Shot Learning Knowledge Distillation Multimodal Learning

Apr 21, 2026

Distillation Traps and Guards: A Calibration Knob for LLM Distillability

Knowledge distillation (KD) transfers capabilities from large language models (LLMs) to smaller students, yet it can fail unpredictably and also underpins model leakage risks. Our analysis revealed several distillation traps: tail noise, off-policy instability, and, most fundamentally, the teacher-student gap, that distort training signals. These traps manifest as overconfident hallucinations, self-correction collapse, and local decoding degradation, causing distillation to fail. Motivated by these findings, we propose a post-hoc calibration method that, to the best of our knowledge, for the first time enables control over a teacher's distillability via reinforcement fine-tuning (RFT). Our objective combines task utility, KL anchor, and across-tokenizer calibration reward. This makes distillability a practical safety lever for foundation models, connecting robust teacher-student transfer with deployment-aware model protection. Experiments across math, knowledge QA, and instruction-following tasks show that students distilled from distillable calibrated teachers outperform SFT and KD baselines, while undistillable calibrated teachers retain their task performance but cause distilled students to collapse, offering a practical knob for both better KD and model IP protection.

Weixiao Zhan, Yongcheng Jing, Leszek Rutkowski +1

Generative AI Lab, College of Computing and Data Science Nanyang Technological University, Singapore 639798 · Systems Research Institute of the Polish Academy of Sciences, AGH University of Krakow, 30-059 Kraków, and the SAN University, 90-113, Łód´z, Poland

Language Model Calibration Language Model Distillation Teacher-Student Learning Knowledge Distillation

Apr 20, 2026

Rethinking Dataset Distillation: Hard Truths about Soft Labels

Despite the perceived success of large-scale dataset distillation (DD) methods, recent evidence finds that simple random image baselines perform on-par with state-of-theart DD methods like SRe2L due to the use of soft labels during downstream model training. This is in contrast with the findings in coreset literature, where high-quality coresets consistently outperform random subsets in the hardlabel (HL) setting. To understand this discrepancy, we perform a detailed scalability analysis to examine the role of data quality under different label regimes, ranging from abundant soft labels (termed as SL+KD regime) to fixed soft labels (SL) and hard labels (HL). Our analysis reveals that high-quality coresets fail to convincingly outperform the random baseline in both SL and SL+KD regimes. In the SL+KD setting, performance further approaches nearoptimal levels relative to the full dataset, regardless of subset size or quality, for a given compute budget. This performance saturation calls into question the widespread practice of using soft labels for model evaluation, where unlike the HL setting, subset quality has negligible influence. A subsequent systematic evaluation of five large-scale and four small-scale DD methods in the HL setting reveals that only RDED reliably outperforms random baselines on ImageNet-1K, but can still lag behind strong coreset methods due to its over-reliance on easy sample patches. Based on this, we introduce CAD-Prune, a compute-aware pruning metric that efficiently identifies samples of optimal difficulty for a given compute budget, and use it to develop CA2D, a compute-aligned DD method, outperforming current DD methods on ImageNet-1K at various IPC settings. Together, our findings uncover many insights into current DD research and establish useful tools to advance dataefficient learning for both coresets and DD.

Priyam Dey, Aditya Sahdev, Sunny Bhati +2

Vision and AI Lab, Indian Institute of Science, Bangalore · University of Maryland, Baltimore County · IIT Hyderabad

Training Data Selection Dataset Distillation Coreset Selection Knowledge Distillation Data-Efficient Learning

Apr 20, 2026

SemLT3D: Semantic-Guided Expert Distillation for Camera-only Long-Tailed 3D Object Detection

Camera-only 3D object detection has emerged as a cost-effective and scalable alternative to LiDAR for autonomous driving, yet existing methods primarily prioritize overall performance while overlooking the severe long-tail imbalance inherent in real-world datasets. In practice, many rare but safety-critical categories such as children, strollers, or emergency vehicles are heavily underrepresented, leading to biased learning and degraded performance. This challenge is further exacerbated by pronounced inter-class ambiguity (e.g., visually similar subclasses) and substantial intra-class diversity (e.g., objects varying widely in appearance, scale, pose, or context), which together hinder reliable long-tail recognition. In this work, we introduce SemLT3D, a Semantic-Guided Expert Distillation framework designed to enrich the representation space for underrepresented classes through semantic priors. SemLT3D consists of: (1) a language-guided mixture-of-experts module that routes 3D queries to specialized experts according to their semantic affinity, enabling the model to better disentangle confusing classes and specialize on tail distributions; and (2) a semantic projection distillation pipeline that aligns 3D queries with CLIP-informed 2D semantics, producing more coherent and discriminative features across diverse visual manifestations. Although motivated by long-tail imbalance, the semantically structured learning in SemLT3D also improves robustness under broader appearance variations and challenging corner cases, offering a principled step toward more reliable camera-only 3D perception.

Hao Vo, Khoa Vo, Thinh Phan +5

AICV Lab, University of Arkansas, USA · University of Utah, USA · University of Houston, USA +1

Autonomous Driving Perception 3D Object Detection Object Detection Knowledge Distillation Monocular 3D Object Detection

Apr 20, 2026

Sharpening Lightweight Models for Generalized Polyp Segmentation: A Boundary Guided Distillation from Foundation Models

Automated polyp segmentation is critical for early colorectal cancer detection and its prevention, yet remains challenging due to weak boundaries, large appearance variations, and limited annotated data. Lightweight segmentation models such as U-Net, U-Net++, and PraNet offer practical efficiency for clinical deployment but struggle to capture the rich semantic and structural cues required for accurate delineation of complex polyp regions. In contrast, large Vision Foundation Models (VFMs), including SAM, OneFormer, Mask2Former, and DINOv2, exhibit strong generalization but transfer poorly to polyp segmentation due to domain mismatch, insufficient boundary sensitivity, and high computational cost. To bridge this gap, we propose LiteBounD, a Lightweight Boundary-guided Distillation framework that transfers complementary semantic and structural priors from multiple VFMs into compact segmentation backbones. LiteBounD introduces (i) a dual-path distillation mechanism that disentangles semantic and boundary-aware representations, (ii) a frequency-aware alignment strategy that supervises low-frequency global semantics and high-frequency boundary details separately, and (iii) a boundary-aware decoder that fuses multi-scale encoder features with distilled semantically rich boundary information for precise segmentation. Extensive experiments on both seen (Kvasir-SEG, CVC-ClinicDB) and unseen (ColonDB, CVC-300, ETIS) datasets demonstrate that LiteBounD consistently outperforms its lightweight baselines by a significant margin and achieves performance competitive with state-of-the-art methods, while maintaining the efficiency required for real-time clinical use. Our code is available at https://github.com/lostinrepo/LiteBounD.

Shivanshu Agnihotri, Snehashis Majhi, Deepak Ranjan Nayak

Dept. of Computer Science and Engineering Malaviya National Institute of Technology Jaipur, India

Image Segmentation Vision Foundation Model Adaptation Polyp Segmentation Medical Image Segmentation Boundary-Aware Segmentation Knowledge Distillation+1

Apr 17, 2026

Improving Reasoning Capabilities in Small Models through Mixture-of-Layers Distillation with Stepwise Attention on Key Information

The significant computational demands of large language models have increased interest in distilling reasoning abilities into smaller models via Chain-of-Thought (CoT) distillation. Current CoT distillation methods mainly focus on transferring teacher-generated rationales for complex reasoning to student models. However, they do not adequately explore teachers' dynamic attention toward critical information during reasoning. We find that language models exhibit progressive attention shifts towards key information during reasoning, which implies essential clues for drawing conclusions. Building on this observation and analysis, we introduce a novel CoT distillation framework that transfers the teacher's stepwise attention on key information to the student model. This establishes structured guidance for the student's progressive concentration on key information during reasoning. More importantly, we develop a Mixture of Layers module enabling dynamic alignment that adapts to different layers between the teacher and student. Our method achieves consistent performance improvements across multiple mathematical and commonsense reasoning datasets. To our knowledge, it is the first method to leverage stepwise attention within CoT distillation to improve small model reasoning.

Yao Chen, Jiawei Sheng, Wenyuan Zhang +1

Institute of Information Engineering, Chinese Academy of Sciences · School of Cyber Security, University of Chinese Academy of Sciences

CoT Distillation Small Language Models Transfer Learning Language Model Distillation Attention Mechanisms Knowledge Distillation

Apr 16, 2026

Weak-to-Strong Knowledge Distillation Accelerates Visual Learning

Large-scale visual learning is increasingly limited by training cost. Existing knowledge distillation methods transfer from a stronger teacher to a weaker student for compression or final-accuracy improvement. We instead investigate distillation to accelerate the training of strong students. We propose a generalizable plug-and-play recipe that freezes a weaker teacher, applies distillation only in early training, and turns it off once the student reaches and surpasses teacher-level performance. For ImageNet and CIFAR classification, this strategy reaches target thresholds much earlier, with up to 4.8 times speedup measured by epochs. We confirm that the method generalizes to other tasks and report 1.7 times epoch speedup for object detection on the COCO dataset, and 2.5 times earlier target-FID crossing for diffusion generation on the CIFAR-10 dataset, measured in steps. These findings validate our method as a universal speedup mechanism for visual learning.

Baiang Li, Wenhao Chai, Felix Heide

Princeton University

Teacher-Student Learning Knowledge Distillation

Apr 16, 2026

DLink: Distilling Layer-wise and Dominant Knowledge from EEG Foundation Models

EEG foundation models (EFMs) achieve strong cross-subject and cross-task generalization through large-scale pretraining and downstream fine-tuning. Through empirical analysis, we observe that (i) task-adapted EFMs provide strong decoding performance but incur substantial overhead when retained as inference backbones, making knowledge distillation a natural route for optimizing compact students; and (ii) direct distillation from a fixed teacher representation underutilizes EFM knowledge, as task-discriminative information is distributed across intermediate layers rather than concentrated in the final layer. These observations motivate DLink (Distilling Layer-wise and Dominant Knowledge), a spectrally guided distillation framework with input-conditioned layer routing for transferring EFM knowledge into compact students. DLink uses a lightweight router to aggregate teacher layers for each input, and aligns magnitude and phase spectra to mitigate compression-induced spectral distortion in learned representations. The routed teacher knowledge is internalized by a project-then-compress student; the teacher and router are used only during training. Experiments on four EEG benchmarks show that DLink improves matched compact students and remains competitive with lightweight baselines, narrowing the gap to fine-tuned EFMs while substantially reducing parameters, FLOPs, and CPU-only inference latency.

Jingyuan Wang, Zhihao Jia, Chenyu Liu +7

1Xiamen University · 2Nanyang Technological University · Institute of Automation, Chinese Academy of Sciences +1

Efficient Neural Network Inference EEG Decoding Knowledge Distillation EEG Foundation Models

Apr 16, 2026

Switch-KD: Visual-Switch Knowledge Distillation for Vision-Language Models

Vision-Language Models (VLMs) have shown remarkable capabilities in joint vision-language understanding, but their large scale poses significant challenges for deployment in resource-constrained scenarios. Knowledge Distillation (KD) offers a viable way to improve model capabilities without increasing model size or data requirements, making deployment more efficient. However, applying KD to VLMs is challenged by modality-specific supervision: although multimodal knowledge in VLMs is fused within the language space, current methods supervise each modality separately without explicitly addressing multimodal alignment, leading to inconsistent multimodal knowledge transfer. To address this, we propose Switch-KD, a visual-switch distillation framework that unifies vision-language knowledge transfer within a shared text-probability space. Switch-KD comprises two key components: (1) Visual-Switch Distillation, which switches the student's visual outputs into the teacher's language pathway to construct cross-modal probabilistic references for implicit visual knowledge transfer; and (2) Dynamic Bi-directional Logits Difference (DBiLD) loss, which adaptively aligns informative probability regions while preserving the distributional structures of teacher and student through bidirectional supervision. Guided by Switch-KD, a 0.5B TinyLLaVA effectively distills rich multimodal knowledge from its 3B teacher, yielding an average improvement of 3.6 points across 10 multimodal benchmarks without any architectural modification.

Haoyi Sun, Xiaoxiao Wang, Ning Mao +5

Li Auto Inc.

Cross-Modal Knowledge Distillation Vision-Language Models VLM Distillation Vision-Language Alignment Knowledge Distillation

Apr 15, 2026

TIP: Token Importance in On-Policy Distillation

On-policy knowledge distillation (OPD) trains a student on its own rollouts under token-level supervision from a teacher. Not all token positions matter equally, but existing views of token importance are incomplete. We ask a direct question: which tokens carry the most useful learning signal in OPD? Our answer is that informative tokens come from two regions: positions with high student entropy, and positions with low student entropy plus high teacher--student divergence, where the student is overconfident and wrong. Empirically, student entropy is a strong first-order proxy: retaining

50\%

of tokens with entropy-based sampling matches or exceeds all-token training while reducing peak memory by up to

47\%

. But entropy alone misses a second important region. When we isolate low-entropy, high-divergence tokens, training on fewer than

10\%

of all tokens nearly matches full-token baselines, showing that overconfident tokens carry dense corrective signal despite being nearly invisible to entropy-only rules. We organize these findings with TIP (Token Importance in on-Policy distillation), a two-axis taxonomy over student entropy and teacher--student divergence, and give a theoretical explanation for why entropy is useful yet structurally incomplete. This view motivates type-aware token selection rules that combine uncertainty and disagreement. We validate this picture across three teacher--student pairs spanning Qwen3, Llama, and Qwen2.5 on MATH-500 and AIME 2024/2025, and on the DeepPlanning benchmark for long-horizon agentic planning, where Q3-only training on

<$$20\%

of tokens surpasses full-token OPD. Our experiments are implemented by extending the OPD repository https://github.com/HJSang/OPSD_OnPolicyDistillation, which supports memory-efficient distillation of larger models under limited GPU budgets.

Yuanda Xu, Hejian Sang, Zhengze Zhou +3

Knowledge Distillation On-Policy Distillation Selective Knowledge Distillation Uncertainty-Aware Knowledge Distillation

Apr 15, 2026

Granularity-Aware Transfer for Tree Instance Segmentation in Synthetic and Real Forests

We address the challenge of synthetic-to-real transfer in forestry perception where real data have only coarse Tree labels while synthetic data provide fine-grained trunk/crown annotations. We introduce MGTD, a mixed-granularity dataset with 53k synthetic and 3.6k real images, and a four-stage protocol isolating domain shift and granularity mismatch. Our core contribution is granularity-aware distillation, which transfers structural priors from fine-grained synthetic teachers to a coarse-label student via logit-space merging and mask unification. Experiments show consistent mask AP gains, especially for small/distant trees, establishing a testbed for Sim-Real transfer under label granularity constraints.

Pankaj Deoli, Atef Tej, Anmol Ashri +2

Robotics Research Lab, University of Kaiserslautern-Landau, Kaiserslautern, Germany · University of Kaiserslautern-Landau, Kaiserslautern, Germany

Instance Segmentation Synthetic-to-Real Domain Adaptation Knowledge Distillation

Apr 14, 2026

Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe

On-policy distillation (OPD) has become a core technique in the post-training of large language models, yet its training dynamics remain poorly understood. This paper provides a systematic investigation of OPD dynamics and mechanisms. We first identify that two conditions govern whether OPD succeeds or fails: (i) the student and teacher should share compatible thinking patterns; and (ii) even with consistent thinking patterns and higher scores, the teacher must offer genuinely new capabilities beyond what the student has seen during training. We validate these findings through weak-to-strong reverse distillation, showing that same-family 1.5B and 7B teachers are distributionally indistinguishable from the student's perspective. Probing into the token-level mechanism, we show that successful OPD is characterized by progressive alignment on high-probability tokens at student-visited states, a small shared token set that concentrates most of the probability mass (97%-99%). We further propose two practical strategies to recover failing OPD: off-policy cold start and teacher-aligned prompt selection. Finally, we show that OPD's apparent free lunch of dense token-level reward comes at a cost, raising the question of whether OPD can scale to long-horizon distillation.

Yaxuan Li, Yuxin Zuo, Bingxiang He +8

Tsinghua University · ShanghaiTech University · University of Illinois Urbana-Champaign +1

Language Model Distillation On-Policy Distillation Knowledge Distillation Teacher-Student Learning

Apr 14, 2026

On the Distillation Loss Functions of Speech VAE for Unified Reconstruction, Understanding, and Generation

Continuous speech representations based on Variational Autoencoders (VAEs) have emerged as a promising alternative to traditional spectrogram or discrete token based features for speech generation and reconstruction. Recent research has tried to enrich the structural information in VAE latent representations by aligning with self-supervised learning (SSL) features, aiming for better generation performance. However, it remains unclear whether the widely-used alignment approach based on time-axis distillation is optimal when considering more tasks. To address this problem, this paper systematically explores different alignment approaches and analyzes their impact on the performances over three axes: reconstruction, understanding, and generation. We investigate various design choices in the distillation loss. Extensive experiments show that the joint-marginal alignment approach with adaptive weighting can achieve the best overall performance while allowing for a controllable balance.

Changhao Cheng, Wei Wang, Wangyou Zhang +4

School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, China · Auditory Cognition and Computational Acoustics Lab, MoE Key Lab of Artificial Intelligence, AI Institute, School of Computer Science, Shanghai Jiao Tong University, Shanghai, China · ByteDance Seed, China

Variational Autoencoders Audio Representation Learning Knowledge Distillation Self-Supervised Speech Representation Learning