Knowledge Distillation

Ask about this topic

Overview

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

CardsList

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

May 28, 2026

SLAD : Shared LoRA Adapters for Task Specific Distillation

In the context of resource-constrained environments such as embedded systems, adapting reduced-size foundation models to downstream tasks has become increasingly popular. This has recently motivated the emerging setting of task-specific distillation, where a larger and a smaller version of the same foundation model are both adapted to the same downstream task, with the goal of transferring knowledge from the former to the latter. Recent work has demonstrated the benefits of using a larger version of the same foundation model to assist the adaptation of a smaller one. Typically, the larger model (teacher) is first adapted via fine-tuning or linear probing before its knowledge is distilled into the smaller model (student). While fine-tuning the teacher often increases its performance, recent work showed that probing it leads to better knowledge distillation to the student. Our findings show that this is mainly due to a mis-alignment in feature representation between the teacher and the student which occurs during the teacher's fine-tuning. Inspired by existing efforts to preserve previously learned knowledge, we first propose leveraging low-rank adaptation, resulting in better feature alignment and therefore better knowledge transfer. Drawing from this insight, we further enhance the feature alignment through a parameter-sharing strategy of the adapters between the two encoders during joint training. Our proposed method, SLAD, shows better feature alignment between the teacher and student, which results in increased performance for not only the student but also the teacher model, while being 2x faster to train than fine-tuning. Through extensive experiments on multiple classification and segmentation datasets, we demonstrate the improved accuracy and transfer efficiency of our method, achieving state-of-the-art performance in the task-specific distillation framework.

Reda Bensaid, Yassir Bendou, Vincent Gripon +1

IMT Atlantique, Brest, France · Polytechnique Montr´eal, Montr´eal, Canada

Vision Foundation Model AdaptationAdapter TuningLow-Rank AdaptationKnowledge DistillationRepresentation Alignment
AskPDF
May 28, 2026

Matching Rates and Optimal Allocation for Federated Probe-Logit Distillation under Heterogeneous Bandwidth Budgets

In federated language modeling, KK nodes each hold nn samples but cannot pool data or exchange full-precision gradients or weights. We study the minimax rate at which a conditional distribution over VV tokens can be estimated when each node may upload at most BB bits per query in a public probe set. In federated probe-logit distillation (FPLD), each node transmits a scalar-quantized logit vector on the probe set, and an aggregator distills a global parametric student. Prior work (Dubey and Huo, 2026) establishes a high-probability KL rate O(d/(Kn)+ρVlog⁡V/m+K−1⋅2−2B/V)O(d/(Kn) + ρ\sqrt{V \log V / m} + K^{-1} \cdot 2^{-2B/V}) plus optimization slack, with the bandwidth term in its trace-sharpened form. Whether this bandwidth-term rate is tight, and how the upper bound generalizes to heterogeneous per-node bandwidths, are left open. We close both gaps. First, the dithered FPLD construction has a matching single-round lower bound Ω(K−1⋅2−2B/V)Ω(K^{-1} \cdot 2^{-2B/V}) under non-degeneracy, pinning the bandwidth-axis rate at Θ(K−1⋅2−2B/V)Θ(K^{-1} \cdot 2^{-2B/V}). TT-round sequential refinement with nested/scaled residual quantizers achieves O(K−1⋅2−2TB/V)O(K^{-1} \cdot 2^{-2TB/V}); vanilla FPLD's TT-independent bandwidth term is suboptimal for every T>1T > 1. Second, we establish a heterogeneous-bandwidth upper bound for per-node budgets BiB_i, paired with a closed-form optimal allocation Bi∗=Btot/K+(V/2)log⁡2(wi/wˉg)B_i^* = B_{\mathrm{tot}}/K + (V/2) \log_2(w_i / \bar{w}_g), a log-tilted water-filling rule that is the per-node analogue of reverse water-filling for distortion-rate optimization. A plug-in adaptive variant estimates the weights from a short warm-up phase and attains 1+O(log⁡(K/δ)/(mT0))1 + O(\sqrt{\log(K/δ)/(m T_0)}) relative suboptimality. Synthetic n-gram simulations confirm that empirical KL is bracketed by the upper and lower bounds and that the optimal allocation strictly dominates uniform and inverse-weighted baselines under heterogeneous clipping.

Prasanjit Dubey, Xiaoming Huo

H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology, Atlanta, GA 30332, U.S.A.

Federated Learning AggregationCommunication-Efficient Distributed TrainingFederated Knowledge DistillationRate-Distortion OptimizationKnowledge DistillationHeterogeneous Federated Learning
AskPDF
May 28, 2026

GAPD: Gold-Action Policy Distillation for Agentic Reinforcement Learning in Knowledge Base Question Answering

Reinforcement learning (RL) is a natural fit for agentic knowledge base question answering (KBQA), where a model must issue executable actions, observe knowledge-base feedback, and eventually return an answer. However, current RL-based KBQA systems mainly optimize sparse rewards from the final answer, leaving intermediate action errors weakly supervised. This is especially limiting for logical-form annotated KBQA benchmarks: gold logical forms can be converted into executable action sequences, but existing pipelines use them mainly for warm-start data construction rather than for on-policy RL updates. We propose GAPD, a training-time Gold-Action Policy Distillation framework that adds dense token-level guidance to outcome-based RL. To align gold actions with on-policy student rollouts, GAPD uses MID-ANCHOR MATCHING: it treats the intermediate entities reached during student exploration and gold execution as state anchors, and matches student states to gold states through these explored entity sets. The current policy conditioned on this aligned gold action serves as a stop-gradient teacher, whose token distribution is distilled back to the ordinary student policy over generated action-token spans. GAPD consistently surpasses the current state of the art on WebQSP, GrailQA, and GraphQ.

Xin Sun, Jianan Xie, Zhongqi Chen +6

University of Science and Technology of China · 2NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences · 3ShanghaiTech University +1

Agentic RLKnowledge Graph Question AnsweringPolicy DistillationKnowledge Distillation
AskPDF
May 28, 2026

LoopFM: Learning frOm HistOrical RePresentations of Foundation Model for Recommendation

Knowledge distillation (KD) transfers a single scalar prediction from a large foundation model (FM) to compact vertical models (VMs), suffering from diminishing transfer ratio -- the fraction of FM improvement captured by the VM -- as a single scalar cannot convey the rich intermediate knowledge that larger FMs learn. To address this bottleneck, we propose LoopFM (Learning frOm HistOrical RePresentations of FM), a framework that opens a high-bandwidth transfer channel by structuring FM intermediate embeddings as input features (e.g., user history sequence) for downstream VMs, without requiring real-time FM inference at serving and architectural coupling between FM and VM. We provide a theoretical framework for LoopFM with a gain decomposition and transfer-ratio analysis. On three public benchmarks, LoopFM demonstrates strong AUC improvements (e.g., 6%+ on TaobaoAd) and complementary knowledge transfer capability with KD. On industrial-scale systems (billions of examples, trillion-parameter FMs), LoopFM approximately doubles the knowledge transfer ratio on top of KD, delivering a +0.5% conversion improvement in the first half after its initial launch, and +1.03% and +1.22% conversion improvement from two individual launches in the subsequent half. Through systematic experiments, LoopFM demonstrates a scaling law in sequence length, embedding dimension, and upstream FM size.

Hua Zheng, Shali Jiang, Boyang Liu +44

Transfer LearningRecommender SystemsKnowledge Distillation
AskPDF
May 27, 2026

Lightweight SAR Ship Detection via Contrastive Distillation

Deep convolutional and transformer-based detectors achieve strong performance for SAR ship detection but are often computationally prohibitive for real-time or onboard deployment. Lightweight models offer improved efficiency yet struggle to capture the complex structural relationships inherent in SAR backscatter. Most existing SAR knowledge-distillation approaches rely on feature or logit matching, which enforces localized activation similarity while neglecting the geometric relationships among object representations. We propose a Structured Unified Relational knowledGE distillation framework for SAR Ship detection (SURGE) that transfers relational geometry from a powerful teacher detector to a compact student detector using a contrastive InfoNCE objective in a shared projection embedding space. To the best of our knowledge, this work presents the first transformer-based SAR ship detector knowledge distillation framework in SAR domain. The framework is architecture-agnostic in the sense that it provides a common region-level distillation interface for two-stage, one-stage and transformer-based detectors without modifying their deployed architectures. Experiments on the SSDD and HRSID benchmarks demonstrate that the proposed method yields substantial improvements for two-stage detectors, achieving up to 6.2 mAP and 8.0 AP75 gains over baseline student and even surpassing teacher performance

Surendar Devasundaram, Banafsheh Saber Latibari, Abhijit Mahalanobis

University of Arizona · Department of Electrical and Computer Engineering · Tucson, Arizona, USA

Efficient Object DetectionRelational Knowledge DistillationObject DetectionSynthetic Aperture RadarKnowledge DistillationShip Detection
AskPDF
May 27, 2026

Skill-Conditioned Gated Self-Distillation for LLM Reasoning

On-policy self-distillation (SD) improves LLM reasoning by using teacher-side privileged information (PI) to turn sparse verifier outcomes into dense token-level supervision. Existing methods usually assume trusted PI, such as reference answers or successful traces. We ask whether PI can instead come from an experience-derived skill bank, where retrieved skills are compact and reusable but may also be irrelevant or misleading. We propose Skill-Conditioned Gated Self-Distillation (SGSD), which formulates skill-based SD as teacher hypothesis validation rather than unconditional imitation. SGSD retrieves skill-mistake pairs, constructs a multi-teacher pool, and lets all skill-conditioned teachers score the same plain-prompt student rollout. The verifier validates each teacher's polarity: supporting a success or suppressing a failure gives positive supervision, while the opposite stance is reversed. A robust gated objective then distills informative teacher-student disagreements while suppressing uncertain or extreme signals. Experiments on multiple mathematical reasoning benchmarks show that SGSD consistently improves over GRPO and remains competitive with answer-conditioned OPSD under a weaker PI assumption. For example, on Qwen3-1.7B, SGSD outperforms GRPO by 6.2% and OPSD by 1.7% on average on AIME24, AIME25, and HMMT25.

Jiazhen Huang, Xiao Chen, Xiao Luo +3

Tsinghua University · University of Wisconsin-Madison · X-Humanoid +2

Mathematical ReasoningSelf-DistillationOn-Policy DistillationTeacher-Student LearningKnowledge DistillationLLM Reasoning
AskPDF
May 27, 2026

Deformable Gaussian Occupancy: Decoupling Rigid and Nonrigid Motion with Factorized Distillation

Understanding dynamic 3D environments is essential for safe autonomous driving, particularly when reasoning about human-centric, nonrigid agents. However, existing weakly supervised occupancy prediction frameworks predominantly assume rigid-body motion and rely on simple frame-to-frame offsets, limiting their ability to capture fine-grained deformations and maintain temporal coherence. To address this issue, we propose DeGO, a deformable Gaussian occupancy framework that unifies decoupled Gaussian deformation with factorized 4D foundation-model distillation. DeGO disentangles rigid and nonrigid motion, enabling each Gaussian primitive to evolve through both deformation and offset-based updates. In parallel, a factorized 4D distillation strategy transfers cross-camera and cross-frame knowledge from the VGGT foundation model, producing foundation-aligned features that enhance temporal consistency. Experiments on the Occ3D-NuScenes benchmark demonstrate that our method achieves state-of-the-art performance under weak supervision, delivering 13.5% gains on human-centric instances and 10.9% overall improvements. These results highlight the effectiveness of deformation-aware and foundation-guided occupancy modeling for dynamic scene understanding. The code is publicly available: https://github.com/vita-epfl/DeGO

Yang Gao, Wuyang Li, Po-Chien Luan +1

´Ecole Polytechnique F´ed´erale de Lausanne (EPFL), Switzerland

3D Gaussian SplattingAutonomous Driving Perception3D Occupancy PredictionKnowledge Distillation3D Scene Understanding
AskPDF
May 27, 2026

Dual-branch Distilled Transformer for Efficient Asymmetric UAV Tracking

Given the real-time demands of UAV tracking, many methods simplify the backbone to reduce computation, but this often weakens feature representation and degrades performance in complex scenarios. To alleviate this issue, we propose EATrack, an efficient and asymmetric UAV tracking framework centered around a teacher-guided dual-branch distillation strategy that enhances the feature expressiveness of the lightweight student model. Specifically, EATrack investigates two complementary perspectives of knowledge transfer: spatially focused feature-level distillation that compensates for weakened representations by guiding the student to learn strong target representations, and prediction-level distillation that enhances spatial localization by learning the teacher's capability for accurate target localization. Furthermore, to enhance robustness against appearance variations, we introduce a fine-grained target-aware distillation strategy that selectively transfers the teacher's target modeling capacity to the student. A temporal adaptation module is incorporated at inference to enhance robustness over time. Experiments on five UAV benchmarks demonstrate that EATrack achieves a favorable balance between accuracy and speed. Code: https://github.com/GXNU-ZhongLab/EATrack

Hongtao Yang, Bineng Zhong, Qihua Liang +4

Key Lab of Education Blockchain and Intelligent Technology, Ministry of Education, Guangxi Normal University, Guilin, 541004, China · Guangxi Key Lab of Multi-Source Information Mining and Security, Guangxi Normal University, Guilin, 541004, China · Nanjing University of Science and Technology +1

Visual Object TrackingEfficient ViTsKnowledge Distillation
AskPDF
May 27, 2026

Multi-Teacher Knowledge Distillation via Teacher-Informed Mixture Priors

Knowledge distillation is a powerful method for model compression, enabling the efficient deployment of complex deep learning models (teachers), including large language models. However, its underlying statistical mechanisms remain unclear, and uncertainty evaluation is often overlooked, especially in real-world scenarios requiring diverse teacher expertise. To address these challenges, we introduce Multi-Teacher Bayesian Knowledge Distillation (MT-BKD), where a distilled student model learns from multiple teachers within the Bayesian framework. Our approach leverages Bayesian inference to capture inherent uncertainty in the distillation process. We introduce a teacher-informed prior, integrating external knowledge from teacher models and task-specific training data, offering better generalization, robustness, and scalability. Additionally, an entropy-based weighting mechanism adaptively adjusts each teacher's influence, allowing the student to combine multiple sources of expertise effectively. MT-BKD enhances the interpretability of the student model's learning process, improves predictive accuracy, and provides uncertainty quantification. We validate MT-BKD on both synthetic and real-world tasks, including protein subcellular location prediction and image classification. Our experiments show improved performance and robust uncertainty quantification, highlighting the strengths of our MT-BKD framework.

Luyang Fang, Yongkai Chen, Jiazhang Cai +2

Department of Statistics, University of Georgia · Department of Statistics, Harvard University

Uncertainty QuantificationUncertainty-Aware Knowledge DistillationMulti-Teacher Knowledge DistillationKnowledge Distillation
AskPDF
May 26, 2026

When the Strongest Teacher Is Not the Best Teacher: Student-Centric Answer Selection

LLM training increasingly relies on teacher-generated supervision, from synthetic responses to reasoning traces and tool-use demonstrations. Current practice often chooses the highest-performing teacher to generate student training data, implicitly treating teacher test performance as a proxy for teaching quality. We show that this assumption can fail: even when multiple teachers provide correct answers to the same question, the answer from the strongest teacher is not necessarily the best supervision for a given student. To address this gap, we propose Student-Centric Answer Sampling (SCAS), a framework that selects from verified teacher-generated answers according to their estimated student-centric learning cost. Motivated by a token-wise gradient decomposition, we derive an efficient forward-only proxy for this cost and use it to guide answer selection during training. Experiments across 30 teacher models, 6 student base models, and 6 tasks show that SCAS consistently improves student performance, suggesting that effective distillation should prioritize supervision matched to the current student rather than teacher strength alone.

Zhengyu Hu, Zheyuan Xiao, Linxin Song +10

University of Washington · University of Texas at Austin · University of Southern California +6

Language Model DistillationLLM Post-TrainingTeacher-Student LearningKnowledge Distillation
AskPDF
May 26, 2026

Generalist Graph Anomaly Detection via Prototype-Based Distillation

Driven by the pressing demand for graph anomaly detection (GAD) in high-stakes domains, the generalist GAD paradigm, which trains a single detector transferable across new graphs, has recently gained growing attention. However, existing methods often rely on scarce and costly annotations for training and sometimes even require few-shot support at inference, which limits their robustness to diverse and unseen anomaly patterns. To address this limitation, we introduce ProMoS, the first unsupervised generalist GAD framework, which detects anomalies by modeling the abundant normality in unlabeled data. ProMoS adopts a knowledge-distillation paradigm to distill normality priors from a frozen self-supervised graph neural network (GNN) teacher to a mixture-of-students model with shared global and lightweight personalized branches, enabling efficient and expressive normality modeling without learning from scratch. We further propose prototype-guided soft-label distillation to align teacher and student in a shared prototype space, enhancing cross-graph generalizability. During inference, ProMoS performs zero-shot anomaly detection on unseen graphs via distillation bias and prototype geometric deviation. Extensive experiments show the effectiveness and efficiency of ProMoS, charting a practical path toward label-free, zero-shot generalist GAD.

Yiming Xu, Zihan Chen, Zhen Peng +4

School of Computer Science and Technology, Xi’an Jiao-tong University, Xi’an, China · National Engineering Research Center for Visual Information and Applications, Xi’an, China · University of Virginia, Charlottesville, USA +3

Unsupervised Anomaly DetectionZero-Shot Anomaly DetectionKnowledge DistillationAnomaly DetectionGraph-Based Anomaly Detection
AskPDF
May 26, 2026

Not All Disagreement Is Learnable: Token Teachability in On-Policy Distillation

On-policy distillation (OPD) trains a student on its own rollouts with token-level teacher supervision. Recent selective OPD methods exploit the non-uniformity of OPD signals by prioritizing high-entropy or high-disagreement tokens. We revisit this principle and ask: which token-level teacher signals are actually learnable? Using a fixed-context diagnostic that measures same-context teacher-student KL reduction, we show that raw KL disagreement is a coarse proxy for learning value. It conflates learnable disagreement, where the teacher assigns corrective mass to the student's top-K candidates, with incompatible disagreement, where the teacher places mass mostly off the student's current support. We formalize this local compatibility as token teachability and show that it better predicts fixed-context improvement than raw KL alone. Motivated by this finding, we propose Teachability-Aware OPD (TA-OPD), a lightweight token-position selection method that applies OPD loss to high-teachability positions without reward models or verifiers. Across Qwen2.5 and Qwen 3 teacher-student settings, TA-OPD often surpasses full-token OPD with only 5% retained tokens and improves over entropy- and divergence-based baselines. Our results reframe selective OPD as selecting learnable teacher signals rather than merely salient tokens.

Yuanyi Wang, Su Lu, Yanggan Gu +6

The Hong Kong Polytechnic University, PolyU · Hong Kong Polytechnic University Daya Bay Technology and Innovation Research Institute

Selective Knowledge DistillationLanguage Model DistillationOn-Policy DistillationTeacher-Student LearningKnowledge Distillation
AskPDF
May 25, 2026

Detail Consistent Stage-Wise Distillation for Efficient 3D MRI Segmentation

Deploying high-performing 3D medical image segmenters (e.g., nnU-Net) is often limited by memory footprint and inference latency. Compression is therefore necessary, but compact 3D encoders tend to lose fine structural cues (small lesions and sharp boundaries) as downsampling repeats across multi-resolution stages. We propose Detail Consistent Distillation (DCD), a stage-wise distillation framework that preserves structural detail across scales by aligning teacher-student features in a wavelet-decomposed representation. At each encoder stage, DCD distills directional detail components in the wavelet domain while leaving the coarse approximation comparatively unconstrained, avoiding over-regularization of global semantics. DCD is used only during training and introduces no inference-time overhead. Experiments on the BraTS 2024 and ISLES 2022 benchmarks demonstrate that our approach achieves superior performance in MRI segmentation using 3D multi-modal data. Code and implementation details for DCD are publicly available at https://github.com/ClinicaAlpha/DCD-3D-MedSeg.

Mengchen Fan, Baocheng Geng, Xi Xiao +5

University of Alabama at Birmingham, Birmingham, AL, USA · Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany · UTHealth Houston, Houston, TX, USA

Image SegmentationMedical Image Segmentation3D Medical Image SegmentationKnowledge Distillation
AskPDF
May 25, 2026

The Bridge-Garden Dilemma in LLM Distillation: Why Mixing Hard and Soft Labels Works

Knowledge distillation (KD) transfers knowledge from a large teacher model to a smaller student. In language modeling, the student is trained either on tokens sampled from the teacher (hard labels) or the teacher's full next-token distribution (soft labels). Despite soft labels appear strictly richer, we find that mixing hard and soft labels consistently yields better results. Crucially, we show that this gain cannot be explained by closer teacher matching during training. Instead, it comes from reduced exposure bias, the mismatch between training and inference distributions. To explain this phenomenon, we introduce the Bridge-Garden Decomposition theory, which categorizes generation steps into two types: Bridges, where the next token must be exact, and Gardens, where it can be flexible. We show that hard-only KD excels in Bridges by avoiding risky deviations, while soft-only KD preserves diversity in Gardens. A hybrid strategy handles both cases and, as a result, reduces exposure bias across the sequence. Guided by this theory, we develop a family of Bridge-Garden hybrid supervision methods that adaptively balance hard and soft labels. Across a primary suite of seven teacher-student pairs (including Qwen, Llama, Gemma, and DeepSeek) and benchmarks in reasoning and coding, our approach outperforms divergence-based and on-policy KD baselines while reducing training cost by 9.7x, enabling efficient model compression. Code is available at https://github.com/ghwang-s/bridge_garden_hybrid_kd_release.

Guanghui Wang, Kaiwen Lv Kacuila, Zhiyong Yang +5

School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing, China · Alibaba Group, Hangzhou, China · State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China +2

LLM CompressionLanguage Model DistillationExposure BiasKnowledge Distillation
AskPDF
May 25, 2026

On-Policy Adversarial Flow Distillation for Autoregressive Video Generation

Autoregressive video generators are attractive for streaming, long-horizon, and interactive applications, but distilling strong black-box teachers into causal students remains difficult. The student must learn under its own rollout distribution, whereas practical teachers may expose only prompt-conditioned completed videos and may differ in architecture, capacity, temporal design, and sampling schedule. This interface makes supervised fine-tuning off-policy, score-based distillation inapplicable, and direct adversarial imitation too sparse for denoising-time credit assignment. We propose Adversarial Flow Distillation (AFD), an on-policy framework for heterogeneous black-box video distillation. AFD queries the teacher and rolls out the current student on the same prompts, trains a prompt-paired Bradley-Terry discriminator to estimate clean-sample teacher-student discrepancy, and converts the resulting on-policy advantage into forward-process flow-matching updates on the student's own noised states. Thus, AFD provides dense velocity-field supervision while requiring no teacher scores, latents, denoising trajectories, step alignment, or reverse-chain reinforcement learning. Experiments across two causal AR student families show that AFD consistently improves motion- and physics-sensitive generation while preserving general video quality, and ablations validate the importance of adaptive on-policy feedback and forward-process credit assignment. The method requires only clean teacher videos and student rollouts, providing a practical route for distilling proprietary or heterogeneous video generators into efficient autoregressive students.

Yang Luo, Shengju Qian, Xiaohang Tang +4

National Univesity of Singapore · LIGHTSPEED · University College London

Autoregressive Video GenerationOn-Policy DistillationTeacher-Student LearningKnowledge DistillationAdversarial Knowledge Distillation
AskPDF
May 25, 2026

SAM3-Assisted Training of Lightweight YOLO Models for Precision Pig Farming

Deep learning-based object detection has revolutionized Precision Livestock Farming (PLF), yet a critical barrier remains: high-performance Foundation Models (such as SAM 3) are too computationally intensive for edge deployment, while lightweight models (like YOLO) require prohibitive manual annotation efforts. This work proposes a fully automated knowledge distillation pipeline that leverages the Segment Anything Model 3 (SAM 3) to generate zero-shot pseudo-labels for training efficient YOLOv8 detectors. By treating SAM 3 as an offline auto-annotator, we eliminate the manual labeling bottleneck, producing models capable of real-time inference on resource-constrained hardware. We systematically evaluate this approach on the PigLife dataset, comparing SAM 3-supervised models against human-annotated baselines. Results demonstrate that a SAM 3-trained YOLOv8m achieves a mean Average Precision (mAP) of 79.4% without human intervention, while reducing inference latency by approximately 200×\times compared to the teacher model. Furthermore, stratified analysis reveals that in low-occlusion scenarios, the automated pipeline achieves detection rates comparable to human benchmarks (AP50>99%AP_{50} > 99\%). These findings indicate that foundation models can serve as effective, zero-annotation-cost supervisors, enabling scalable edge computing solutions for smart agriculture.

Marcos Vinicius Mendes Faria, Thiago Borges Pereira, Isabella C. F. S. Condotta +2

Coordenadoria de Inform´atica, Instituto Federal do Esp´ırito Santo (IFES), Serra, ES · Department of Animal Sciences, University of Illinois at Urbana-Champaign, USA

Edge InferencePseudo-LabelingObject DetectionReal-Time Object DetectionKnowledge DistillationPrecision Livestock Farming
AskPDF
May 25, 2026

Extreme Region Policy Distillation

Reinforcement learning for large language models faces a fundamental trade-off between sample efficiency and asymptotic performance: strictly on-policy methods discard trajectories after a single update, while off-policy reuse introduces distribution mismatch that existing trust-region techniques mitigate primarily by enforcing conservative optimization, often leaving rich training signals underutilized. To investigate this, we perform extensive off-policy updates on fixed data. Our experiments reveal that aggressive multi-step optimization brings rapid initial gains, but excessive updates cause trajectory probabilities to deviate and entropy to collapse, with performance plateauing early. Tightening KL constraints merely lowers the ceiling without resolving the degradation. This motivates Extreme Region Policy Distillation (ERPD), a two-stage framework that decouples sample efficiency from KL efficiency. The first stage performs weakly constrained off-policy optimization on fixed data to maximally extract training signals. The resulting policy provides token-level supervision. In the second stage, we distill these signals into the base policy under trust-region constraints, filtering harmful drift while preserving useful signals. The distilled policy achieves comparable or better performance with substantially smaller KL divergence, indicating that much of the first-stage divergence was spent on unnecessary drift rather than genuine improvement. Crucially, ERPD accommodates both strong and weak teachers: when aggressive optimization yields no stronger policy, even degenerate teachers provide effective supervision via alternative signal construction strategies. We validate ERPD on mathematical reasoning, showing gains for strong base models where on-policy training plateaus, and reliable improvements with weak teachers.

Changyu Chen, Xiting Wang, Rui Yan

Gaoling School of Artificial Intelligence, Renmin University of China · Wuhan University

KL-Regularized RLRL for Language Model ReasoningLanguage Model DistillationKnowledge DistillationOff-Policy RL
AskPDF
May 24, 2026

When Does Adaptive Guidance Help? Belief-Aware Privileged Distillation for Autonomous Driving Under Partial Observability

Guided Soft Actor-Critic (GSAC) distills knowledge from a privileged full-state teacher to a partial-observation student for autonomous driving, but uses a fixed distillation coefficient lambda regardless of the agent's uncertainty. We present Belief-Aware GSAC (BA-GSAC), which modulates lambda via ensemble disagreement, and use it as a testbed for a systematic empirical study asking: when does adaptive guidance actually help? Evaluating five strategies (fixed lambda in {0.01, 0.1}, adaptive, linear decay, and vanilla SAC) across three POMDP difficulty levels on Highway-Env, we find that preliminary single-seed runs suggest benefits under mild and moderate partial observability, but under severe occlusion (evaluated with 3 seeds for all methods) the adaptive coefficient collapses to lambda_min within about 3K steps. We trace this to an observability blindness phenomenon: because the ensemble predicts partial observations, it achieves low disagreement even under heavy occlusion, modeling what is visible but unable to detect what is missing. We diagnose the root cause and propose an architectural fix (training the ensemble on full-state predictions using the guiding actor's privileged access); while not validated here, we show that even with current limitations, the warmup phase provides measurable stabilization (CV=13.3% vs. 29.8% for constant lambda=0.01). In fact, a simple deterministic linear decay schedule achieves the best severe-POMDP performance across all metrics (mean 116.5, CV=8.9%), suggesting that the scheduling effect, not the ensemble, drives the stability benefit. These findings provide practical guidance for designing uncertainty-aware teacher-student frameworks and highlight ensemble prediction targets as an important design choice.

Mehmet Haklidir

TUBITAK BILGEM Artificial Intelligence Institute, Turkey

Partially Observable RLAutonomous DrivingSoft Actor-CriticLearning Using Privileged InformationKnowledge Distillation
AskPDF
May 24, 2026

PANDO: Efficient Multimodal AI Agents via Online Skill Distillation

Recent advances in multimodal web agents often rely on increased inference-time computation, including rollout search, verifier passes, offline skill discovery, and specialist model stacks. This raises a central question: can a web agent become more efficient as it accumulates experience, rather than more expensive? We first analyze trajectories from VisualWebArena and identify three recurring sources of inefficiency: repeat-action loops, hidden discovery costs, and low prompt-cache reuse. We then introduce PANDO, a single-rollout online skill-distillation framework that maintains a structured Skill Library and combines progress reflection, confidence-based skill demotion, hierarchical routing, visual compression, and cache-aware prompting. On the full set of 910 VisualWebArena tasks, PANDO achieves a 58.3% success rate, outperforming SGV (54.0%) and our WALT reproduction (45.2%), while using 58% fewer tokens than SGV and 61% fewer tokens than WALT, without any pre-evaluation discovery budget. A 300-task ablation further shows that rules and routines provide most of the success gains, while routing, compression, and cache-aware prompting convert the larger skill library into lower marginal token cost. Finally, we introduce three trajectory-level efficiency metrics -- Action Repetition Rate, Step Overhead Ratio, and Prompt Cache Utilization -- to make efficiency visible beyond terminal success.

Yubo Li, Yidi Miao, Yuntian Shen +1

LLM Agent Workflow OptimizationLLM Agent Skill LearningLLM AgentsKnowledge DistillationWeb Agents
AskPDF
May 23, 2026

DisDop: Distillation with Domain Priors for Open-Vocabulary Aerial Object Detection

With the widespread application of drones in recent years, object detection of aerial images has attracted increasing attention, especially open-vocabulary aerial detection which is not restricted to predefined categories. Due to the scarcity of drone's viewpoint images and their significant differences from natural images, it is difficult to achieve satisfying results by directly applying vanilla open-vocabulary detection methods designed for natural scenarios. Some studies propose to transfer knowledge from pre-trained models by using lightweight networks or generating pseudo labels, but they tend to rely on models trained on natural images, neglecting the potential of foundation models specifically tailored for remote sensing and aerial imagery. To address this limitation, we propose DisDop, a unified framework that systematically distills multi-level domain priors from remote sensing foundation models (e.g., RemoteCLIP and DINOv3) into a lightweight detector. Specifically, we first distill visual priors through a teacher fusion strategy that combines RemoteCLIP's cross-modal alignment capability with DINOv3's fine-grained local feature extraction ability, transferring their complementary strengths to the detector's backbone. Second, we distill textual priors embedded in RemoteCLIP's text encoder by explicitly modeling inter-category semantic relationships, while incorporating global contextual priors to enhance local feature representation for small objects. Through this multi-level prior distillation framework, our DisDop achieves new state-of-the-art performance on open-vocabulary aerial detection benchmarks. Extensive ablation analysis also demonstrates the rationality and effectiveness of our proposed modules.

Ruihao Xu, Yong Liu, Yansong Tang +6

1Tsinghua Shenzhen International Graduate School, Tsinghua University · 2Tsinghua University

Open-Vocabulary Object DetectionKnowledge DistillationAerial Object Detection
AskPDF
May 22, 2026

Strong Teacher Not Needed? On Distillation in LLM Pretraining

Knowledge distillation generally assumes a strong-to-weak relationship where stronger teachers yield better students. In this work, we examine this assumption about distillation in large language model pretraining. By varying architecture sizes and training token budgets, we create strong-to-weak, same-level, and weak-to-strong teacher-student relationships, and study distillation's effectiveness under each. We find that the teacher need not be strong: with proper mixing of the language modeling and knowledge distillation losses, even small and undertrained teachers improve larger students. At the same time, a stronger teacher is not always better: pushing the teacher further, through more parameters or more training tokens, can saturate or even reverse the distillation gains. We further observe that distillation improves generalization (out-of-distribution and downstream performance) more readily than in-domain fitting. Together, these results challenge the common belief that distillation pretraining always requires a strong teacher.

Taiming Lu, Zhuang Liu

Princeton University

Language Model PretrainingLanguage Model DistillationTeacher-Student LearningKnowledge Distillation
AskPDF
May 22, 2026

Learning Through Noise: Why Subliminal Learning Works and When It Fails

In the context of artificial neural networks, subliminal learning refers to the transfer of task-relevant knowledge or unintended biases from teacher to student models through distillation on task-unrelated input\unicodex2013\unicode{x2013}output pairs. Prior explanations tie this effect to shared or closely matched teacher\unicodex2013\unicode{x2013}student initialization. We show that a closely matched initialization is not necessary. Instead, subliminal learning is governed by compatible output heads. Using a controlled MNIST setting, we split outputs into an auxiliary head (for auxiliary, task-unrelated noise signals) and a class head (for classification) to demonstrate subliminal learning occurs\unicodex2014\unicode{x2014}even when we randomly initialize hidden layers and remove layers, add new layers, or change the architecture (MLP-to-CNN). Compatible auxiliary heads enable transfer of a recoverable teacher signal, bringing the student's representations closer to the teacher's. When the class heads remain compatible as well, students trained only on task-unrelated noise can approach, and in favorable regimes match, teacher-level task performance. Our setting enables us to develop a theory that explains the mechanism of subliminal learning and to derive upper bounds on when subliminal learning fails. Together, our results turn subliminal learning from a surprising transfer effect into a theoretically grounded mechanism with predictable limits.

Vincent C. Brockers, Roman D. Ventzke, Valentin Neuhaus +2

Max Planck Institute for Dynamics and Self-Organization, Göttingen · Faculty of Physics, Institute for the Dynamics of Complex Systems, University of Göttingen

Subliminal LearningCross-Architecture Knowledge DistillationTeacher-Student LearningKnowledge DistillationRepresentation Alignment
AskPDF
May 22, 2026

HARNESS-LM: A Three-Phase Training Recipe for Harnessing SLMs in Sponsored Search Retrieval

In the competitive landscape of sponsored search, balancing retrieval quality with production latency is a critical challenge. While large retrieval models based on Small Language Models (SLMs) such as Qwen3-Embedding-4B/8B set strong upper bounds on public benchmarks, their deployment in high-throughput, latency-sensitive environments remains impractical. In this paper, we present HARNESS-LM (HLM), a three-phase training framework for transferring the capabilities of large-scale retrievers into compact, cost-efficient models. The approach comprises: (1) training a high-performance reference ("teacher") retriever by fine-tuning a billion-parameter-scale SLM; (2) aligning query representations via an L2 objective to distill knowledge into a sub-600M parameter student encoder; and (3) applying a final contrastive refinement stage to optimize the student for retrieval performance. We also present a comprehensive empirical study of key design choices, including alignment objectives, embedding dimensionality, model scale, architecture, and optimization strategies, to identify configurations that are most effective in production settings. On a real-world Bing Ads evaluation benchmark, HLM recovers over 98% of the reference retriever's precision across multiple settings, while delivering up to 27x lower online query-encoder latency and 20x higher throughput on NVIDIA A100 GPUs. Online A/B testing on Bing Ads further shows a +1% Revenue, +0.6% Impression, and +0.4% Click uplift over the current ensemble of retrievers running in production with the deployed 190M parameter model, clearly highlighting the practical efficacy of the HLM recipe in a real-world sponsored search setting.

Vipul Gupta, Shikhar Mohan, Lakshya Kumar +4

Microsoft AI India

Online AdvertisingSmall Language ModelsInformation RetrievalEfficient Language Model InferenceKnowledge DistillationDense Retrieval
AskPDF
May 22, 2026

One-Forcing: Towards Stable One-Step Autoregressive Video Generation

Recent advances in autoregressive diffusion-based video generation have substantially improved the quality of real-time video synthesis. However, most existing methods still require multiple denoising steps, while reducing sampling to a single step often leads to severe quality degradation: trajectory-based consistency distillation methods often produce videos with weak dynamics, whereas DMD-based methods, such as Self-Forcing, tend to generate blurry frames. We attribute this limitation to the teacher trajectories exhibiting highly concentrated curvature near the high-noise endpoint, which poses a fundamental geometric challenge for one-step distillation under the consistency distillation framework. To address these limitations, we propose One-Forcing, a simple yet effective approach that augments the DMD objective with an auxiliary GAN loss for high-quality and efficient one-step video generation. We further find that framewise autoregression stabilizes adversarial training, enabling higher-quality generation with substantially fewer training iterations than chunkwise autoregression. Experiments on VBench show that One-Forcing achieves a total score of 83.76, establishing state-of-the-art performance among one-step causal video generation methods while remaining competitive with strong multi-step approaches.

Jiaqi Feng, Justin Cui, Yuanhao Ban +1

Tsinghua University · UCLA

Autoregressive Video GenerationOne-Step Generative ModelingKnowledge DistillationAdversarial Knowledge DistillationVideo Diffusion ModelsSingle-Step Video Generation+1
AskPDF
May 22, 2026

TaxDistill: Improving Metagenomic Taxonomic Annotation via Distilled Genomic Foundation Models

Metagenomic taxonomic annotation is essential for interpreting complex microbial communities, yet reliable annotation remains challenging under reference database incompleteness and ambiguous taxonomic boundaries. Existing similarity-based tools are efficient, but they often produce noisy pseudo-labels in complex environments; learning-based post-hoc correction methods can further inherit this noise when trained on hard pseudo-labels generated by upstream classifiers. We propose TaxDistill, a plug-and-play knowledge distillation framework for reliable metagenomic taxonomic recalibration. TaxDistill uses the genomic foundation model GenomeOcean as a semantic teacher and performs one-way online distillation, where the teacher is updated by hierarchical taxonomic supervision while its detached soft probability distributions guide a lightweight student network. This design mitigates the student's overfitting to noisy hard pseudo-labels and supports confidence based rejection of potentially unreliable predictions. Comprehensive experiments on seven diverse CAMI2 datasets with MMseqs2, Metabuli, and Kraken2 show that TaxDistill improves annotation reliability over existing baselines in most scenarios. For example, on the Gastrointestinal dataset, TaxDistill improves the F1 score of MMseqs2 from 0.763 to 0.941, surpassing the Taxometer baseline. Our results suggest that large genomic foundation models can provide semantic supervision for noisy scientific annotations, enabling lightweight post-hoc recalibration of existing metagenomic annotation results.

Rongye Ye, Lun Li, Zheng Luo +3

National Genomics Data Center, China National Center for Bioinformation, Beijing 100101, China · Beijing Institute of Genomics, Chinese Academy of Sciences, Beijing 100101, China · University of Chinese Academy of Sciences, Beijing 100049, China +1

Learning with Noisy LabelsKnowledge Distillation
AskPDF
May 21, 2026

Safeguarding Text-to-Image Generative Models Against Unauthorized Knowledge Distillation

Closed-weight generative services are increasingly deployed through query-based APIs, where users can obtain generated outputs while model parameters remain inaccessible. However, such deployment does not prevent model stealing: an attacker can repeatedly query the service, collect large volumes of released synthetic images, and use them as training data for a private substitute model. This query-output-driven process enables unauthorized knowledge distillation and capability replication without direct access to the original weights. To mitigate this threat, a practical defense should preserve the visual fidelity of released images, provide explicit control over perturbation magnitude, and scale efficiently to large-volume output release. We present WaveGuard, a single-pass, generator-based protection framework that safeguards released synthetic images under a user-specified perturbation budget. WaveGuard employs a frequency-aware perturbation generator to inject structured, imperceptible perturbations that maintain perceptual utility for benign viewers while reducing the usefulness of protected images as training data for unauthorized student models. Extensive experiments under WikiArt-related synthetic-output distillation settings show that WaveGuard achieves a favorable efficacy--fidelity--efficiency trade-off, with explicit imperceptibility control and substantial gains in protection efficiency.

Yilan Gao, Sida Huang, Hongyuan Zhang +1

School of Artificial Intelligence, OPtics and ElectroNics (iOPEN), Northwestern Polytechnical University, Xi’an 710072, P. R. China · Institute of Artificial Intelligence (TeleAI), China Telecom, P. R. China · The University of Hong Kong

Model Extraction AttacksT2I GenerationAdversarial PerturbationsKnowledge Distillation
AskPDF
May 21, 2026

Toward Understanding Adversarial Distillation: Why Robust Teachers Fail

Adversarial Distillation aims to enhance student robustness by guiding the student with a robust teacher's soft labels within the min-max adversarial training framework, yet its success is notoriously inconsistent: a more robust teacher often fails to improve, or even harms, the student's robust generalization. In this paper, we identify a key mechanism of this teacher dependency: the misalignment between the teacher's supervisory confidence and the student's representational limitations on a consistent subset of training data -- the Robustly Unlearnable Set. We present a theoretical framework analyzing the feature learning dynamics of a two-layer neural network, demonstrating that this mismatch creates a dichotomy in distillation outcomes. We prove that when a teacher provides confident supervision on unlearnable samples, it compels the student to memorize spurious noise patterns that eventually overpower the learned robust signal, thereby driving robust overfitting. Conversely, a teacher that exhibits high uncertainty on these samples effectively suppresses noise memorization, allowing the student to rely solely on the learnable signal for robust generalization. We empirically validate our theory across both synthetic simulations and real-image classification datasets, confirming that robust overfitting is driven by the teacher's interaction with unlearnable samples. Finally, we demonstrate that a teacher's predictive entropy on unlearnable samples serves as a strong indicator of student robustness, validating our theoretical framework and offering a principled guideline for robust teacher selection.

Hongsin Lee, Hye Won Chung

School of Electrical Engineering, KAIST, Daejeon, Korea.

Adversarial TrainingNeural Network RobustnessAdversarial RobustnessAdversarial ExamplesTeacher-Student LearningKnowledge Distillation+2
AskPDF
May 21, 2026

Visual-Advantage On-Policy Distillation for Vision-Language Models

On-policy knowledge distillation has proven effective for language models, yet its application to vision-language models (VLMs) remains underexplored. We observe that standard on-policy distillation can improve a student's output quality while failing to strengthen its reliance on visual input: on vision-critical tokens, the student's predictions remain largely unchanged whether or not fine-grained visual detail is present, even though the teacher's predictions depend heavily on it.To make this difference observable, we introduce visual advantage (VA), the token-level log-probability difference when the teacher scores a student-generated rollout with versus without access to fine-grained visual detail. VA is concentrated in a small minority of tokens, and these high-VA tokens are the ones that actually carry the visual supervision signal. This motivates a distillation objective that treats them differently from language scaffolding, so their contribution is not diluted by the abundant surrounding language tokens.We propose Visual-Advantage On-Policy Distillation (VA-OPD), which uses VA at two granularities: rollout-level reweighting by trajectory-averaged VA, and token-level KL averaged within high-VA and low-VA groups separately. We train on two math datasets (Geometry3K and ViRL39K) and evaluate on eight benchmarks covering both mathematical reasoning and visual understanding, across three teacher sizes (4B, 8B, and 32B) on the Qwen3-VL family. VA-OPD improves over standard on-policy distillation on every benchmark, with the gain growing monotonically along both the teacher-size and data-scale axes, suggesting that these factors compound consistently.

Ruiqi Liu, Xiaolei Lv, Gengsheng Li +8

Institute of Automation, CAS · School of Advanced Interdisciplinary Sciences, UCAS · 3Hello +1

Vision-Language ModelsVLM DistillationVLM ReasoningVisually Grounded ReasoningOn-Policy DistillationTeacher-Student Learning+1
AskPDF
May 20, 2026

X-Token: Projection-Guided Cross-Tokenizer Knowledge Distillation

Cross-tokenizer knowledge distillation allows a student model to learn from teachers with incompatible vocabularies. Prior work operates on hidden states or logits; the latter is preferred as a drop-in replacement requiring no auxiliary components. Logit-based methods either use only the correct-token probability, missing the full 'dark knowledge' in the teacher's distribution, or operate on the full output distribution, relying on strict token partitioning and/or unprincipled heuristic ranking. We identify two key shortcomings of full-distribution, logit-based methods: (i) an uncommon-token failure, where critical tokens fall into the unmatched subset (e.g., Llama's 1100 multi-digit numerals under digit-splitting Qwen supervision) and are suppressed during training, reducing GSM8k from 12.89 to 2.56 compared to same-tokenizer KD from a weaker teacher; and (ii) over-conservative matching, where strict 1-to-1 matching excludes near-equivalent tokens across surface forms. These failures require distinct remedies: eliminating the partition when critical tokens are misaligned, and refining it when alignment is reliable. We propose X-Token, an approach with two complementary loss formulations targeting these issues. P-KL removes partitioning and aligns the student's distribution with the teacher's via a sparse projection matrix W (initialized from tokenizer-level string rules) to address the uncommon-token failure. H-KL retains the hybrid form while relaxing matching to align each student token with its top-ranked teacher mapping under W. Both objectives share W and extend naturally to multiple teachers. Empirically, on Llama-3.2-1B, X-Token outperforms the current state of the art GOLD by +3.82 average points with a Qwen3-4B teacher and by +0.5 with a Phi-4-Mini teacher. Further, a two-teacher setup (Phi-4-mini + Llama-3B) improves over single-teacher distillation by +1.3 points.

Sharath Turuvekere Sreenivas, Adithyakrishna Venkatesh Hanasoge, Mingyu Yang +4

Cross-Tokenizer Knowledge DistillationLanguage Model DistillationTeacher-Student LearningKnowledge Distillation
AskPDF
May 20, 2026

When Are Teacher Tokens Reliable? Position-Weighted On-Policy Self-Distillation for Reasoning

On-policy self-distillation (OPSD) trains a student on its own rollouts using a privileged teacher, but its standard objective weights all generated tokens equally, implicitly treating the privileged teacher target as equally reliable at every student-visited prefix. Existing entropy-based OPD methods relax this uniformity by modulating token-level supervision with teacher entropy, but high teacher entropy in reasoning has an ambiguous reliability meaning: it can reflect either non-viable uncertainty or benign solution diversity. To identify this phenomenon, we introduce a branch-viability diagnostic. Specifically, we record next-token alternatives from the privileged-answer teacher prompt, force each alternative after the student prompt plus its on-policy spine prefix, and test whether the resulting student-template continuation recovers the correct answer. On Qwen3-4B, we find that an oriented within-sequence position score is the strongest tested predictor of teacher-token reliability, reaching an area-under-ROC-curve (AUROC) of 0.83; local uncertainty scores are at most 0.57. Motivated by this trajectory-level structure, we propose Position-Weighted On-Policy Self-Distillation (PW-OPSD), which applies an increasing position weight while keeping the same student rollout, privileged teacher pass, and clipped forward-KL target as OPSD. In our comprehensive evaluations with different random seeds, the diagnostic-derived PW-OPSD improves AIME 2024 and AIME 2025 Avg@12 by +1.0 and +1.1 points, and a generalization evaluation on two larger-scale models from different families, DeepSeek-R1-Distill-Llama-8B and Olmo-3-7B-Think, also demonstrates consistent aggregate Avg@12 improvements. These results show that teacher-token reliability in reasoning distillation is trajectory-structured and can be utilized without additional teacher computation.

Xiaogeng Liu, Xinyan Wang, Yingzi Ma +2

1Johns Hopkins University · University of Wisconsin–Madison · 3Nanyang Technological University

CoT DistillationOn-Policy Self-DistillationMathematical ReasoningLanguage Model DistillationTeacher-Student LearningKnowledge Distillation
AskPDF