Chengkai Huang

Yuhang Yao, Jianyi Zhang, Junda Wu et al.

Large language models are rapidly gaining popularity and have been widely adopted in real-world applications. While the quality of training data is essential, privacy concerns arise during data collection. Federated learning offers a solution by allowing multiple clients to collaboratively train LLMs without sharing local data. However, FL introduces new challenges, such as model convergence issues due to heterogeneous data and high communication costs. A comprehensive study is required to address these challenges and guide future research. This paper surveys Federated learning for LLMs (FedLLM), highlighting recent advances and future directions. We focus on two key aspects: fine-tuning and prompt learning in a federated setting, discussing existing work and associated research challenges. We finally propose potential directions for federated LLMs, including pre-training, federated agents, and LLMs for federated learning.

Juexiang Ye, Xue Li, Xinyu Yang et al.

Large language model-based agents operating in long-horizon interactions require memory systems that support temporal consistency, multi-hop reasoning, and evidence-grounded reuse across sessions. Existing approaches largely rely on unstructured retrieval or coarse abstractions, which often lead to temporal conflicts, brittle reasoning, and limited traceability. We propose MemWeaver, a unified memory framework that consolidates long-term agent experiences into three interconnected components: a temporally grounded graph memory for structured relational reasoning, an experience memory that abstracts recurring interaction patterns from repeated observations, and a passage memory that preserves original textual evidence. MemWeaver employs a dual-channel retrieval strategy that jointly retrieves structured knowledge and supporting evidence to construct compact yet information-dense contexts for reasoning. Experiments on the LoCoMo benchmark demonstrate that MemWeaver substantially improves multi-hop and temporal reasoning accuracy while reducing input context length by over 95\% compared to long-context baselines.

Chuhan Wang, Xintong Li, Jennifer Yuntong Zhang et al.

Multimodal large language models often struggle with faithful reasoning in complex visual scenes, where intricate entities and relations require precise visual grounding at each step. This reasoning unfaithfulness frequently manifests as hallucinated entities, mis-grounded relations, skipped steps, and over-specified reasoning. Existing preference-based approaches, typically relying on textual perturbations or answer-conditioned rationales, fail to address this challenge as they allow models to exploit language priors to bypass visual grounding. To address this, we propose SceneAlign, a framework that leverages scene graphs as structured visual information to perform controllable structural interventions. By identifying reasoning-critical nodes and perturbing them through four targeted strategies that mimic typical grounding failures, SceneAlign constructs hard negative rationales that remain linguistically plausible but are grounded in inaccurate visual facts. These contrastive pairs are used in Direct Preference Optimization to steer models toward fine-grained, structure-faithful reasoning. Across seven visual reasoning benchmarks, SceneAlign consistently improves answer accuracy and reasoning faithfulness, highlighting the effectiveness of grounding-aware alignment for multimodal reasoning.

Ruhan Wang, Chengkai Huang, Zhiyong Wang et al.

Large language models (LLMs) exhibit strong reasoning capabilities when guided by high-quality demonstrations, yet such data is often distributed across organizations that cannot centralize it due to regulatory, proprietary, or institutional constraints. We study federated reasoning, where a server improves multi-step reasoning by coordinating with heterogeneous clients holding private demonstrations, without centralized training or raw data sharing. The key challenge is that client reliability is query-dependent, while the server cannot inspect client data to determine which contributions are trustworthy. To address this, we propose Uncertainty-Aware Federated Reasoning (FERA), a training-free framework based on iterative server-client co-refinement. Across communication rounds, clients generate reasoning traces with lightweight uncertainty estimates, and the server synthesizes them into improved reasoning that is redistributed as context for the next round, progressively improving both server outputs and client-side reasoning. Within each round, Uncertainty-Aware Self-Critique Aggregation (UA-SCA) resolves conflicts among heterogeneous client traces through query-dependent trust weighting and structured cross-client verification. Rather than simply discarding low-quality traces, UA-SCA revises flawed reasoning steps to recover useful information. We provide theoretical guarantees showing that the proposed iterative protocol converges and that uncertainty-aware weighting accelerates convergence. Experiments on multiple reasoning benchmarks show that FERA consistently outperforms both federated training and training-free baselines, achieving progressively higher accuracy across rounds while maintaining communication and computational efficiency.

Chengkai Huang, Shoujin Wang, Xianzhi Wang et al.

Sequential recommender systems (SRSs) aim to predict the subsequent items which may interest users via comprehensively modeling users' complex preference embedded in the sequence of user-item interactions. However, most of existing SRSs often model users' single low-level preference based on item ID information while ignoring the high-level preference revealed by item attribute information, such as item category. Furthermore, they often utilize limited sequence context information to predict the next item while overlooking richer inter-item semantic relations. To this end, in this paper, we proposed a novel hierarchical preference modeling framework to substantially model the complex low- and high-level preference dynamics for accurate sequential recommendation. Specifically, in the framework, a novel dual-transformer module and a novel dual contrastive learning scheme have been designed to discriminatively learn users' low- and high-level preference and to effectively enhance both low- and high-level preference learning respectively. In addition, a novel semantics-enhanced context embedding module has been devised to generate more informative context embedding for further improving the recommendation performance. Extensive experiments on six real-world datasets have demonstrated both the superiority of our proposed method over the state-of-the-art ones and the rationality of our design.

Xihang Wang, Zihan Wang, Chengkai Huang et al.

Multimodal Retrieval-Augmented Generation (MRAG) is widely adopted for Multimodal Large Language Models (MLLMs) with external evidence to reduce hallucinations. Despite its success, most existing MRAG frameworks treat retrieved evidence as indivisible documents, implicitly assuming that all content within a document is equally informative. In practice, however, sometimes only a small fraction of a document is relevant to a given query, while the remaining content introduces substantial noise that may lead to performance degradation. We address this fundamental limitation by reframing MRAG as a fine-grained evidence selection problem. We propose Fragment-level Evidence Selection for RAG (FES-RAG), a framework that selects atomic multimodal fragments rather than entire documents as grounding evidence. FES-RAG decomposes retrieved multimodal documents into sentence-level textual fragments and region-level visual fragments, enabling precise identification of evidence that directly supports generation. To guide fragment selection, we introduce Fragment Information Gain (FIG), a principled metric that measures the marginal contribution of each fragment to the MLLM's generation confidence. Based on FIG, we distill fragment-level utility judgments from a high-capacity MLLM into a lightweight selector, achieving accurate evidence selection with low inference overhead. Experiments on the M2RAG benchmark show that FES-RAG consistently outperforms state-of-the-art document-level MRAG methods, achieving up to 27 percent relative improvement in CIDEr. By selecting fewer yet more informative fragments, our approach substantially reduces context length while improving factual accuracy and generation coherence.

Tianqi Gao, Chengkai Huang, Zihan Wang et al.

Large language models (LLMs) have recently been adopted for recommendation by framing user preference modeling as a language generation problem. However, existing latent reasoning approaches typically represent user intent with a single latent vector, which struggles to capture the inherently multi-faceted nature of user preferences. We propose Factorized Latent Reasoning (FLR), a novel framework for LLM-based sequential recommendation that decomposes latent reasoning into multiple disentangled preference factors. FLR introduces a lightweight multi-factor attention module that iteratively refines a latent thought representation, where each factor attends to distinct aspects of the user's interaction history. To encourage diversity and specialization, we design orthogonality, attention diversity, and sparsity regularization objectives, and dynamically aggregate factor contributions for the final prediction. We further integrate FLR with an efficient reinforcement learning strategy based on group-relative policy optimization, enabling stable alignment directly in the latent reasoning space. Experiments on multiple benchmarks show that FLR consistently outperforms strong baselines while improving robustness and interpretability.

Chengkai Huang, Shoujin Wang, Xianzhi Wang et al.

Sequential recommendation aims to predict the next item which interests users via modeling their interest in items over time. Most of the existing works on sequential recommendation model users' dynamic interest in specific items while overlooking users' static interest revealed by some static attribute information of items, e.g., category, or brand. Moreover, existing works often only consider the positive excitation of a user's historical interactions on his/her next choice on candidate items while ignoring the commonly existing negative excitation, resulting in insufficient modeling dynamic interest. The overlook of static interest and negative excitation will lead to incomplete interest modeling and thus impede the recommendation performance. To this end, in this paper, we propose modeling both static interest and negative excitation for dynamic interest to further improve the recommendation performance. Accordingly, we design a novel Static-Dynamic Interest Learning (SDIL) framework featured with a novel Temporal Positive and Negative Excitation Modeling (TPNE) module for accurate sequential recommendation. TPNE is specially designed for comprehensively modeling dynamic interest based on temporal positive and negative excitation learning. Extensive experiments on three real-world datasets show that SDIL can effectively capture both static and dynamic interest and outperforms state-of-the-art baselines.

Chengkai Huang, Junda Wu, Yu Xia et al.

Recent breakthroughs in Large Language Models (LLMs) have led to the emergence of agentic AI systems that extend beyond the capabilities of standalone models. By empowering LLMs to perceive external environments, integrate multimodal information, and interact with various tools, these agentic systems exhibit greater autonomy and adaptability across complex tasks. This evolution brings new opportunities to recommender systems (RS): LLM-based Agentic RS (LLM-ARS) can offer more interactive, context-aware, and proactive recommendations, potentially reshaping the user experience and broadening the application scope of RS. Despite promising early results, fundamental challenges remain, including how to effectively incorporate external knowledge, balance autonomy with controllability, and evaluate performance in dynamic, multimodal settings. In this perspective paper, we first present a systematic analysis of LLM-ARS: (1) clarifying core concepts and architectures; (2) highlighting how agentic capabilities -- such as planning, memory, and multimodal reasoning -- can enhance recommendation quality; and (3) outlining key research questions in areas such as safety, efficiency, and lifelong personalization. We also discuss open problems and future directions, arguing that LLM-ARS will drive the next wave of RS innovation. Ultimately, we foresee a paradigm shift toward intelligent, autonomous, and collaborative recommendation experiences that more closely align with users' evolving needs and complex decision-making processes.

Xihang Wang, Zihan Wang, Chengkai Huang et al.

Multimodal Retrieval-Augmented Generation (MRAG) addresses key limitations of Multimodal Large Language Models (MLLMs), such as hallucination and outdated knowledge. However, current MRAG systems struggle to distinguish whether retrieved multimodal data truly supports the semantic core of an answer or merely provides superficial relevance. Existing metrics often rely on heuristic position-based confidence, which fails to capture the informational density of multimodal entities. To address this, we propose Multi-modal Evidence Grounding (MEG), a semantic-aware metric that quantifies the contribution of retrieved evidence. Unlike standard confidence measures, MEG utilizes Semantic Certainty Anchoring, focusing on high-IDF information-bearing tokens that better capture the semantic core of the answer. Building on MEG, we introduce MEG-RAG, a framework that trains a multimodal reranker to align retrieved evidence with the semantic anchors of the ground truth. By prioritizing high-value content based on semantic grounding rather than token probability distributions, MEG-RAG improves the accuracy and multimodal consistency of generated outputs. Extensive experiments on the M$^2$RAG benchmark show that MEG-RAG consistently outperforms strong baselines and demonstrates robust generalization across different teacher models.

Hongtao Huang, Chengkai Huang, Tong Yu et al.

Recent advancements in diffusion models have shown promising results in sequential recommendation (SR). Existing approaches predominantly rely on implicit conditional diffusion models, which compress user behaviors into a single representation during the forward diffusion process. While effective to some extent, this oversimplification often leads to the loss of sequential and contextual information, which is critical for understanding user behavior. Moreover, explicit information, such as user-item interactions or sequential patterns, remains underutilized, despite its potential to directly guide the recommendation process and improve precision. However, combining implicit and explicit information is non-trivial, as it requires dynamically integrating these complementary signals while avoiding noise and irrelevant patterns within user behaviors. To address these challenges, we propose Dual Conditional Diffusion Models for Sequential Recommendation (DCRec), which effectively integrates implicit and explicit information by embedding dual conditions into both the forward and reverse diffusion processes. This allows the model to retain valuable sequential and contextual information while leveraging explicit user-item interactions to guide the recommendation process. Specifically, we introduce the Dual Conditional Diffusion Transformer (DCDT), which employs a cross-attention mechanism to dynamically integrate explicit signals throughout the diffusion stages, ensuring contextual understanding and minimizing the influence of irrelevant patterns. This design enables precise and contextually relevant recommendations. Extensive experiments on public benchmark datasets demonstrate that DCRec significantly outperforms state-of-the-art methods in both accuracy and computational efficiency.

Chengkai Huang, Hongtao Huang, Tong Yu et al. · amazon-science

Recommender systems (RS) have become essential in filtering information and personalizing content for users. RS techniques have traditionally relied on modeling interactions between users and items as well as the features of content using models specific to each task. The emergence of foundation models (FMs), large scale models trained on vast amounts of data such as GPT, LLaMA and CLIP, is reshaping the recommendation paradigm. This survey provides a comprehensive overview of the Foundation Models for Recommender Systems (FM4RecSys), covering their integration in three paradigms: (1) Feature-Based augmentation of representations, (2) Generative recommendation approaches, and (3) Agentic interactive systems. We first review the data foundations of RS, from traditional explicit or implicit feedback to multimodal content sources. We then introduce FMs and their capabilities for representation learning, natural language understanding, and multi-modal reasoning in RS contexts. The core of the survey discusses how FMs enhance RS under different paradigms. Afterward, we examine FM applications in various recommendation tasks. Through an analysis of recent research, we highlight key opportunities that have been realized as well as challenges encountered. Finally, we outline open research directions and technical challenges for next-generation FM4RecSys. This survey not only reviews the state-of-the-art methods but also provides a critical analysis of the trade-offs among the feature-based, the generative, and the agentic paradigms, outlining key open issues and future research directions.

Chengkai Huang, Yu Xia, Rui Wang et al.

Retrieval-augmented large language models (LLMs) have been remarkably competent in various NLP tasks. However, it was observed by previous works that retrieval is not always helpful, especially when the LLM is already knowledgeable on the query to answer. Motivated by this, Adaptive Retrieval-Augmented Generation (ARAG) studies retrieving only when the knowledge asked by the query is absent in the LLM. Previous works of ARAG either require accessing the pre-training corpus or prompting with additional model inferences. Aiming to avoid such drawbacks, we propose to determine whether the model is knowledgeable on a query via inspecting the (contextualized) pre-trained token embeddings of LLMs. We hypothesize that such embeddings capture rich information on the model's intrinsic knowledge base, which enables an efficient way of judging the necessity to retrieve from an external corpus. Extensive experiments demonstrate our ARAG approach's superior performance across various benchmarks.

Shuguang Jiao, Chengkai Huang, Shuhan Qi et al.

Agentic Retrieval-Augmented Generation (Agentic RAG) has become a widely adopted paradigm for multi-hop question answering and complex knowledge reasoning, where retrieval and reasoning are interleaved at inference time. As reasoning trajectories grow longer, failures become increasingly common. Existing approaches typically address such failures by either stopping at diagnostic analysis or rerunning the entire retrieval-reasoning pipeline, which leads to substantial computational overhead and redundant reasoning. In this paper, we propose Doctor-RAG (DR-RAG), a unified diagnose-and-repair framework that corrects failures in Agentic RAG through explicit error localization and prefix reuse, enabling minimal-cost intervention. DR-RAG decomposes failure handling into two consecutive stages: (i) trajectory-level failure diagnosis and localization, which attributes errors to a coverage-gated taxonomy and identifies the earliest failure point in the reasoning trajectory; and (ii) tool-conditioned local repair, which intervenes only at the diagnosed failure point while maximally reusing validated reasoning prefixes and retrieved evidence. By explicitly separating error attribution from correction, DR-RAG enables precise error localization, thereby avoiding expensive full-pipeline reruns and enabling targeted, efficient repair. We evaluate DR-RAG across three multi-hop question answering benchmarks, multiple agentic RAG baselines, and different backbone models. Experimental results demonstrate that DR-RAG substantially improves answer accuracy while significantly reducing reasoning token consumption compared to rerun-based repair strategies.

Ruhan Wang, Zhiyong Wang, Chengkai Huang et al.

For question-answering (QA) tasks, in-context learning (ICL) enables language models to generate responses without modifying their parameters by leveraging examples provided in the input. However, the effectiveness of ICL heavily depends on the availability of high-quality examples, which are often scarce due to data privacy constraints, annotation costs, and distribution disparities. A natural solution is to utilize examples stored on client devices, but existing approaches either require transmitting model parameters - incurring significant communication overhead - or fail to fully exploit local datasets, limiting their effectiveness. To address these challenges, we propose Federated In-Context Learning (Fed-ICL), a general framework that enhances ICL through an iterative, collaborative process. Fed-ICL progressively refines responses by leveraging multi-round interactions between clients and a central server, improving answer quality without the need to transmit model parameters. We establish theoretical guarantees for the convergence of Fed-ICL and conduct extensive experiments on standard QA benchmarks, demonstrating that our proposed approach achieves strong performance while maintaining low communication costs.

Chengkai Huang, Junda Wu, Zhouhang Xie et al.

Personalized preference alignment for LLMs with diverse human preferences requires evaluation and alignment methods that capture pluralism. Most existing preference alignment datasets are logged under policies that differ substantially from the evaluated LLMs, and existing off-policy estimators focus solely on overall utility while ignoring preference pluralism. Extending Off-Policy Evaluation (OPE) to pluralistic preference alignment, therefore, remains an open question. Thus, we propose the Pluralistic Off-Policy Evaluation (POPE), the first framework for offline pluralistic preference evaluation and alignment in LLMs. POPE includes a unified reward function that combines (1) a collaborative utility component derived from human preference signals (e.g., upvotes or relevance scores) and (2) a diversity component inspired by entropy-based coverage measures, together reflecting pluralistic alignment. Furthermore, to estimate this reward from logged interactions, we derive decomposable inverse propensity scoring (IPS) estimators that separately evaluate relevance and diversity. Theoretically, we prove that our decomposed IPS estimators establish a lower bound on their variance. With the off-policy evaluated value function, we can directly enable off-policy optimization to further enhance pluralistic alignment. Empirical results demonstrate that POPE efficiently enhances pluralistic response generation and maintains the models' general capabilities on downstream tasks

Xiaoxin Ye, Chengkai Huang, Hongtao Huang et al.

Cross-Domain Sequential Recommendation (CDSR) leverages user behaviors across domains to enhance recommendation quality. However, naive aggregation of sequential signals can introduce conflicting domain-specific preferences, leading to negative transfer. While Sequential Recommendation (SR) already suffers from noisy behaviors such as misclicks and impulsive actions, CDSR further amplifies this issue due to domain heterogeneity arising from diverse item types and user intents. The core challenge is disentangling three intertwined signals: domain-invariant preferences, domain-specific preferences, and noise. Diffusion Models (DMs) offer a generative denoising framework well-suited for disentangling complex user preferences and enhancing robustness to noise. Their iterative refinement process enables gradual denoising, making them effective at capturing subtle preference signals. However, existing applications in recommendation face notable limitations: sequential DMs often conflate shared and domain-specific preferences, while cross-domain collaborative filtering DMs neglect temporal dynamics, limiting their ability to model evolving user preferences. To bridge these gaps, we propose \textbf{DPG-Diff}, a novel Disentangled Preference-Guided Diffusion Model, the first diffusion-based approach tailored for CDSR, to or best knowledge. DPG-Diff decomposes user preferences into domain-invariant and domain-specific components, which jointly guide the reverse diffusion process. This disentangled guidance enables robust cross-domain knowledge transfer, mitigates negative transfer, and filters sequential noise. Extensive experiments on real-world datasets demonstrate that DPG-Diff consistently outperforms state-of-the-art baselines across multiple metrics.

Le Pan, Yuanjiang Cao, Chengkai Huang et al.

Recommender Systems (RSs) aim to provide personalized recommendations for users. A newly discovered bias, known as sentiment bias, uncovers a common phenomenon within Review-based RSs (RRSs): the recommendation accuracy of users or items with negative reviews deteriorates compared with users or items with positive reviews. Critical users and niche items are disadvantaged by such unfair recommendations. We study this problem from the perspective of counterfactual inference with two stages. At the model training stage, we build a causal graph and model how sentiment influences the final rating score. During the inference stage, we decouple the direct and indirect effects to mitigate the impact of sentiment bias and remove the indirect effect using counterfactual inference. We have conducted extensive experiments, and the results validate that our model can achieve comparable performance on rating prediction for better recommendations and effective mitigation of sentiment bias. To the best of our knowledge, this is the first work to employ counterfactual inference on sentiment bias mitigation in RSs.

Shutian Gu, Chengkai Huang, Ruoyu Wang et al.

Vision-and-Language Navigation (VLN) requires an agent to follow natural-language instructions and navigate through previously unseen environments. Recent approaches increasingly employ large language models (LLMs) as high-level navigators due to their flexibility and reasoning capability. However, prompt-based LLM navigation often suffers from inefficient decision-making, as the model must repeatedly interpret instructions from scratch and reason over noisy and verbose navigable candidates at each step. In this paper, we propose a retrieval-augmented framework to improve the efficiency and stability of LLM-based VLN without modifying or fine-tuning the underlying language model. Our approach introduces retrieval at two complementary levels. At the episode level, an instruction-level embedding retriever selects semantically similar successful navigation trajectories as in-context exemplars, providing task-specific priors for instruction grounding. At the step level, an imitation-learned candidate retriever prunes irrelevant navigable directions before LLM inference, reducing action ambiguity and prompt complexity. Both retrieval modules are lightweight, modular, and trained independently of the LLM. We evaluate our method on the Room-to-Room (R2R) benchmark. Experimental results demonstrate consistent improvements in Success Rate, Oracle Success Rate, and SPL on both seen and unseen environments. Ablation studies further show that instruction-level exemplar retrieval and candidate pruning contribute complementary benefits to global guidance and step-wise decision efficiency. These results indicate that retrieval-augmented decision support is an effective and scalable strategy for enhancing LLM-based vision-and-language navigation.

Xingmei Wang, Xiaoyu Hu, Chengkai Huang et al.

Bridging 2D and 3D sensor modalities is critical for robust perception in autonomous systems. However, image-to-point cloud (I2P) registration remains challenging due to the semantic-geometric gap between texture-rich but depth-ambiguous images and sparse yet metrically precise point clouds, as well as the tendency of existing methods to converge to local optima. To overcome these limitations, we introduce CrossI2P, a self-supervised framework that unifies cross-modal learning and two-stage registration in a single end-to-end pipeline. First, we learn a geometric-semantic fused embedding space via dual-path contrastive learning, enabling annotation-free, bidirectional alignment of 2D textures and 3D structures. Second, we adopt a coarse-to-fine registration paradigm: a global stage establishes superpoint-superpixel correspondences through joint intra-modal context and cross-modal interaction modeling, followed by a geometry-constrained point-level refinement for precise registration. Third, we employ a dynamic training mechanism with gradient normalization to balance losses for feature alignment, correspondence refinement, and pose estimation. Extensive experiments demonstrate that CrossI2P outperforms state-of-the-art methods by 23.7% on the KITTI Odometry benchmark and by 37.9% on nuScenes, significantly improving both accuracy and robustness.

Zhiqiang Liu, Chengkai Huang, Yanxia Liu

Knowledge distillation has become an important approach to obtain a compact yet effective model. To achieve this goal, a small student model is trained to exploit the knowledge of a large well-trained teacher model. However, due to the capacity gap between the teacher and the student, the student's performance is hard to reach the level of the teacher. Regarding this issue, existing methods propose to reduce the difficulty of the teacher's knowledge via a proxy way. We argue that these proxy-based methods overlook the knowledge loss of the teacher, which may cause the student to encounter capacity bottlenecks. In this paper, we alleviate the capacity gap problem from a new perspective with the purpose of averting knowledge loss. Instead of sacrificing part of the teacher's knowledge, we propose to build a more powerful student via adversarial collaborative learning. To this end, we further propose an Adversarial Collaborative Knowledge Distillation (ACKD) method that effectively improves the performance of knowledge distillation. Specifically, we construct the student model with multiple auxiliary learners. Meanwhile, we devise an adversarial collaborative module (ACM) that introduces attention mechanism and adversarial learning to enhance the capacity of the student. Extensive experiments on four classification tasks show the superiority of the proposed ACKD.

Zhiqiang Liu, Yanxia Liu, Chengkai Huang

Knowledge distillation is an effective and stable method for model compression via knowledge transfer. Conventional knowledge distillation (KD) is to transfer knowledge from a large and well pre-trained teacher network to a small student network, which is a one-way process. Recently, deep mutual learning (DML) has been proposed to help student networks learn collaboratively and simultaneously. However, to the best of our knowledge, KD and DML have never been jointly explored in a unified framework to solve the knowledge distillation problem. In this paper, we investigate that the teacher model supports more trustworthy supervision signals in KD, while the student captures more similar behaviors from the teacher in DML. Based on these observations, we first propose to combine KD with DML in a unified framework. Furthermore, we propose a Semi-Online Knowledge Distillation (SOKD) method that effectively improves the performance of the student and the teacher. In this method, we introduce the peer-teaching training fashion in DML in order to alleviate the student's imitation difficulty, and also leverage the supervision signals provided by the well-trained teacher in KD. Besides, we also show our framework can be easily extended to feature-based distillation methods. Extensive experiments on CIFAR-100 and ImageNet datasets demonstrate the proposed method achieves state-of-the-art performance.