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.

Jun 4, 2026

Knowledge Distillation for Visual Autoregressive Models

Autoregressive (AR) image generation models are highly expressive but computationally intensive, motivating effective model compression. Knowledge distillation (KD) is a natural approach for model compression and has been widely studied in language modeling, yet its behavior in visual AR generation remains underexplored. In this work, we present the first systematic study of distillation strategies for AR image models. Our analysis shows that while standard distillation can yield meaningful gains, recent methods developed for language do not directly transfer to images: long decoding horizons and visual token ambiguity make teacher supervision unreliable especially under student-conditioned contexts. To address this, we propose VarKD, a distillation framework for visual autoregressive models that distills on student samples while selectively applying teacher supervision and reducing token-level ambiguity. Experiments on ImageNet across multiple AR backbones show that VarKD consistently outperforms prior distillation baselines, narrowing the gap to large-scale models.

Elia Peruzzo, Aritra Bhowmik, Guillaume Sautiere +2

Qualcomm AI Research∗ · Qualcomm AI Research · University of Technology Nuremberg

Visual Autoregressive Models Autoregressive Image Generation Knowledge Distillation

Jun 4, 2026

OPRD: On-Policy Representation Distillation

On-policy distillation (OPD) supervises the student exclusively in the output space by matching next-token distributions. This paradigm suffers from two limitations: (i) a high-variance gradient estimator whose signal-to-noise ratio collapses as the student approaches the teacher, and (ii) an LM-head information bottleneck that discards the teacher's intermediate hidden states. We propose On-Policy Representation Distillation (OPRD), the first method to lift on-policy distillation into the hidden-state space. OPRD aligns student and teacher representations across selected layers on the same on-policy rollouts, providing dense, deterministic, per-layer supervision while bypassing the LM head entirely. Theoretically, OPRD provides a deterministic per-sample gradient, removing the token-level estimation variance that plagues OPD, and exposes structural information that any output-space objective necessarily discards. Empirically, OPRD closes the student-teacher gap on competition mathematics benchmarks (AIME 2024, AIME 2025, and AIMO), where every output-space baseline plateaus below the teacher, while training 1.44x faster and using up to 54% less memory. We further extend OPRD to the cross-architecture setting via OPRD-Bridge. By exploiting the observation that heterogeneous models share a low-rank representational structure, we construct a frozen projector pair that aligns representations across arbitrary depth and width mismatches, shifting the alignment from the output space (which depends on a shared vocabulary) to the representation space. We validate OPRD-Bridge on both cross-architecture (Qwen3-4B -> Qwen3-1.7B-Base) and cross-tokenizer (Phi-4-mini-reasoning -> Qwen3-1.7B-Base) settings, demonstrating successful knowledge transfer even when the vocabulary-based alignment channel is unavailable. Code: https://github.com/ShenzhiYang2000/OPRD.

Shenzhi Yang, Guangcheng Zhu, Bowen Song +8

Zhejiang University · Ant Group

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

Jun 4, 2026

Compress-Distill: Reasoning Trace Compression for Efficient Knowledge Distillation

Reasoning models produce long chain-of-thought traces that are costly to distill and encourage verbose student outputs. We study post-hoc compression of such traces before knowledge distillation. Two teachers, Qwen3.5-397B-A17B and gpt-oss-120B, generate about 283k correct traces each; two instruction-tuned models then compress them to 8.6-21.0% of their original character length. Across a 48-run main grid plus seven Qwen-teacher truncation ablations, compressed traces reduce training tokens to 12-30% of raw, speed up training by 2.0-7.6x, and shorten inference outputs by 3-19x with smaller reductions under the shorter gpt-oss teacher. However, raw traces retain the highest downstream accuracy at every scale and for both teachers. A length-matched raw-trace truncation ablation shows that compression is not merely benefiting from a smaller token budget: model-compressed traces usually beat or match naive truncation, especially for smaller students, while maintaining shorter inference outputs. Overall, reasoning-trace compression offers an accuracy-efficiency trade-off rather than a free improvement: students retain up to 96% of raw-trace accuracy while gaining up to 18x higher per-token efficiency, and at the 0.8B scale under LoRA compressed traces narrow the raw-vs-compressed gap but do not exceed raw.

Maxime Griot, Paul Steven Scotti, Tanishq Mathew Abraham

Université catholique de Louvain · Sophont Inc.

CoT Distillation Teacher-Student Learning Knowledge Distillation CoT Compression Large Reasoning Models

Jun 4, 2026

LiAuto-GeoX: Efficient Grounded Driving Transformer

Dense 3D reconstruction has demonstrated immense potential for spatial understanding, yet its viability as a real-time, onboard representation for autonomous driving remains an open challenge. Existing large-scale visual geometry models typically require substantial computational resources and lack the long-range geometric fidelity, surround-view consistency, and real-time efficiency demanded by dynamic driving environments. To bridge this gap, we present LiAuto-GeoX, an efficient grounded driving transformer designed for deployable, ego-centric 3D scene understanding. Our approach begins by learning a high-capacity driving geometry model from large-scale surround-view data, utilizing sparse LiDAR priors to provide robust geometric grounding in distant, ambiguous, or structure-sparse regions. We then instantiate this capability into a highly compact 155M-parameter onboard model through a novel geometry-preserving distillation framework. This framework employs mask-guided depth-aware distillation to retain fine-grained metric structures by emphasizing geometrically informative regions, and relative-pose relational distillation to enforce cross-view spatial consistency through pose-induced geometric relations. Extensive evaluations reveal that LiAuto-GeoX runs at 220 FPS on KITTI while maintaining high-fidelity dense reconstruction, enabling real-time deployment. The learned geometry transfers seamlessly to downstream autonomy tasks, achieving 90.6 PDMS in trajectory prediction, 24.63 mIoU in occupancy prediction, and 47.67 IoU in future-frame prediction. These all demonstrate that efficient dense 3D reconstruction can transcend its traditional role as a perception target to serve as a scalable, foundational geometric representation for next-generation autonomous driving.

Jiawei Lian, Haoyi Sun, Yang Wu +8

Nanjing University of Science and Technology · Li Auto Inc. · Northwestern Polytechnical University +1

3D ViTs 3D Reconstruction Autonomous Driving Perception Knowledge Distillation

Jun 4, 2026

ViCuR: Visual Cues as Recoverable Privilege for Multimodal On-Policy Distillation

On-policy distillation (OPD) improves reasoning by training a student on trajectories sampled from its own policy under supervision from a teacher. In multimodal reasoning, a common extension is to use a privileged teacher that observes training-time-only signals such as reference answers or rationales. However, such answer-side privilege creates a train-test mismatch: the teacher's supervision may depend on signals unavailable to the student, encouraging shortcut imitation rather than visually grounded reasoning. We propose ViCuR, a visually grounded privileged-teacher distillation framework that replaces answer-side privilege with visual cues (query-related evidence in the input). Because these cues are derived from the same visual input available at inference, their evidence is recoverable by the student. To support this, ViCuR introduces a lightweight cue recovery module that uses dedicated sink-token cross-attention during prefill to aggregate task-relevant visual evidence into an internal representation, without changing the inference interface or requiring auxiliary cue-generation losses. Across seven benchmarks with Qwen3-VL-2B and 8B students, ViCuR consistently improves over answer-based on-policy self-distillation by +1.19 and +1.24 on overall average performance. It also extends naturally to stronger-teacher OPD, surpassing OPD baselines by +0.64 and +1.08, with consistent out-of-domain gains at the 8B scale. These results show that, in multimodal on-policy distillation, the design of teacher privilege is as important as teacher strength.

Kanghui Tian, Siyuan Liu, Ziang Yan +3

1Shanghai AI Laboratory · 3Nanjing University · 2Fudan University

VLM Distillation Multimodal Reasoning On-Policy Distillation Learning Using Privileged Information Teacher-Student Learning Knowledge Distillation

Jun 4, 2026

Beyond Output Matching: Preserving Internal Geometry in NVFP4 LLM Distillation

Demand for low-precision inference, including NVFP4-based approaches, has grown as large language models are increasingly deployed in latency and cost constrained production environments. Quantization-aware distillation (QAD) helps recover accuracy lost under low bit quantization by training a quantized student to match the output distribution of a frozen higher precision teacher via a KL-divergence loss. In this work, we first provide a representation level diagnosis of QAD: output matching alone can mask internal degradation, because many intermediate activation geometries can yield similar teacher-aligned logits. Using CKA, we show that KL-only QAD can reduce layerwise representational similarity relative to the BF16 teacher, with especially severe drift in RL-post-trained models. This drift correlates with downstream bottlenecks on reasoning and coding tasks, suggesting that low bit recovery requires preserving internal geometry rather than matching outputs alone. Motivated by this finding, we propose CKA-QAD, a CKA-guided representational alignment method for NVFP4 QAD and low bit LLM accuracy recovery. The method adds a lightweight regularizer that preserves internal representational geometry during distillation by aligning layerwise Gram matrices through CKA. Across Nemotron 3 Nano and Qwen3-4B-Thinking-2507, CKA-QAD substantially improves representational alignment and improves downstream reasoning and coding accuracy with modest training overhead. Our findings position CKA-guided representational alignment as a practical complement to output matching for quantized LLM recovery.

Fangbo Tu, Junhua Zhao, Chi Liu +4

PayPal, Inc.

LLM Quantization Quantization-Aware Distillation Knowledge Distillation Representation Alignment

Jun 3, 2026

LoRi: Low-Rank Distillation for Implicit Reasoning

Implicit chain-of-thought (iCoT) methods aim to internalize reasoning in large language models, but often underperform explicit CoT prompting. We empirically find that hidden-state reasoning trajectories exhibit low-rank structure. Motivated by this observation, we propose a low-rank distillation framework that transfers reasoning by aligning teacher and student trajectories in a shared low-rank tensor subspace using first- and second-order statistics. The resulting formulation captures the global structure of reasoning while supporting a compact latent reasoning process. We evaluate the method across multiple model families, including LLaMA and Qwen, at different scales on mathematical reasoning benchmarks. Our approach consistently improves performance, especially on challenging multi-step tasks, approaching explicit CoT accuracy and outperforming prior iCoT distillation methods.

Ryan Solgi, Jiayi Tian, Zheng Zhang

University of California-Santa Barbara, USA

CoT Distillation Mathematical Reasoning Implicit Reasoning in Language Models Language Model Distillation Knowledge Distillation Latent CoT Reasoning

Jun 3, 2026

Invariant Gradient Alignment for Robust Reasoning Distillation

Large language models (LLMs) suffer from shortcut learning: they systematically fail on out-of-distribution (OOD) inputs whose semantic surface differs from training data, even when the logical structure is identical. This undermines knowledge distillation pipelines that transfer chain-of-thought reasoning to smaller students. We introduce Invariant Gradient Alignment (IGA), a training framework that aligns gradient updates across semantically diverse but logically isomorphic examples via three innovations: (i) Logical Isomer Sets, groups of problems sharing identical logical structure across distinct semantic domains (mathematics, medicine, law, science); (ii) a differentiable Continuous Gradient Conflict Mask, that suppresses parameter dimensions with high cross-domain gradient variance while preserving invariant directions; and (iii) a truncated SVD projection of the masked gradient back onto the LoRA low-rank manifold, maintaining parameter efficiency throughout. Theoretically, IGA yields tighter OOD generalization bounds than ERM, scaling with the number of isomer domains, and converges at the standard SGD rate under mild regularity. Empirically, IGA outperforms eight baselines across four benchmarks with accuracy gains up to 14.3 pp over ERM-SFT and a Logical Consistency Score of 0.031 versus 0.142 -- a fourfold improvement in representational invariance.

Zehua Cheng, Wei Dai, Jiahao Sun

University of Oxford, Oxford, UK

CoT Distillation Domain Generalization OOD Generalization Knowledge Distillation

Jun 3, 2026

Geometry-Aware Distillation for Prompt Tuning Biomedical Vision-Language Models

Current prompt-based and adapter-based tuning of vision-language models (VLMs) is attractive for medical imaging, where clinical data sensitivity favors frozen backbones and annotations are limited. However, these methods typically optimize only the ground-truth class, treating all other classes as equally incorrect, ignoring clinically meaningful class relations and yielding unstable decision boundaries in limited-supervision settings. We propose Omni-Geometry Knowledge Distillation (OGKD), a new framework that injects class-relation structure into the teacher to produce directional targets that preserve the ground truth while respecting inter-class geometry. Using these targets, we develop two distillation losses: Global Geometry-Aware Distillation (GAD) operates on the global image token, and Label-Guided Geometry Distillation (LGD) applies the same geometry to attentive patch tokens to improve fine-grained alignment. Across comprehensive experiments and analyses on 11 widely-used medical datasets for base-to-novel and few-shot evaluations, our OGKD achieves substantially better performance, consistently improving accuracy by an average absolute gain of 1.7%-2.8% over all prior state-of-the-art VLM adaptation counterparts. It also robustly generalizes to unseen classes and yields more reliable predictions than other approaches. Our code is available at https://github.com/tientrandinh/OGKD.

Tran Dinh Tien, Zhiqiang Shen

Department of Machine Learning Mohamed bin Zayed University of Artificial Intelligence

VLM Adaptation VLM Distillation Medical VLMs Relational Knowledge Distillation Knowledge Distillation Visual Prompt Learning

Jun 3, 2026

DuDi: Dual-Signal Distillation with Cross-Lingual Verbalizer

Small language models (SLMs) are efficient and scalable, but their multilingual capabilities degrade severely at sub-billion scales, especially for Southeast Asian (SEA) languages. We introduce DuDi, a dual-signal multilingual distillation framework that combines an online sequence-level signal with off-policy and on-policy token-level signals. DuDi further uses a cross-lingual verbalizer to refine teacher feedback and improve teacher-student transferability in multilingual settings. Experiments on SEA-HELM across multiple model families, scales, and teacher-student settings show that DuDi consistently outperforms competitive distillation baselines. Ablations and analyses confirm that sequence-level optimization, token-level supervision, and cross-lingual verbalization provide complementary and transferable learning signals for multilingual SLMs.

Patomporn Payoungkhamdee, Tinnakit Udsa, Jian Gang Ngui +3

School of Information Science and Technology, VISTEC · AI Singapore · MBZUAI

Multilingual Language Models Small Language Models Language Model Distillation Low-Resource Language Processing Knowledge Distillation Cross-Lingual Transfer

Jun 3, 2026

Hyper-ICL: Attention Calibration with Hyperbolic Anchor Distillation for Multimodal In-Context Learning

Multimodal In-Context Learning (ICL) has emerged as a practical inference paradigm for Multimodal Large Language Models, where a small set of interleaved image-text In-Context Demonstrations (ICDs) conditions the model to solve new tasks. Despite its flexibility, multimodal ICL incurs high inference latency and suffers from instability due to sensitivity to demonstration formatting, ordering, and content. To address these limitations, we propose Hyper-ICL, a lightweight, training-based framework for demonstration-free multimodal ICL that reconstructs demonstration effects directly without requiring ICDs at inference time. Hyper-ICL learns a parameter-efficient low-rank logit-level adapter that calibrates attention distributions to better match demonstration-induced attention redistribution. To capture how demonstration influence varies across queries, we introduce a query-adaptive modulation mechanism that adaptively controls intervention strength at token level across layers and heads based on the current query. Finally, we propose a layer-wise hyperbolic anchor distillation loss that aligns intermediate student features to a demonstration-conditioned teacher via Lorentz geodesic distance. This loss encourages the student to reconstruct the demonstration-query relationships induced by ICDs. Extensive experiments across six different multimodal benchmarks (including VQAv2, OK-VQA, and COCO Caption) demonstrate that Hyper-ICL consistently improves accuracy and stability over vanilla ICL and existing state-of-the-art methods.

Niloufar Alipour Talemi, Hossein Kashiani, Fatemeh Afghah

Holcombe Department of Electrical and Computer Engineering, Clemson University, Clemson, United States.

Efficient Multimodal Inference Multimodal ICL Multimodal Large Language Models In-Context Learning Attention Mechanisms Multimodal Foundation Models+2

Jun 2, 2026

Derivative Informed Learning of Exchange-Correlation Functionals

Machine-learned (ML) exchange-correlation (XC) functionals aim to replace human-designed density functional approximations by learning directly from reference data, but they still do not consistently outperform traditional

\mathcal{O}(N^4)

-scaling hybrid functionals. We study a hybrid-distillation setting in which

\mathcal{O}(N^3)

-scaling ML-XC functionals are trained to reproduce B3LYP/def2-SVP targets. We introduce Derivative Informed XC-Loss (DI-Loss), a loss that incorporates additional information from the reference hybrid functional by supervising first and second derivatives of the energy on the Grassmannian of admissible density matrices. Rather than only matching the self-consistent fixed point, DI-Loss aligns the local first- and second-order response of the learned functional with that of the target functional. Across four evaluated architectures, DI-Loss consistently improves the main energy metrics. Averaged uniformly across architectures, the total-energy MAE decreases by 66% relative to energy and density supervision alone. The density-sensitive mean-field energy metric

E_ρ

improves from

1.2

to

0.8

mEh on average, while dipole and

\mathcal{L}_2

density errors do not improve uniformly. We further show that densities from the distilled functionals reduce hybrid-functional SCF iterations by up to 50%. In downstream TDDFT calculations, Hessian supervision improves excited-state predictions, with XCdiff reducing the mean excitation-energy MAE by 19 - 35%.

Eike S. Eberhard, Luca A. Thiede, Abdul Aldossary +5

Technical University of Munich, Munich, Germany · Munich Data Science Institute, Munich, Germany · Munich Center for Machine Learning, Munich, Germany +4

Quantum Chemistry Density Functional Theory Knowledge Distillation Computational Materials Science

Jun 2, 2026

Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories

The past few decades have witnessed significant advances in the design of machine learning algorithms, from early studies on task-specific shallow models to more general deep Large Language Models (LLMs). Despite showing promising results in tasks that require instant prediction or in-context learning, existing models lack the ability to continually learn and effectively transfer their temporal in-context knowledge to their long-term parameters. Inspired by human learning process, we introduce a ''Sleep'' paradigm that allows the models to continually learn, distill their short-term fragile memories into stable long-term knowledge with replay, and recursively improve themselves with ''Dreaming'' process. In more detail, sleep consists of two stages: (1) Memory Consolidation: an upward distillation process, called Knowledge Seeding, where the memories of a smaller-self are distilled into a larger network to provide more capacity while preserving the knowledge. As a proof of concept, we present a new Generalized Distillation process for {Knowledge Seeding} (i.e., the combination of on-policy distillation with Reinforcement Learning (RL)-based imitation learning); (2) Dreaming: a self-improvement phase, where the model uses RL to generate a curriculum of synthetic data to rehearse new knowledge and refine existing capabilities without human supervision. Our experiments on long-horizon, continual learning, knowledge incorporation, and few-shot generalization tasks support the importance of the sleep stage.

Ali Behrouz, Farnoosh Hashemi, Adel Javanmard +1

Google

Continual Learning for LLMs Persistent Memory for Language Models Replay-Based Continual Learning Language Model Self-Improvement Knowledge Distillation

Jun 2, 2026

AAD-1: Asymmetric Adversarial Distillation for One-Step Autoregressive Video Generation

We present AAD-1, an Asymmetric Adversarial Distillation framework for One-step autoregressive image-to-video generation. State-of-the-art methods adopt adversarial distillation but suffer from motion collapse and training instability, resulting in static videos. AAD-1 addresses these challenges through two key designs in architecture and training strategy. Our key architectural insight is to break the symmetry between generator and discriminator. While the generator remains causal to preserve autoregressive sampling capability, the discriminator attends bidirectionally over the full spatiotemporal context and produces a single holistic realism score for the entire video sequence. This asymmetric design enables the discriminator to effectively detect global temporal failures and long-range drift that cause motion collapse in autoregressive generation. To stabilize training, we introduce a phased strategy that first uses distribution matching to bootstrap a stable one-step generator, providing a warm-up phase that brings the student distribution closer to the teacher before adversarial distillation begins. Extensive experiments on VBench demonstrate that AAD-1 achieves state-of-the-art performance in one-step autoregressive video generation.

Haobo Li, Yanhong Zeng, Yunhong Lu +6

AutoLab, SAI, SJTU · Ant Group · Department of Automation, Tsinghua University +2

Image-to-Video Generation Autoregressive Video Generation Knowledge Distillation Temporal Consistency in Video Generation Adversarial Knowledge Distillation Single-Step Video Generation

Jun 2, 2026

BAHSD: Bridging the Long-tail Gap via Adaptive Distillation in Black-box Sequential Recommendation

Sequential recommendation systems are widely adopted but often deployed as black-box APIs, which has driven recent interest in model extraction to replicate their capabilities locally. However, the long-tail distribution induces severe signal heterogeneity: dense head sequences trigger the solidification of teacher preference, biasing extraction toward local patterns, while sparse tail sequences yield flat, noisy predictions. Existing one-size-fits-all extraction overlooks this disparity, resulting in noise overfitting and suboptimal knowledge transfer. We propose BAHSD, a black-box adaptive distillation framework that handles signal heterogeneity via a multi-scale consistency probing mechanism to implicitly quantify signal reliability. Based on this, an adaptive hierarchical objective is designed: dynamic-temperature KL divergence mitigates preference solidification for high-confidence signals, while ranking consistency and InfoNCE contrastive learning provide noise-robust enhancement for low-confidence signals. BAHSD consistently outperforms baselines, achieving up to 4.98% gain over the teacher and 80%+ improvement on tail users, offering a plug-and-play solution for high-fidelity black-box recommendation extraction.

Xi Zhou, Famin Wu, Mingming Li +4

Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China · Beijing Institute for General Artificial Intelligence, Beijing, China

Adaptive Knowledge Distillation Sequential Recommendation Long-Tail Learning Recommender Systems Knowledge Distillation

Jun 2, 2026

What Do Students Learn? A Feature-Level Analysis of Dark Knowledge

Knowledge Distillation (KD) is a powerful tool for model compression, yet the precise mechanisms by which student models acquire feature representations remain underexplored. In this work, we analyze student feature learning using the Interaction Tensor framework. Our analysis reveals that effective KD acts as a regularizer that prunes low-frequency, sample-specific features, encouraging the student to rely on a compact set of highly reusable features. Crucially, we observe that the dataset-level confusion matrix contains structural information analogous to the teacher's "Dark Knowledge." Leveraging this insight, we propose Confusion Distillation (CD), a teacher-free self-distillation method that utilizes the model's own evolving )confusion patterns as dynamic soft targets. CD achieves competitive performance on ResNet-34 and ResNet-50 for CIFAR-100, outperforming existing self-distillation methods like CS-KD and PS-KD by 1.2% while offering a computationally efficient alternative to standard KD.

Seungu Kang, Songkuk Kim

Yonsei University, Seoul 03722, Republic of Korea

Representation Learning Self-Distillation Teacher-Student Learning Knowledge Distillation Image Classification

Jun 1, 2026

Pathway-Structured Privileged Distillation for Deployable Computational Pathology

Integrating transcriptomics and histopathology can improve cancer risk modelling, yet practical use is constrained by the limited availability of RNA profiling in routine settings. Here we introduce Mixture of Pathway Experts (MoPE), a knowledge-distillation framework that reframes multimodal learning as privileged distillation for histology-only inference. MoPE is motivated by the partial observability between RNA profiles and whole-slide images: histology can capture morphology-linked consequences of certain molecular programmes, but cannot be expected to reconstruct the full transcriptomic state. MoPE encodes RNA-derived pathways and transfers the molecular supervision to pathway-indexed pathology experts through memory-usage alignment. Across diverse public benchmarks and two independent breast cancer cohorts, MoPE consistently improved WSI-only inference performance relative to baseline methods. Pathway-usage analyses and human-audited visual inspection provide bounded inspection of model behaviour and candidate morphology-linked readouts. These results support pathway-structured privileged distillation as a promising route to using molecular information during training while preserving RNA-free inference.

Yongxin Guo, Hao Lu, Onur Koyun +2

School of Medicine, Wake Forest University, Winston-Salem, NC, USA.

Computational Pathology Oncology Mixture-of-Experts Models Learning Using Privileged Information Knowledge Distillation Whole-Slide Image Analysis

Jun 1, 2026

Filter, Then Reweight: Rethinking Optimization Granularity in On-Policy Distillation

On-Policy distillation (OPD) in large language models is shifting from full-trace KL supervision toward more selective training paradigms. Recent OPD methods increasingly focus on selecting which trajectories to learn from, which tokens are most informative, and which supervision signals are most reliable. Motivated by this trend, we rethink optimization granularity of OPD and propose \fireicon\ FiRe-OPD (Filter, then Reweight), which jointly adjusts supervision signals at both trajectory and token levels. In details, FiRe-OPD first filters trajectories to remove low-quality rollout samples, and then applies soft reweighting within the retained trajectories to emphasize informative tokens. Compared with hard token selection, FiRe-OPD leverages a soft-weighting mechanism to effectively mitigate information loss and enhance optimization stability, thereby achieving finer-grained OPD optimization. We validate the effectiveness of FiRe-OPD across strong-to-weak, single-teacher, and multi-teacher settings, and demonstrate its superiority over recent token-level OPD methods ( (e.g., +6.25 on AIME 2024 in strong-to-weak, +18.81 on Miner in multi-teacher). Our code is available at https://github.com/YuYingLi0/FiRe-OPD.

Yuying Li, Leqi Zheng, Yongzi Yu +6

THU · Meituan · HKUST +2

Selective Knowledge Distillation Language Model Distillation On-Policy Distillation Knowledge Distillation

Jun 1, 2026

SafeSteer: Localized On-Policy Distillation for Efficient Safety Alignment

Aligning Large Language Models (LLMs) with human values often degrades their general capabilities, termed the alignment tax. Existing methods mitigate this by balancing dual objectives, which heavily rely on massive general-purpose data or auxiliary reward models. In this paper, we argue that, because safety features are inherently sparse within the output distribution, alignment requires localized modifications rather than global trade-offs. To this end, we propose SafeSteer, which performs on-policy distillation confined to safety tokens. First, we construct a safety teacher via activation steering. Based on this teacher, we develop a safety token selection algorithm. Consequently, SafeSteer restricts the reverse KL penalty to these tokens during training to preserve general capabilities. Experimental results across diverse models show that our SafeSteer achieves a superior trade-off between safety and general capability compared with existing methods, attaining strong safety performance on seven safety benchmarks with only minimal degradation on five general capability benchmarks. Notably, SafeSteer requires only 100 harmful samples without using any general-purpose data, less than 1% of what previous baselines used, considerably reducing alignment cost. More details are on our project page at https://anjingkun.github.io/SafeSteer.

Hao Li, Jingkun An, Zijun Song +8

Beihang University · Beijing Institute of Technology · Beijing University of Posts and Telecommunications +4

LLM Safety Alignment On-Policy Distillation Knowledge Distillation

Jun 1, 2026

Cross-Domain Dead Tree Detection via Knowledge Distillation in Aerial Imagery

Detecting dead trees in aerial imagery is vital for assessing forest health, especially as tree mortality increases globally due to climate change, but domain variability and scarce labeled data often limit model generalization. This study advances the TreeMort-1T-UNet (Tree Mortality 1-Task U-Net) model, initially trained on Finnish aerial imagery (source domain), by applying knowledge distillation (KD) to adapt it to various target domains, including Polish, German, and Estonian datasets representing diverse forest types. We assess four KD variants: Basic, Self, Feature-level, and Ensemble, against a fine-tuning baseline, using Mean Tree IoU, Instance F1-score, Instance Precision, and Mean Centroid Error as key metrics, alongside representational analyses (e.g., cosine similarity, CKA, SSIM, t-SNE, and linear probing) for domain invariance. Feature-level KD outperforms others, yielding a Mean Tree IoU of 0.106, Instance F1-score of 0.63, Instance Precision of 0.55, and Mean Centroid Error of 3.039 on the Polish dataset, with robust precision across other target domains (e.g., 0.15 on Finnish, 0.67 on Polish, 0.60 on German, 0.59 on Estonian). It excels in low-data scenarios with fewer false positives and shows superior representational invariance (e.g., higher deep-layer CKA/SSIM, better domain mixing in t-SNE, and linear probing AUC of 0.95), making it ideal for precision-critical forestry applications. Additional ablation studies confirm that key components like feature alignment enhance its performance balance across metrics. Our findings demonstrate KD's potential to enhance transfer learning in remote sensing, offering a scalable, domain-robust tool for ecological monitoring and sustainable forest management.

Anis Ur Rahman, Mete Ahishali, Einari Heinaro +1

CSC – IT Center for Science Ltd., Espoo, Finland · Department of Forest Sciences, University of Helsinki · KOKO Forest Ltd. +1

Remote Sensing Image Understanding Cross-Domain Transfer Learning Object Detection Knowledge Distillation Aerial Object Detection

Jun 1, 2026

Pool-Select-Refine for Allocation-Aware Generative Dataset Distillation

Diffusion-based dataset distillation has recently emerged as a promising paradigm for condensing large-scale datasets into compact synthetic sets. By leveraging pretrained generative priors, these methods can produce realistic class-conditional samples more efficiently than traditional matching-based approaches. However, most existing diffusion-based methods still adopt a rigid

Generate-and-Use'' strategy, where the generated samples are directly treated as the final distilled set under a fixed images-per-class budget. Such a design tightly couples candidate generation with final budget allocation, which may result in redundant waste of the limited budget or insufficiently informative samples. In this paper, we propose

Pool-Select-Refine'', a two-stage framework for allocation-aware generative dataset distillation. First, instead of directly using a fixed number of generated samples, we construct an over-complete candidate pool and select a compact subset under the target budget. Second, we refine the selected samples in latent space using soft-label supervision derived from the teacher model, improving semantic alignment while preserving the generative prior. This design explicitly decouples generation, selection, and refinement, enabling more effective use of the distillation budget. Experiments on large-scale and fine-grained image classification benchmarks show that the proposed framework delivers consistent gains over diffusion-based baselines. The results suggest that introducing a curation stage before refinement is a simple yet effective way to improve diffusion-based dataset distillation.

Wenmin Li, Shunsuke Sakai, Zhongkai Zhao +1

Graduate School of Engineering, University of Fukui, 3-9-1 Bunkyo, Fukui, 910-8507, Fukui, Japan · College of Computer Science and Artificial Intelligence, Southwest Minzu University, No.168, Wenxing Section, Dajian Road, Konggang Development Zone, Shuangliu District, Chengdu, 610225, Sichuan, China

Dataset Distillation Knowledge Distillation Synthetic Data Selection Diffusion-Based Dataset Distillation

May 31, 2026

What Makes a Strong Model? A Unified Spectral Analysis of Knowledge Transfer over High-dimensional Linear Regression

Teacher-Student Knowledge Transfer (KT) is ubiquitous in modern machine learning, ranging from classical model compression via Knowledge Distillation (KD) to the emergent phenomenon of Weak-to-Strong (W2S) generalization. While existing studies offer isolated insights, a unified theoretical framework explaining the efficacy of KT across these disparate regimes remains lacking. In this work, we establish a unified spectral analysis of SGD dynamics in high-dimensional linear regression, elucidating the efficiency of KT across seemingly disparate regimes. We characterize KT efficiency through two distinct mechanisms: Spectral Horizon Expansion in KD, which enables the capture of statistically inaccessible high-frequency signals, and Spectral Denoising in W2S, where the student acts as a filter for optimization noise. Our framework unifies these phenomena, revealing that the efficacy of transfer is governed by the interplay between implicit regularization and heterogeneous spectral learning speeds over the spectrum.

Wendao Wu, Fangqing Zhang, Haihan Zhang +1

State Key Lab of General AI, School of Intelligence Science and Technology, Peking University, China · School of Mathematical Sciences, Peking University, China

High-Dimensional Linear Regression Weak-to-Strong Generalization Teacher-Student Learning Knowledge Distillation Gradient Descent Dynamics

May 31, 2026

SWARD: Stochastic Window-Attention-Based Relational Distillation for Cross-Architectural Semantic Segmentation

Large-scale vision foundation models have driven substantial gains on dense prediction tasks such as semantic segmentation, but their size makes deployment impractical in resource-constrained settings, motivating knowledge distillation as a means of transferring their capabilities to lightweight student networks. However, modern foundation teachers are predominantly transformer-based that encode global context, whereas efficient students are typically convolutional networks with locally biased receptive fields. Existing distillation methods largely assume architectural homogeneity and rely on direct feature mimicry, which fails to bridge this representational gap and neglects the structured spatial dependencies and discriminative organization required for accurate semantic segmentation. In this paper, we propose SWARD, a knowledge distillation framework that addresses this gap through two complementary mechanisms. First, we introduce a Multi-Scale Windowed Attention Distillation (MWAD) module that aligns teacher-student attention-based relations within stochastically shifted window partitions whose offsets are randomly resampled at every training iteration. This removes window boundary bias, and, combined with the multi-scale design, captures both short- and long-range spatial dependencies. Second, we introduce Prototype Discriminative Regularization (PDR), a loss that helps shape the student's feature distribution by enforcing inter-class separation and intra-class compactness, further sharpening the discriminative structure beyond what feature mimicry alone can produce under the student's reduced capacity. Experiments across different vision applications (i.e., urban scene parsing and medical image segmentation) show that SWARD achieves state-of-the-art performance.

Aditya Makineni, Qing Tian

Department of Computer Science University of Alabama at Birmingham Birmingham, US

Cross-Architecture Knowledge Distillation Relational Knowledge Distillation Knowledge Distillation Semantic Segmentation

May 29, 2026

Rethinking the Role of Temperature in Large Language Model Distillation

Reverse Kullback-Leibler (RKL) divergence is widely favored over forward KL (FKL) in large language models (LLM) distillation, yet this preference is largely based on comparisons that omit the temperature

τ

, overlooking its central role in softening teacher distributions and improving knowledge transfer. In this work, we revisit temperature in LLM distillation and show that it fundamentally changes the comparison between FKL and RKL. Our analysis reveals an asymmetric effect: temperature substantially enriches FKL with non-dominant token signals, whereas it mainly rescales RKL gradients, causing FKL to benefit much more from

τ

scaling than RKL. This asymmetry overturns the standard empirical conclusion: although RKL outperforms FKL at

τ=1

, FKL consistently surpasses RKL at higher temperatures across instruction-following benchmarks. Moreover, the impact of temperature is not limited to FKL; it improves a broader family of distillation objectives, enabling simple KL-based methods to achieve competitive performance against recent state-of-the-art LLM distillation approaches.

Hoang-Chau Luong, Lingwei Chen

Golisano College of Computing and Information Sciences Rochester Institute of Technology Rochester, NY, United States

Language Model Distillation Knowledge Distillation Kullback-Leibler Divergence

May 29, 2026

Bridging Reasoning Trajectories in On-Policy Distillation via Near-Future Guidance

On-Policy Distillation (OPD) improves large language model reasoning by training a student model on trajectories sampled from its own policy under teacher supervision. Although OPD operates on trajectories, its learning signal remains token-level: it identifies deviations through high-loss tokens and repairs them through local reverse-KL correction. We show that this "trajectory-sampled but token-learned" mechanism cannot reliably bridge student trajectories toward teacher trajectories. About 30% of high-loss tokens fall into the low-divergence regime, indicating that many are surface-form mismatches rather than real reasoning forks. Moreover, even truly divergent tokens are difficult to repair with isolated token-level supervision, since reasoning failures often unfold as short-horizon distributional drift. We propose Trajectory-aware OPD (TOPD), which uses near-future trajectory information to identify real divergent states and distribute guidance across multiple future tokens. Experiments show that suppressing non-divergent high-loss tokens improves standard OPD from 47.8% to 48.2% average accuracy, while TOPD further improves performance to 52.2%, with gains on AIME24 from 60.0% to 63.3% and AIME25 from 46.7% to 53.3%.

Yuxuan Jiang, Francis Ferraro

University of Maryland, Baltimore County

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

May 29, 2026

COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation

LLM agents are increasingly expected not only to complete isolated tasks, but also to carry bounded representations of human expertise, judgment, and interaction style. Building such person-grounded agents remains difficult because actionable knowledge associated with a person or role is usually embedded in heterogeneous traces rather than written as clean instructions. Existing memory and persona systems capture fragments of this evidence, while skill frameworks provide portable packaging formats; however, there is no end-to-end workflow for distilling these traces into inspectable, correctable, and agent-usable skills. We present an automated trace-to-skill distillation system for generating person-grounded AI skills via expert knowledge distillation. Given materials from a target person or role, COLLEAGUE.SKILL produces a versioned skill package with two coordinated tracks: a capability track for practices, mental models, and decision heuristics, and a bounded behavior track for communication style, interaction rules, and correction history. The package can be inspected, invoked, updated through natural-language feedback, rolled back, installed across agent hosts, and optionally prepared for controlled distribution. We describe the artifact contract, generation workflow, correction lifecycle, deployment surface, and domain presets implemented in the open-source system. At the time of writing, the public repository has approximately 18.5k GitHub stars; the gallery lists 215 skills from 165 contributors and more than 100k cumulative stars across listed skill cards. The system illustrates how person-grounded skills can be represented as portable, correctable packages rather than opaque prompts or hidden memories.

Tianyi Zhou, Dongrui Liu, Leitao Yuan +2

Shanghai Artificial Intelligence Laboratory

LLM Agent Skill Learning Persona-Conditioned Generation Knowledge Distillation Personal AI Agents

May 29, 2026

Student Capacity Moderates Knowledge Distillation Effectiveness: A Systematic Study Across ResNet Teacher-Student Pairs on CIFAR-10

We investigate how teacher-student capacity relationships modulate knowledge distillation (KD) effectiveness in ResNet-based image classification on CIFAR-10. Across four teacher-student pairs (R50->R18, R34->R18, R50->R34, and R101->R34) we compare Logit-KD and Feature-KD under a strict evaluation protocol: hyperparameters and checkpoints are selected on a held-out validation split, selected configurations are re-run with five seeds, and the test set is used exclusively for final reporting. Beyond accuracy, we measure distillation fidelity directly via teacher-student agreement and KL divergence. We report four findings. First, the student-capacity pattern survives the corrected protocol at reduced magnitude: the only statistically significant gains occur for R34 students under Feature-KD (+0.19 and +0.21 pp, p<0.05), in two pairs whose teachers differ two-fold in parameters but not in accuracy, localizing the moderating variable on the student side, while no KD gain for R18 students is distinguishable from zero. Second, Feature-KD matches or outperforms Logit-KD in all four pairs, and its students land closer to the teacher's output distribution (KL at T=1) than Logit-KD students despite never observing teacher logits. Third, top-1 teacher-student agreement is flat across all pairs, decoupling fidelity from accuracy gains. Fourth, architecture dominates KD: correcting the ResNet stem for 32x32 inputs is worth +5.5 to +7.2 pp, more than 25x the largest KD gain. We also retract an attribution made in v1: a controlled re-run shows the reported gradient-clipping bug had no measurable effect, and v1's larger gains are explained by test-set selection. Code and results: github.com/umutonuryasar/kd-capacity-gap (tag v2.0).

Umut Onur Yasar

Teacher-Student Learning Feature-Based Knowledge Distillation Knowledge Distillation Image Classification

May 29, 2026

Trust-Region Behavior Blending for On-Policy Distillation

On-policy distillation (OPD) trains a student on prefixes sampled from its own policy while matching a stronger teacher. This addresses the prefix mismatch of offline distillation, but early student rollouts can still be poor, placing teacher supervision on weak or low-quality prefixes. We propose Trust-Region behavior Blending (TRB), a warmup method that replaces the early rollout policy with the closest-to-teacher behavior policy inside a student-centered KL trust region, while keeping the per-prefix reverse-KL OPD loss unchanged. The KL budget is annealed to zero, so training returns to pure student rollouts after warmup. Across two math-reasoning distillation settings, TRB attains the strongest average among the compared methods.

Daniil Plyusov, Alexey Gorbatovski, Alexey Malakhov +4

T-Tech

Language Model Distillation Trust-Region Optimization On-Policy Distillation Teacher-Student Learning Policy Learning Knowledge Distillation

May 29, 2026

DEM: A Distilled Explanation Model for Interpretable Anomaly Detection in Physiological Sensor Networks

Anomaly detection in physiological sensor data from Wireless Body Area Networks (WBANs) can be caused by sensor faults, network disruptions, or missing data, leading to false alarms. Hence, it demands both high predictive accuracy and clinically interpretable explanations. Existing approaches rely either on black-box models that achieve strong performance but offer no transparency, or on post-prediction explanation methods such as SHAP and LIME. In this paper, we propose the Distilled Explanation Model (DEM), a three-stage glass-box framework that distills the non-linear knowledge of a gradient boosting expert into an interpretable decision tree operating on residuals relative to a linear baseline, so that the explanation is not an approximation but the prediction itself. DEM introduces a novel distillation fidelity metric that quantifies how faithfully the explanation tree captures the expert model's non-linear contribution, providing a principled measure of explanation trustworthiness absent from prior interpretable models. Evaluated across four physiological datasets, including MIMIC-IV, WESAD, eICU, and an in-house SmartNet WBAN corpus, DEM achieves an AUC of 0.9964 on clinical contextual anomaly detection and 0.9047 on wearable stress detection while producing human-readable if-then rules at a controllable depth. Inference requires 0.17ms per 1000 samples, rendering DEM 1235x faster than SHAP-based post-hoc explanation and suitable for real-time physiological monitoring. Ablation studies confirm that the XGBoost distillation step provides measurable gains over naive residual fitting, and depth-sensitivity analysis demonstrates an explicit, user-controlled accuracy-interpretability trade-off unique to DEM among existing intrinsically interpretable models.

Jyotirmoy Singh, Anushka Roy, Shreea Bose +1

Interpretable ML Interpretable Anomaly Detection Explainability Evaluation Knowledge Distillation Anomaly Detection

May 28, 2026

EVL-ECG: Efficient ECG Interpretation With Multi-Aspect Heterogeneous Knowledge Distillation

High-fidelity ECG interpretation is increasingly reliant on massive foundation models, yet their deployment in clinical edge-care remains hindered by extreme computational demands. While knowledge distillation (KD) is a promising solution, traditional methods fail to capture the complex spatio-temporal dependencies of ECG signals when transferring knowledge across heterogeneous architectures. In this paper, we propose EVL-ECG, a framework specifically designed for cross-architecture distillation of cardiac diagnostic logic. EVL-ECG introduces three ECG-aware innovations: (1) Multi-Head Cross-Attention Alignment, which harmonizes architectural discrepancies to preserve fine-grained morphological features; (2) Optimal Transport-based Visual Feature Matching, utilizing optimal transport to maintain global structural relationships across ECG leads despite mismatched token representations; and (3) Geometric Intra-Architecture Relation Matching, which distills the latent diagnostic reasoning of the teacher model. Evaluations across ECG benchmarks demonstrate that EVL-ECG yields improvements of up to 2.4% AUC and 1.1% clinical accuracy over existing baselines. Notably, EVL-ECG establishes an efficient 2B-parameter ECG foundation model, suitable for resource-constrained clinical environments.

Dang Nguyen Hong, Nhi Ngoc-Yen Nguyen, Huy-Hieu Pham

Hanoi University of Science and Technology, Hanoi, Vietnam · VinUni-Illinois Smart Health Center, VinUniversity, Hanoi, Vietnam · College of Engineering and Computer Science, VinUniversity, Hanoi, Vietnam +1

Efficient Neural Network Inference Electrocardiogram Foundation Models Electrocardiogram Classification Knowledge Distillation Time-Series Foundation Models