On-policy distillation (OPD), which supervises a student on its own sampled trajectories, has emerged as a data-efficient post-training method for improving reasoning while avoiding the reward dependence of reinforcement learning and the catastrophic forgetting often observed in standard supervised fine-tuning. However, standard OPD typically computes teacher supervision under noisy student-generated contexts and often relies on a single stochastic teacher rollout per prompt. As a result, the supervision signal can be high-variance: the sampled teacher trajectory can be incorrect, uninformative, or poorly matched to the student's current reasoning behavior. To address this limitation, we propose BRTS, a Best-of-N Rollout Teacher Selection framework for on-policy distillation. BRTS augments standard student-context OPD with a teacher-context supervision branch constructed from the curated teacher trajectory. Rather than distilling from the first sampled teacher rollout, BRTS samples a small pool of teacher trajectories and selects the auxiliary trajectory using a simple priority rule: correctness first, student alignment second. When multiple correct teacher trajectories are available, BRTS chooses the one most aligned with the student's current behavior; when unconditioned teacher samples fail on harder prompts, it invokes a ground-truth-conditioned recovery step to elicit a natural derivation. The selected trajectory is then used to provide reliable teacher-context supervision inside the OPD loop, augmented with an auxiliary loss on the teacher trajectory. Experiments on AIME 2024, AIME 2025, and AMC 2023 show that BRTS improves over standard OPD on challenging reasoning benchmarks, with the largest gains on harder datasets. Our code is available at https://github.com/BWGZK-keke/BRTS.
Knowledge distillation~(KD) is an effective learning paradigm for improving the performance of lightweight student networks by utilizing additional supervision knowledge distilled from teacher networks. Most pioneering studies either learn from only a single teacher in their distillation learning methods, neglecting the potential that a student can learn from multiple teachers simultaneously, or simply treat each teacher to be equally important, unable to reveal the different importance of teachers for specific examples. To bridge this gap, we propose a novel adaptive multi-teacher multi-level…
SenseTime Research（商汤科技研究院）；CUHK-SenseTime Joint Laboratory, The Chinese University of Hong Kong（香港中文大学-商汤科技联合实验室，香港中文大学）
Training a small student network with the guidance of a larger teacher network is an effective way to promote the performance of the student. Despite the different types, the guided knowledge used to distill is always kept unchanged for different teacher and student pairs in previous knowledge distillation methods. However, we find that teacher and student models with different networks or trained from different initialization could have distinct feature representations among different channels. (e.g. the high activated channel for different categories). We name this incongruous representation…
Knowledge distillation which learns a lightweight student model by distilling knowledge from a cumbersome teacher model is an attractive approach for learning compact deep neural networks (DNNs). Recent works further improve student network performance by leveraging multiple teacher networks. However, most of the existing knowledge distillation-based multi-teacher methods use separately pretrained teachers. This limits the collaborative learning between teachers and the mutual learning between teachers and student. Network quantization is another attractive approach for learning compact DNNs. …
With the success of deep neural networks, knowledge distillation which guides the learning of a small student network from a large teacher network is being actively studied for model compression and transfer learning. However, few studies have been performed to resolve the poor learning issue of the student network when the student and teacher model sizes significantly differ. In this paper, we propose a densely guided knowledge distillation using multiple teacher assistants that gradually decreases the model size to efficiently bridge the large gap between the teacher and student networks. To…
School of Computer Science, The University of Sydney（悉尼大学计算机科学学院）
Recent years have witnessed dramatically improvements in the knowledge distillation, which can generate a compact student model for better efficiency while retaining the model effectiveness of the teacher model. Previous studies find that: more accurate teachers do not necessary make for better teachers due to the mismatch of abilities. In this paper, we aim to analysis the phenomenon from the perspective of model calibration. We found that the larger teacher model may be too over-confident, thus the student model cannot effectively imitate. While, after the simple model calibration of the tea…
Teacher-Guided Student Self-Knowledge Distillation Using Diffusion Model
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences（人工智能安全国家重点实验室，计算技术研究所，中国科学院）；University of Chinese Academy of Sciences（中国科学院大学）
Simon Fraser University（西蒙弗雷泽大学）；Borealis AI（Borealis人工智能研究院）
Knowledge distillation is a widely applicable technique for training a student neural network under the guidance of a trained teacher network. For example, in neural network compression, a high-capacity teacher is distilled to train a compact student; in privileged learning, a teacher trained with privileged data is distilled to train a student without access to that data. The distillation loss determines how a teacher's knowledge is captured and transferred to the student. In this paper, we propose a new form of knowledge distillation loss that is inspired by the observation that semantically…
There is a need for an on-the-fly computational process with very low performance system such as system-on-chip (SoC) and embedded device etc. This paper presents pacemaker knowledge distillation as intermediate ensemble teacher to use convolutional neural network in these systems. For on-the-fly system, we consider student model using 1xN shape on-the-fly filter and teacher model using normal NxN shape filter. We note three points about training student model, caused by applying on-the-fly filter. First, same depth but unavoidable thin model compression. Second, the large capacity gap and par…
Zhejiang University（浙江大学）；Shanghai Institute for Advanced Study of Zhejiang University（浙江大学上海高等研究院）；ZJU-Bangsun Joint Research Center（浙江大学-邦盛联合研究中心）；Zhejiang University of Technology（浙江工业大学）
Knowledge distillation aims to compress a powerful yet cumbersome teacher model into a lightweight student model without much sacrifice of performance. For this purpose, various approaches have been proposed over the past few years, generally with elaborately designed knowledge representations, which in turn increase the difficulty of model development and interpretation. In contrast, we empirically show that a simple knowledge distillation technique is enough to significantly narrow down the teacher-student performance gap. We directly reuse the discriminative classifier from the pre-trained …
The University of Sydney（悉尼大学）；SenseTime Research（商汤科技研究院）；Shanghai AI Lab（上海人工智能实验室）
Knowledge distillation (KD) has shown very promising capabilities in transferring learning representations from large models (teachers) to small models (students). However, as the capacity gap between students and teachers becomes larger, existing KD methods fail to achieve better results. Our work shows that the `prior knowledge' is vital to KD, especially when applying large teachers. Particularly, we propose the dynamic prior knowledge (DPK), which integrates part of teacher's features as the prior knowledge before the feature distillation. This means that our method also takes the teacher'…
Knowledge distillation usually transfers the knowledge from a pre-trained cumbersome teacher network to a compact student network, which follows the classical teacher-teaching-student paradigm. Based on this paradigm, previous methods mostly focus on how to efficiently train a better student network for deployment. Different from the existing practices, in this paper, we propose a novel student-helping-teacher formula, Teacher Evolution via Self-Knowledge Distillation (TESKD), where the target teacher (for deployment) is learned with the help of multiple hierarchical students by sharing the st…
In this paper, we present a thorough evaluation of the efficacy of knowledge distillation and its dependence on student and teacher architectures. Starting with the observation that more accurate teachers often don't make good teachers, we attempt to tease apart the factors that affect knowledge distillation performance. We find crucially that larger models do not often make better teachers. We show that this is a consequence of mismatched capacity, and that small students are unable to mimic large teachers. We find typical ways of circumventing this (such as performing a sequence of knowledge…
The outpouring of various pre-trained models empowers knowledge distillation by providing abundant teacher resources, but there lacks a developed mechanism to utilize these teachers adequately. With a massive model repository composed of teachers pre-trained on diverse tasks, we must surmount two obstacles when using knowledge distillation to learn a new task. First, given a fixed computing budget, it is not affordable to try each teacher and train the student repeatedly, making it necessary to seek out the most contributive teacher precisely and efficiently. Second, semantic gaps exist betwee…
Knowledge Distillation (KD) is an effective framework for compressing deep learning models, realized by a student-teacher paradigm requiring small student networks to mimic the soft target generated by well-trained teachers. However, the teachers are commonly assumed to be complex and need to be trained on the same datasets as students. This leads to a time-consuming training process. The recent study shows vanilla KD plays a similar role as label smoothing and develops teacher-free KD, being efficient and mitigating the issue of learning from heavy teachers. But because teacher-free KD relies…
The performance of a distillation-based compressed network is governed by the quality of distillation. The reason for the suboptimal distillation of a large network (teacher) to a smaller network (student) is largely attributed to the gap in the learning capacities of given teacher-student pair. While it is hard to distill all the knowledge of a teacher, the quality of distillation can be controlled to a large extent to achieve better performance. Our experiments show that the quality of distillation is largely governed by the quality of teacher's response, which in turn is heavily affected by…
Peking University（北京大学）；Megvii (Face++) Technology Inc（旷视科技（Face++））；National Engineering Laboratory for Big Data Analysis and Applications（大数据分析与应用技术国家工程实验室）；DiDi AI Labs（滴滴人工智能实验室）
Knowledge distillation has become increasingly important in model compression. It boosts the performance of a miniaturized student network with the supervision of the output distribution and feature maps from a sophisticated teacher network. Some recent works introduce multi-teacher distillation to provide more supervision to the student network. However, the effectiveness of multi-teacher distillation methods are accompanied by costly computation resources. To tackle with both the efficiency and the effectiveness of knowledge distillation, we introduce the feature aggregation to imitate the m…
Armed Forces Yangju Hospital（杨州武装部队医院）；Seoul National University（首尔国立大学）
For more clinical applications of deep learning models for medical image segmentation, high demands on labeled data and computational resources must be addressed. This study proposes a coarse-to-fine framework with two teacher models and a student model that combines knowledge distillation and cross teaching, a consistency regularization based on pseudo-labels, for efficient semi-supervised learning. The proposed method is demonstrated on the abdominal multi-organ segmentation task in CT images under the MICCAI FLARE 2022 challenge, with mean Dice scores of 0.8429 and 0.8520 in the validation …
Knowledge Distillation is a technique which aims to utilize dark knowledge to compress and transfer information from a vast, well-trained neural network (teacher model) to a smaller, less capable neural network (student model) with improved inference efficiency. This approach of distilling knowledge has gained popularity as a result of the prohibitively complicated nature of such cumbersome models for deployment on edge computing devices. Generally, the teacher models used to teach smaller student models are cumbersome in nature and expensive to train. To eliminate the necessity for a cumberso…
Knowledge distillation is a standard teacher-student learning framework to train a light-weight student network under the guidance of a well-trained large teacher network. As an effective teaching strategy, interactive teaching has been widely employed at school to motivate students, in which teachers not only provide knowledge but also give constructive feedback to students upon their responses, to improve their learning performance. In this work, we propose an InterActive Knowledge Distillation (IAKD) scheme to leverage the interactive teaching strategy for efficient knowledge distillation. …
This paper proposes a novel knowledge distillation-based learning method to improve the classification performance of convolutional neural networks (CNNs) without a pre-trained teacher network, called exit-ensemble distillation. Our method exploits the multi-exit architecture that adds auxiliary classifiers (called exits) in the middle of a conventional CNN, through which early inference results can be obtained. The idea of our method is to train the network using the ensemble of the exits as the distillation target, which greatly improves the classification performance of the overall network.…
Traditional knowledge distillation transfers "dark knowledge" of a pre-trained teacher network to a student network, and ignores the knowledge in the training process of the teacher, which we call teacher's experience. However, in realistic educational scenarios, learning experience is often more important than learning results. In this work, we propose a novel knowledge distillation method by integrating the teacher's experience for knowledge transfer, named experience ensemble knowledge distillation (EEKD). We save a moderate number of intermediate models from the training process of the tea…
The knowledge of a well-trained deep neural network (a.k.a. the "teacher") is valuable for learning similar tasks. Knowledge distillation extracts knowledge from the teacher and integrates it with the target model (a.k.a. the "student"), which expands the student's knowledge and improves its learning efficacy. Instead of enforcing the teacher to work on the same task as the student, we borrow the knowledge from a teacher trained from a general label space -- in this "Generalized Knowledge Distillation (GKD)", the classes of the teacher and the student may be the same, completely different, or …
Although deep convolution neural networks (DCNN) have achieved excellent performance in human pose estimation, these networks often have a large number of parameters and computations, leading to the slow inference speed. For this issue, an effective solution is knowledge distillation, which transfers knowledge from a large pre-trained network (teacher) to a small network (student). However, there are some defects in the existing approaches: (I) Only a single teacher is adopted, neglecting the potential that a student can learn from multiple teachers. (II) The human segmentation mask can be reg…
Beijing University of Posts and Telecommunications（北京邮电大学）；ByteDance（字节跳动）；USTC（中国科学技术大学）；Beijing Key Laboratory of Network System and Network Culture（北京网络系统与网络文化重点实验室）；Key Laboratory of Interactive Technology and Experience System, Ministry of Culture and Tourism（文化和旅游部互动技术与体验系统重点实验室）；Zhongguancun Academy（中关村科学城）
Graduate School of Convergence Science and Technology（融合科学与技术研究生院）；Seoul National University（首尔国立大学）
Knowledge Distillation (KD) aims to transfer knowledge in a teacher-student framework, by providing the predictions of the teacher network to the student network in the training stage to help the student network generalize better. It can use either a teacher with high capacity or {an} ensemble of multiple teachers. However, the latter is not convenient when one wants to use feature-map-based distillation methods. For a solution, this paper proposes a versatile and powerful training algorithm named FEature-level Ensemble for knowledge Distillation (FEED), which aims to transfer the ensemble kno…
The crux of knowledge distillation is to effectively train a resource-limited student model with the guide of a pre-trained larger teacher model. However, when there is a large difference between the model complexities of teacher and student (i.e., capacity gap), knowledge distillation loses its strength in transferring knowledge from the teacher to the student, thus training a weaker student. To mitigate the impact of the capacity gap, we introduce knowledge distillation via intermediate heads. By extending the intermediate layers of the teacher (at various depths) with classifier heads, we c…
HorizonDrive: Self-Corrective Autoregressive World Model for Long-horizon Driving Simulation
Wuhan University（武汉大学）；The University of Tokyo（东京大学）；Horizon Robotics；Tsinghua University（清华大学）；University of Science and Technology of China（中国科学技术大学）；The Chinese University of Hong Kong（香港中文大学）
Knowledge Distillation (KD) has been used in image classification for model compression. However, rare studies apply this technology on single-stage object detectors. Focal loss shows that the accumulated errors of easily-classified samples dominate the overall loss in the training process. This problem is also encountered when applying KD in the detection task. For KD, the teacher-defined hard samples are far more important than any others. We propose ADL to address this issue by adaptively mimicking the teacher's logits, with more attention paid on two types of hard samples: hard-to-learn sa…
