Guihai Chen

Chengfei Lv, Chaoyue Niu, Renjie Gu et al.

To break the bottlenecks of mainstream cloud-based machine learning (ML) paradigm, we adopt device-cloud collaborative ML and build the first end-to-end and general-purpose system, called Walle, as the foundation. Walle consists of a deployment platform, distributing ML tasks to billion-scale devices in time; a data pipeline, efficiently preparing task input; and a compute container, providing a cross-platform and high-performance execution environment, while facilitating daily task iteration. Specifically, the compute container is based on Mobile Neural Network (MNN), a tensor compute engine along with the data processing and model execution libraries, which are exposed through a refined Python thread-level virtual machine (VM) to support diverse ML tasks and concurrent task execution. The core of MNN is the novel mechanisms of operator decomposition and semi-auto search, sharply reducing the workload in manually optimizing hundreds of operators for tens of hardware backends and further quickly identifying the best backend with runtime optimization for a computation graph. The data pipeline introduces an on-device stream processing framework to enable processing user behavior data at source. The deployment platform releases ML tasks with an efficient push-then-pull method and supports multi-granularity deployment policies. We evaluate Walle in practical e-commerce application scenarios to demonstrate its effectiveness, efficiency, and scalability. Extensive micro-benchmarks also highlight the superior performance of MNN and the Python thread-level VM. Walle has been in large-scale production use in Alibaba, while MNN has been open source with a broad impact in the community.

Haolin Zhou, Junwei Pan, Xinyi Zhou et al.

User response prediction is essential in industrial recommendation systems, such as online display advertising. Among all the features in recommendation models, user behaviors are among the most critical. Many works have revealed that a user's behavior reflects her interest in the candidate item, owing to the semantic or temporal correlation between behaviors and the candidate. While the literature has individually examined each of these correlations, researchers have yet to analyze them in combination, that is, the semantic-temporal correlation. We empirically measure this correlation and observe intuitive yet robust patterns. We then examine several popular user interest models and find that, surprisingly, none of them learn such correlation well. To fill this gap, we propose a Temporal Interest Network (TIN) to capture the semantic-temporal correlation simultaneously between behaviors and the target. We achieve this by incorporating target-aware temporal encoding, in addition to semantic encoding, to represent behaviors and the target. Furthermore, we conduct explicit 4-way interaction by deploying target-aware attention and target-aware representation to capture both semantic and temporal correlation. We conduct comprehensive evaluations on two popular public datasets, and our proposed TIN outperforms the best-performing baselines by 0.43% and 0.29% on GAUC, respectively. During online A/B testing in Tencent's advertising platform, TIN achieves 1.65% cost lift and 1.93% GMV lift over the base model. It has been successfully deployed in production since October 2023, serving the WeChat Moments traffic. We have released our code at https://github.com/zhouxy1003/TIN.

Yang Luo, Shiyu Wang, Zhemeng Yu et al.

The surging demand for cloud computing resources, driven by the rapid growth of sophisticated large-scale models and data centers, underscores the critical importance of efficient and adaptive resource allocation. As major tech enterprises deploy massive infrastructures with thousands of GPUs, existing cloud platforms still struggle with low resource utilization due to key challenges: capturing hierarchical indicator structures, modeling non-Gaussian distributions, and decision-making under uncertainty. To address these challenges, we propose HRAMONY, an adaptive Hierarchical Attention-based Resource Modeling and Decision-Making System. HARMONY combines hierarchical multi-indicator distribution forecasting and uncertainty-aware Bayesian decision-making. It introduces a novel hierarchical attention mechanism that comprehensively models complex inter-indicator dependencies, enabling accurate predictions that can adapt to evolving environment states. By transforming Gaussian projections into adaptive non-Gaussian distributions via Normalizing Flows. Crucially, HARMONY leverages the full predictive distributions in an adaptive Bayesian process, proactively incorporating uncertainties to optimize resource allocation while robustly meeting SLA constraints under varying conditions. Extensive evaluations across four large-scale cloud datasets demonstrate HARMONY's state-of-the-art performance, significantly outperforming nine established methods. A month-long real-world deployment validated HARMONY's substantial practical impact, realizing over 35,000 GPU hours in savings and translating to $100K+ in cost reduction, showcasing its remarkable economic value through adaptive, uncertainty-aware scaling. Our code is available at https://github.com/Floating-LY/HARMONY1.

Zheng Chen, Yulun Zhang, Jinjin Gu et al.

Image super-resolution (SR) methods typically model degradation to improve reconstruction accuracy in complex and unknown degradation scenarios. However, extracting degradation information from low-resolution images is challenging, which limits the model performance. To boost image SR performance, one feasible approach is to introduce additional priors. Inspired by advancements in multi-modal methods and text prompt image processing, we introduce text prompts to image SR to provide degradation priors. Specifically, we first design a text-image generation pipeline to integrate text into the SR dataset through the text degradation representation and degradation model. By adopting a discrete design, the text representation is flexible and user-friendly. Meanwhile, we propose the PromptSR to realize the text prompt SR. The PromptSR leverages the latest multi-modal large language model (MLLM) to generate prompts from low-resolution images. It also utilizes the pre-trained language model (e.g., T5 or CLIP) to enhance text comprehension. We train the PromptSR on the text-image dataset. Extensive experiments indicate that introducing text prompts into SR, yields impressive results on both synthetic and real-world images. Code: https://github.com/zhengchen1999/PromptSR.

Jiarui Zhang, Xiangyu Liu, Yong Hu et al.

Multi-turn dialogue is the predominant form of interaction with large language models (LLMs). While LLM routing is effective in single-turn settings, existing methods fail to maximize cumulative performance in multi-turn dialogue due to interaction dynamics and delayed rewards. To address this challenge, we move from myopic, single-turn selection to long-horizon sequential routing for multi-turn dialogue. Accordingly, we propose DialRouter, which first performs MCTS to explore dialogue branches induced by different LLM selections and collect trajectories with high cumulative rewards. DialRouter then learns a lightweight routing policy from search-derived data, augmented with retrieval-based future state approximation, enabling multi-turn routing without online search. Experiments on both open-domain and domain-specific dialogue tasks across diverse candidate sets of both open-source and closed-source LLMs demonstrate that DialRouter significantly outperforms single LLMs and existing routing baselines in task success rate, while achieving a superior performance-cost trade-off when combined with a cost-aware reward.

Chen Gong, Zhenzhe Zheng, Yunfeng Shao et al.

Machine learning models have been deployed in mobile networks to deal with massive data from different layers to enable automated network management and intelligence on devices. To overcome high communication cost and severe privacy concerns of centralized machine learning, federated learning (FL) has been proposed to achieve distributed machine learning among networked devices. While the computation and communication limitation has been widely studied, the impact of on-device storage on the performance of FL is still not explored. Without an effective data selection policy to filter the massive streaming data on devices, classical FL can suffer from much longer model training time ($4\times$) and significant inference accuracy reduction ($7\%$), observed in our experiments. In this work, we take the first step to consider the online data selection for FL with limited on-device storage. We first define a new data valuation metric for data evaluation and selection in FL with theoretical guarantees for speeding up model convergence and enhancing final model accuracy, simultaneously. We further design {\ttfamily ODE}, a framework of \textbf{O}nline \textbf{D}ata s\textbf{E}lection for FL, to coordinate networked devices to store valuable data samples. Experimental results on one industrial dataset and three public datasets show the remarkable advantages of {\ttfamily ODE} over the state-of-the-art approaches. Particularly, on the industrial dataset, {\ttfamily ODE} achieves as high as $2.5\times$ speedup of training time and $6\%$ increase in inference accuracy, and is robust to various factors in practical environments.

Yucheng Ding, Chaoyue Niu, Fan Wu et al.

Many large vision models have been deployed on the cloud for real-time services. Meanwhile, fresh samples are continuously generated on the served mobile device. How to leverage the device-side samples to improve the cloud-side large model becomes a practical requirement, but falls into the dilemma of no raw sample up-link and no large model down-link. Specifically, the user may opt out of sharing raw samples with the cloud due to the concern of privacy or communication overhead, while the size of some large vision models far exceeds the mobile device's runtime capacity. In this work, we propose a device-cloud collaborative controlled learning framework, called DC-CCL, enabling a cloud-side large vision model that cannot be directly deployed on the mobile device to still benefit from the device-side local samples. In particular, DC-CCL vertically splits the base model into two submodels, one large submodel for learning from the cloud-side samples and the other small submodel for learning from the device-side samples and performing device-cloud knowledge fusion. Nevertheless, on-device training of the small submodel requires the output of the cloud-side large submodel to compute the desired gradients. DC-CCL thus introduces a light-weight model to mimic the large cloud-side submodel with knowledge distillation, which can be offloaded to the mobile device to control its small submodel's optimization direction. Given the decoupling nature of two submodels in collaborative learning, DC-CCL also allows the cloud to take a pre-trained model and the mobile device to take another model with a different backbone architecture.

Qiyang Li, Rui Kong, Yuchen Li et al.

The integration of dynamic, sparse structures like Mixture-of-Experts (MoE) with parameter-efficient adapters (e.g., LoRA) is a powerful technique for enhancing Large Language Models (LLMs). However, this architectural enhancement comes at a steep cost: despite minimal increases in computational load, the inference latency often skyrockets, leading to decoding speeds slowing by over 2.5 times. Through a fine-grained performance analysis, we pinpoint the primary bottleneck not in the computation itself, but in the severe overhead from fragmented, sequential CUDA kernel launches required for conventional dynamic routing. To address this challenge, we introduce AdaFuse, a framework built on a tight co-design between the algorithm and the underlying hardware system to enable efficient dynamic adapter execution. Departing from conventional layer-wise or block-wise routing, AdaFuse employs a token-level pre-gating strategy, which makes a single, global routing decision for all adapter layers before a token is processed. This "decide-once, apply-everywhere" approach effectively staticizes the execution path for each token, creating an opportunity for holistic optimization. We capitalize on this by developing a custom CUDA kernel that performs a fused switching operation, merging the parameters of all selected LoRA adapters into the backbone model in a single, efficient pass. Experimental results on popular open-source LLMs show that AdaFuse achieves accuracy on par with state-of-the-art dynamic adapters while drastically cutting decoding latency by a factor of over 2.4x, thereby bridging the gap between model capability and inference efficiency.

Shiju Zhao, Junhao Hu, Jiaqi Zheng et al.

The Key-Value (KV) cache of Large Language Models (LLMs) is prefix-based, making it highly inefficient for processing contexts retrieved in arbitrary order. Position-Independent Caching (PIC) has been proposed to enable KV reuse without positional constraints; however, existing approaches often incur substantial accuracy degradation, limiting their practical adoption. To address this issue, we propose native PIC by reintroducing the encoder to prevalent decoder-only LLMs and explicitly training it to support PIC. We further develop COMB, a PIC-aware caching system that integrates seamlessly with existing inference frameworks. Experimental results show that COMB reduces Time-to-First-Token (TTFT) by 51-94% and increases throughput by 3$\times$ with comparable accuracy. Furthermore, the quality improvement when using DeepSeek-V2-Lite-Chat demonstrates the applicability of COMB to other types of decoder-only LLMs. Our code is available at https://github.com/shijuzhao/Comb.

Zuowu Zheng, Xiaofeng Gao, Junwei Pan et al.

In Click-through rate (CTR) prediction models, a user's interest is usually represented as a fixed-length vector based on her history behaviors. Recently, several methods are proposed to learn an attentive weight for each user behavior and conduct weighted sum pooling. However, these methods only manually select several fields from the target item side as the query to interact with the behaviors, neglecting the other target item fields, as well as user and context fields. Directly including all these fields in the attention may introduce noise and deteriorate the performance. In this paper, we propose a novel model named AutoAttention, which includes all item/user/context side fields as the query, and assigns a learnable weight for each field pair between behavior fields and query fields. Pruning on these field pairs via these learnable weights lead to automatic field pair selection, so as to identify and remove noisy field pairs. Though including more fields, the computation cost of AutoAttention is still low due to using a simple attention function and field pair selection. Extensive experiments on the public dataset and Tencent's production dataset demonstrate the effectiveness of the proposed approach.

Yucheng Ding, Chaoyue Niu, Fan Wu et al.

To meet the practical requirements of low latency, low cost, and good privacy in online intelligent services, more and more deep learning models are offloaded from the cloud to mobile devices. To further deal with cross-device data heterogeneity, the offloaded models normally need to be fine-tuned with each individual user's local samples before being put into real-time inference. In this work, we focus on the fundamental click-through rate (CTR) prediction task in recommender systems and study how to effectively and efficiently perform on-device fine-tuning. We first identify the bottleneck issue that each individual user's local CTR (i.e., the ratio of positive samples in the local dataset for fine-tuning) tends to deviate from the global CTR (i.e., the ratio of positive samples in all the users' mixed datasets on the cloud for training out the initial model). We further demonstrate that such a CTR drift problem makes on-device fine-tuning even harmful to item ranking. We thus propose a novel label correction method, which requires each user only to change the labels of the local samples ahead of on-device fine-tuning and can well align the locally prior CTR with the global CTR. The offline evaluation results over three datasets and five CTR prediction models as well as the online A/B testing results in Mobile Taobao demonstrate the necessity of label correction in on-device fine-tuning and also reveal the improvement over cloud-based learning without fine-tuning.

Yan Zhuang, Zhenzhe Zheng, Yunfeng Shao et al.

Pervasive mobile AI applications primarily employ one of the two learning paradigms: cloud-based learning (with powerful large models) or on-device learning (with lightweight small models). Despite their own advantages, neither paradigm can effectively handle dynamic edge environments with frequent data distribution shifts and on-device resource fluctuations, inevitably suffering from performance degradation. In this paper, we propose ECLM, an edge-cloud collaborative learning framework for rapid model adaptation for dynamic edge environments. We first propose a novel block-level model decomposition design to decompose the original large cloud model into multiple combinable modules. By flexibly combining a subset of the modules, this design enables the derivation of compact, task-specific sub-models for heterogeneous edge devices from the large cloud model, and the seamless integration of new knowledge learned on these devices into the cloud model periodically. As such, ECLM ensures that the cloud model always provides up-to-date sub-models for edge devices. We further propose an end-to-end learning framework that incorporates the modular model design into an efficient model adaptation pipeline including an offline on-cloud model prototyping and training stage, and an online edge-cloud collaborative adaptation stage. Extensive experiments over various datasets demonstrate that ECLM significantly improves model performance (e.g., 18.89% accuracy increase) and resource efficiency (e.g., 7.12x communication cost reduction) in adapting models to dynamic edge environments by efficiently collaborating the edge and the cloud models.

Yikai Yan, Chaoyue Niu, Fan Wu et al.

The mainstream workflow of image recognition applications is first training one global model on the cloud for a wide range of classes and then serving numerous clients, each with heterogeneous images from a small subset of classes to be recognized. From the cloud-client discrepancies on the range of image classes, the recognition model is desired to have strong adaptiveness, intuitively by concentrating the focus on each individual client's local dynamic class subset, while incurring negligible overhead. In this work, we propose to plug a new intra-client and inter-image attention (ICIIA) module into existing backbone recognition models, requiring only one-time cloud-based training to be client-adaptive. In particular, given a target image from a certain client, ICIIA introduces multi-head self-attention to retrieve relevant images from the client's historical unlabeled images, thereby calibrating the focus and the recognition result. Further considering that ICIIA's overhead is dominated by linear projection, we propose partitioned linear projection with feature shuffling for replacement and allow increasing the number of partitions to dramatically improve efficiency without scarifying too much accuracy. We finally evaluate ICIIA using 3 different recognition tasks with 9 backbone models over 5 representative datasets. Extensive evaluation results demonstrate the effectiveness and efficiency of ICIIA. Specifically, for ImageNet-1K with the backbone models of MobileNetV3-L and Swin-B, ICIIA can improve the testing accuracy to 83.37% (+8.11%) and 88.86% (+5.28%), while adding only 1.62% and 0.02% of FLOPs, respectively.

Zhiying Xu, Jiafan Xu, Hongding Peng et al.

Deep learning models rely on highly optimized tensor libraries for efficient inference on heterogeneous hardware. Current deep compilers typically predetermine layouts of tensors and then optimize loops of operators. However, such unidirectional and one-off workflow strictly separates graph-level optimization and operator-level optimization into different system layers, missing opportunities for unified tuning. This paper proposes ALT, a compiler that performs joint graph- and operator-level optimizations for deep models. ALT provides a generic transformation module to manipulate layouts and loops with easy-to-use primitive functions. ALT further integrates an auto-tuning module that jointly optimizes graph-level data layouts and operator-level loops while guaranteeing efficiency. Experimental results show that ALT significantly outperforms state-of-the-art compilers (e.g., Ansor) in terms of both single operator performance (e.g., 1.5x speedup on average) and end-to-end inference performance (e.g., 1.4x speedup on average).

Chenxin Mao, Shangyu Liu, Zhenzhe Zheng et al.

Retrieval-Augmented Generation (RAG) empowers LLMs with external knowledge, making cross-institutional domain-specific knowledge base integration a highly promising deployment paradigm. Despite this potential, strict privacy regulations create severe "data silos" that obstruct such collaboration. Building federated RAG systems requires distributed inference, but the Transformer's self-attention mechanism fundamentally conflicts with this by mandating cross-node access to distributed Key-Value caches. To address this challenge, we present FedRAG, a high-throughput, privacy-preserving federated RAG framework. At its core is a novel Scrambled Distributed Attention protocol that utilizes numerically stable feature scrambling and token permutation. By dynamically delegating scrambled computations to collaborating nodes, our system successfully decouples attention execution from data localization without exposing plaintext. Crucially, our approach requires no specialized hardware or model retraining, circumventing the prohibitive latency and communication overheads of cryptographic solutions while robustly defending against intermediate state inversion attacks. Extensive evaluations demonstrate our framework preserves negligible (<0.1\%) model utility degradation and achieves up to a 62$\times$ latency reduction over existing secure baselines, sustaining practical, human-reading throughput for cross-institutional knowledge synergy.

Hang Zeng, Xiangyu Liu, Yong Hu et al.

Large language models (LLMs) offer strong capabilities but raise cost and privacy concerns, whereas small language models (SLMs) facilitate efficient and private local inference yet suffer from limited capacity. To synergize the complementary strengths, we introduce a dynamic collaboration framework, where an SLM learns to proactively decide how to request an LLM during multi-step reasoning, while the LLM provides adaptive feedback instead of acting as a passive tool. We further systematically investigate how collaboration strategies are shaped by SLM and LLM capabilities as well as efficiency and privacy constraints. Evaluation results reveal a distinct scaling effect: stronger SLMs become more self-reliant, while stronger LLMs enable fewer and more informative interactions. In addition, the learned dynamic collaboration strategies significantly outperform static pipelines and standalone inference, and transfer robustly to unseen LLMs.

Hongxuan Zhang, Zhining Liu, Yao Zhao et al.

In this work, we propose FastCoT, a model-agnostic framework based on parallel decoding without any further training of an auxiliary model or modification to the LLM itself. FastCoT uses a size-varying context window whose size changes with position to conduct parallel decoding and auto-regressive decoding simultaneously, thus fully utilizing GPU computation resources. In FastCoT, the parallel decoding part provides the LLM with a quick glance of the future composed of approximate tokens, which could lead to faster answers compared to regular autoregressive decoding used by causal transformers. We also provide an implementation of parallel decoding within LLM, which supports KV-cache generation and batch processing. Through extensive experiments, we demonstrate that FastCoT saves inference time by nearly 20% with only a negligible performance drop compared to the regular approach. Additionally, we show that the context window size exhibits considerable robustness for different tasks.

Yumou Liu, Zhenzhe Zheng, Jiang Rong et al.

Modern content platforms offer paid promotion to mitigate cold start by allocating exposure via auctions. Our empirical analysis reveals a counterintuitive flaw in this paradigm: while promotion rescues low-to-medium quality content, it can harm high-quality content by forcing exposure to suboptimal audiences, polluting engagement signals and downgrading future recommendation. We recast content promotion as a dual-objective optimization that balances short-term value acquisition with long-term model improvement. To make this tractable at bid time in content promotion, we introduce a decomposable surrogate objective, gradient coverage, and establish its formal connection to Fisher Information and optimal experimental design. We design a two-stage auto-bidding algorithm based on Lagrange duality that dynamically paces budget through a shadow price and optimizes impression-level bids using per-impression marginal utilities. To address missing labels at bid time, we propose a confidence-gated gradient heuristic, paired with a zeroth-order variant for black-box models that reliably estimates learning signals in real time. We provide theoretical guarantees, proving monotone submodularity of the composite objective, sublinear regret in online auction, and budget feasibility. Extensive offline experiments on synthetic and real-world datasets validate the framework: it outperforms baselines, achieves superior final AUC/LogLoss, adheres closely to budget targets, and remains effective when gradients are approximated zeroth-order. These results show that strategic, information-aware promotion can improve long-term model performance and organic outcomes beyond naive impression-maximization strategies.

Yuchen Li, Haoyi Xiong, Linghe Kong et al.

Learning to rank (LTR) is widely employed in web searches to prioritize pertinent webpages from retrieved content based on input queries. However, traditional LTR models encounter two principal obstacles that lead to suboptimal performance: (1) the lack of well-annotated query-webpage pairs with ranking scores covering a diverse range of search query popularities, which hampers their ability to address queries across the popularity spectrum, and (2) inadequately trained models that fail to induce generalized representations for LTR, resulting in overfitting. To address these challenges, we propose a \emph{\uline{G}enerative \uline{S}emi-\uline{S}upervised \uline{P}re-trained} (GS2P) LTR model. We conduct extensive offline experiments on both a publicly available dataset and a real-world dataset collected from a large-scale search engine. Furthermore, we deploy GS2P in a large-scale web search engine with realistic traffic, where we observe significant improvements in the real-world application.

Kai Liu, Qian Zheng, Kaiwen Tao et al.

With unprecedented rapid development, deep neural networks (DNNs) have deeply influenced almost all fields. However, their heavy computation costs and model sizes are usually unacceptable in real-world deployment. Model quantization, an effective weight-lighting technique, has become an indispensable procedure in the whole deployment pipeline. The essence of quantization acceleration is the conversion from continuous floating-point numbers to discrete integer ones, which significantly speeds up the memory I/O and calculation, i.e., addition and multiplication. However, performance degradation also comes with the conversion because of the loss of precision. Therefore, it has become increasingly popular and critical to investigate how to perform the conversion and how to compensate for the information loss. This article surveys the recent five-year progress towards low-bit quantization on DNNs. We discuss and compare the state-of-the-art quantization methods and classify them into 8 main categories and 24 sub-categories according to their core techniques. Furthermore, we shed light on the potential research opportunities in the field of model quantization. A curated list of model quantization is provided at https://github.com/Kai-Liu001/Awesome-Model-Quantization.

Jiarui Zhang, Xiangyu Liu, Yong Hu et al.

Retrieval-Augmented Generation (RAG) significantly improves the performance of Large Language Models (LLMs) on knowledge-intensive tasks. However, varying response quality across LLMs under RAG necessitates intelligent routing mechanisms, which select the most suitable model for each query from multiple retrieval-augmented LLMs via a dedicated router model. We observe that external documents dynamically affect LLMs' ability to answer queries, while existing routing methods, which rely on static parametric knowledge representations, exhibit suboptimal performance in RAG scenarios. To address this, we formally define the new retrieval-augmented LLM routing problem, incorporating the influence of retrieved documents into the routing framework. We propose RAGRouter, a RAG-aware routing design, which leverages document embeddings and RAG capability embeddings with contrastive learning to capture knowledge representation shifts and enable informed routing decisions. Extensive experiments on diverse knowledge-intensive tasks and retrieval settings, covering open and closed-source LLMs, show that RAGRouter outperforms the best individual LLM and existing routing methods. With an extended score-threshold-based mechanism, it also achieves strong performance-efficiency trade-offs under low-latency constraints. The code and data are available at https://github.com/OwwO99/RAGRouter.

Zhiteng Li, Hanxuan Li, Junyi Wu et al.

Diffusion Transformers (DiTs) have emerged as the state-of-the-art architecture for video generation, yet their computational and memory demands hinder practical deployment. While post-training quantization (PTQ) presents a promising approach to accelerate Video DiT models, existing methods suffer from two critical limitations: (1) dependence on computation-heavy and inflexible calibration procedures, and (2) considerable performance deterioration after quantization. To address these challenges, we propose DVD-Quant, a novel Data-free quantization framework for Video DiTs. Our approach integrates three key innovations: (1) Bounded-init Grid Refinement (BGR) and (2) Auto-scaling Rotated Quantization (ARQ) for calibration data-free quantization error reduction, as well as (3) $δ$-Guided Bit Switching ($δ$-GBS) for adaptive bit-width allocation. Extensive experiments across multiple video generation benchmarks demonstrate that DVD-Quant achieves an approximately 2$\times$ speedup over full-precision baselines on advanced DiT models while maintaining visual fidelity. Notably, DVD-Quant is the first to enable W4A4 PTQ for Video DiTs without compromising video quality. Code and models will be available at https://github.com/lhxcs/DVD-Quant.

Chenzhi Hu, Qinzhe Hu, Yuhang Xu et al.

Large reasoning models (LRMs) like OpenAI o1 and DeepSeek-R1 achieve high accuracy on complex tasks by adopting long chain-of-thought (CoT) reasoning paths. However, the inherent verbosity of these processes frequently results in redundancy and overthinking. To address this issue, existing works leverage Group Relative Policy Optimization (GRPO) to reduce LRM output length, but their static length reward design cannot dynamically adapt according to the relative problem difficulty and response length distribution, causing over-compression and compromised accuracy. Therefore, we propose SmartThinker, a novel GRPO-based efficient reasoning method with progressive CoT length calibration. SmartThinker makes a two-fold contribution: First, it dynamically estimates the optimal length with peak accuracy during training and guides overlong responses toward it to reduce response length while sustaining accuracy. Second, it dynamically modulates the length reward coefficient to avoid the unwarranted penalization of correct reasoning paths. Extensive experiment results show that SmartThinker achieves up to 52.5% average length compression with improved accuracy, and achieves up to 16.6% accuracy improvement on challenging benchmarks like AIME25. The source code can be found at https://github.com/SJTU-RTEAS/SmartThinker.

Junyi Chen, Shihao Bai, Zaijun Wang et al.

Extensive LLM applications demand efficient structured generations, particularly for LR(1) grammars, to produce outputs in specified formats (e.g., JSON). Existing methods primarily parse LR(1) grammars into a pushdown automaton (PDA), leading to runtime execution overhead for context-dependent token processing, especially inefficient under large inference batches. To address these issues, we propose Pre$^3$ that exploits deterministic pushdown automata (DPDA) to optimize the constrained LLM decoding efficiency. First, by precomputing prefix-conditioned edges during the preprocessing, Pre$^3$ enables ahead-of-time edge analysis and thus makes parallel transition processing possible. Second, by leveraging the prefix-conditioned edges, Pre$^3$ introduces a novel approach that transforms LR(1) transition graphs into DPDA, eliminating the need for runtime path exploration and achieving edge transitions with minimal overhead. Pre$^3$ can be seamlessly integrated into standard LLM inference frameworks, reducing time per output token (TPOT) by up to 40% and increasing throughput by up to 36% in our experiments. Our code is available at https://github.com/ModelTC/lightllm.

Gucongcong Fan, Chaoyue Niu, Chengfei Lyu et al.

Mobile agents rely on Large Language Models (LLMs) to plan and execute tasks on smartphone user interfaces (UIs). While cloud-based LLMs achieve high task accuracy, they require uploading the full UI state at every step, exposing unnecessary and often irrelevant information. In contrast, local LLMs avoid UI uploads but suffer from limited capacity, resulting in lower task success rates. We propose $\textbf{CORE}$, a $\textbf{CO}$llaborative framework that combines the strengths of cloud and local LLMs to $\textbf{R}$educe UI $\textbf{E}$xposure, while maintaining task accuracy for mobile agents. CORE comprises three key components: (1) $\textbf{Layout-aware block partitioning}$, which groups semantically related UI elements based on the XML screen hierarchy; (2) $\textbf{Co-planning}$, where local and cloud LLMs collaboratively identify the current sub-task; and (3) $\textbf{Co-decision-making}$, where the local LLM ranks relevant UI blocks, and the cloud LLM selects specific UI elements within the top-ranked block. CORE further introduces a multi-round accumulation mechanism to mitigate local misjudgment or limited context. Experiments across diverse mobile apps and tasks show that CORE reduces UI exposure by up to 55.6% while maintaining task success rates slightly below cloud-only agents, effectively mitigating unnecessary privacy exposure to the cloud. The code is available at https://github.com/Entropy-Fighter/CORE.

Minyuan Zhou, Yuning Chen, Jiaqi Zheng et al.

Operating large-scale anycast networks is challenging because client-to-site mappings often misalign with operator's expectation due to opaque inter-domain routing. We present AnyPro, the first system to unlock the full potential of AS-path prepending (ASPP), efficiently deriving globally optimal configurations to steer clients toward performance-optimal sites at scale. AnyPro first employs an efficient polling mechanism to identify all clients sensitive to ASPP. By analyzing the routing changes during the process, the system derives a set of ASPP constraints that guide client traffic toward the desired sites. We then formulate the anycast optimization problem as a constraint-based program and compute optimal ASPP configurations. Extensive evaluation on a global testbed with 20 PoPs demonstrates the effectiveness of AnyPro: it reduces the 90th percentile latency by 37.7% compared to baseline configurations without ASPP. Furthermore, we show that AnyPro can be integrated with PoP-level anycast optimization techniques to achieve additional performance gains.

Yang Luo, Haoyang Luan, Haoyun Pan et al.

Accurate quality prediction in multi-process manufacturing is critical for industrial efficiency but hindered by three core challenges: time-lagged process interactions, overlapping operations with mixed periodicity, and inter-process dependencies in shared frequency bands. To address these, we propose PAF-Net, a frequency decoupled time series prediction framework with three key innovations: (1) A phase-correlation alignment method guided by frequency domain energy to synchronize time-lagged quality series, resolving temporal misalignment. (2) A frequency independent patch attention mechanism paired with Discrete Cosine Transform (DCT) decomposition to capture heterogeneous operational features within individual series. (3) A frequency decoupled cross attention module that suppresses noise from irrelevant frequencies, focusing exclusively on meaningful dependencies within shared bands. Experiments on 4 real-world datasets demonstrate PAF-Net's superiority. It outperforms 10 well-acknowledged baselines by 7.06% lower MSE and 3.88% lower MAE. Our code is available at https://github.com/StevenLuan904/PAF-Net-Official.

Jiatong Li, Libo Zhu, Haotong Qin et al.

Diffusion models have been achieving remarkable performance in face restoration. However, the heavy computations of diffusion models make it difficult to deploy them on devices like smartphones. In this work, we propose QuantFace, a novel low-bit quantization for one-step diffusion face restoration models, where the full-precision (\ie, 32-bit) weights and activations are quantized to 4$\sim$6-bit. We first analyze the data distribution within activations and find that they are highly variant. To preserve the original data information, we employ rotation-scaling channel balancing. Furthermore, we propose Quantization-Distillation Low-Rank Adaptation (QD-LoRA) that jointly optimizes for quantization and distillation performance. Finally, we propose an adaptive bit-width allocation strategy. We formulate such a strategy as an integer programming problem, which combines quantization error and perceptual metrics to find a satisfactory resource allocation. Extensive experiments on the synthetic and real-world datasets demonstrate the effectiveness of QuantFace under 6-bit and 4-bit. QuantFace achieves significant advantages over recent leading low-bit quantization methods for face restoration. The code is available at https://github.com/jiatongli2024/QuantFace.

Kai Liu, Haotong Qin, Yong Guo et al.

Low-bit quantization has become widespread for compressing image super-resolution (SR) models for edge deployment, which allows advanced SR models to enjoy compact low-bit parameters and efficient integer/bitwise constructions for storage compression and inference acceleration, respectively. However, it is notorious that low-bit quantization degrades the accuracy of SR models compared to their full-precision (FP) counterparts. Despite several efforts to alleviate the degradation, the transformer-based SR model still suffers severe degradation due to its distinctive activation distribution. In this work, we present a dual-stage low-bit post-training quantization (PTQ) method for image super-resolution, namely 2DQuant, which achieves efficient and accurate SR under low-bit quantization. The proposed method first investigates the weight and activation and finds that the distribution is characterized by coexisting symmetry and asymmetry, long tails. Specifically, we propose Distribution-Oriented Bound Initialization (DOBI), using different searching strategies to search a coarse bound for quantizers. To obtain refined quantizer parameters, we further propose Distillation Quantization Calibration (DQC), which employs a distillation approach to make the quantized model learn from its FP counterpart. Through extensive experiments on different bits and scaling factors, the performance of DOBI can reach the state-of-the-art (SOTA) while after stage two, our method surpasses existing PTQ in both metrics and visual effects. 2DQuant gains an increase in PSNR as high as 4.52dB on Set5 (x2) compared with SOTA when quantized to 2-bit and enjoys a 3.60x compression ratio and 5.08x speedup ratio. The code and models will be available at https://github.com/Kai-Liu001/2DQuant.

Zixuan Yang, Jiaqi Zheng, Guihai Chen

Ensuring verifiable and interpretable safety of deep reinforcement learning (DRL) is crucial for its deployment in real-world applications. Existing approaches like verification-in-the-loop training, however, face challenges such as difficulty in deployment, inefficient training, lack of interpretability, and suboptimal performance in property satisfaction and reward performance. In this work, we propose a novel verification-driven interpretation-in-the-loop framework Reintrainer to develop trustworthy DRL models, which are guaranteed to meet the expected constraint properties. Specifically, in each iteration, this framework measures the gap between the on-training model and predefined properties using formal verification, interprets the contribution of each input feature to the model's output, and then generates the training strategy derived from the on-the-fly measure results, until all predefined properties are proven. Additionally, the low reusability of existing verifiers and interpreters motivates us to develop Reinfier, a general and fundamental tool within Reintrainer for DRL verification and interpretation. Reinfier features breakpoints searching and verification-driven interpretation, associated with a concise constraint-encoding language DRLP. Evaluations demonstrate that Reintrainer outperforms the state-of-the-art on six public benchmarks in both performance and property guarantees. Our framework can be accessed at https://github.com/Kurayuri/Reinfier.

Fei Wang, Xiaofeng Gao, Guihai Chen et al.

Importance sampling (IS) as an elegant and efficient variance reduction (VR) technique for the acceleration of stochastic optimization problems has attracted many researches recently. Unlike commonly adopted stochastic uniform sampling in stochastic optimizations, IS-integrated algorithms sample training data at each iteration with respect to a weighted sampling probability distribution $P$, which is constructed according to the precomputed importance factors. Previous experimental results show that IS has achieved remarkable progresses in the acceleration of training convergence. Unfortunately, the calculation of the sampling probability distribution $P$ causes a major limitation of IS: it requires the input data to be well-structured, i.e., the feature vector is properly defined. Consequently, recurrent neural networks (RNN) as a popular learning algorithm is not able to enjoy the benefits of IS due to the fact that its raw input data, i.e., the training sequences, are often unstructured which makes calculation of $P$ impossible. In considering of the the popularity of RNN-based learning applications and their relative long training time, we are interested in accelerating them through IS. This paper propose a novel Fast-Importance-Mining algorithm to calculate the importance factor for unstructured data which makes the application of IS in RNN-based applications possible. Our experimental evaluation on popular open-source RNN-based learning applications validate the effectiveness of IS in improving the convergence rate of RNNs.

Chen Gong, Zhenzhe Zheng, Yiliu Chen et al.

Machine learning models are widely integrated into modern mobile apps to analyze user behaviors and deliver personalized services. Ensuring low-latency on-device model execution is critical for maintaining high-quality user experiences. While prior research has primarily focused on accelerating model inference with given input features, we identify an overlooked bottleneck in real-world on-device model execution pipelines: extracting input features from raw application logs. In this work, we explore a new direction of feature extraction optimization by analyzing and eliminating redundant extraction operations across different model features and consecutive model inferences. We then introduce AutoFeature, an automated feature extraction engine designed to accelerate on-device feature extraction process without compromising model inference accuracy. AutoFeature comprises three core designs: (1) graph abstraction to formulate the extraction workflows of different input features as one directed acyclic graph, (2) graph optimization to identify and fuse redundant operation nodes across different features within the graph; (3) efficient caching to minimize operations on overlapping raw data between consecutive model inferences. We implement a system prototype of AutoFeature and integrate it into five industrial mobile services spanning search, video and e-commerce domains. Online evaluations show that AutoFeature reduces end-to-end on-device model execution latency by 1.33x-3.93x during daytime and 1.43x-4.53x at night.

Yuhang Xu, Kaibin Tian, Yang Tian et al.

Reinforcement Learning (RL) has become a cornerstone for improving the performance of Large Language Models (LLMs). However, its rollout phase constitutes a significant efficiency bottleneck, mainly arising from the long-tail bubbles across data parallel ranks, particularly in long-context scenarios where faster GPUs remain idle while waiting for stragglers. Existing solutions, such as partial rollout or asynchronous RL, mitigate these bubbles by compromising the algorithm's strict synchronous nature. Instead, we propose BubbleSpec, a novel framework that accelerates RL rollouts while strictly keeping the mathematical exactness. Instead of attempting to eliminate bubbles, BubbleSpec exploits them. We exploit the idle time windows of faster ranks to pre-generate rollout results for subsequent steps, serving as drafts for speculative decoding. Unlike prior speculative methods that rely on historical epoch similarity and warm-ups, BubbleSpec is agnostic to dataset size and provides immediate acceleration from the onset of training. Extensive evaluations demonstrate that BubbleSpec reduces decoding steps by 50% and increases rollout throughput by up to 1.8x. Critically, BubbleSpec is seamlessly compatible with various RL frameworks and strategies as it sustains the strict synchronous property of RL algorithms.

Chengying Huan, Ziheng Meng, Yongchao Liu et al.

Graph Chain-of-Thought (Graph-CoT) enables large language models (LLMs) to perform step-by-step reasoning over graph-structured knowledge, but existing pipelines suffer from low accuracy, excessive token usage, high latency, and low throughput due to single-agent monolithic prompts, repeated context re-encoding, and inefficient serving execution. We present GLM, the first multi-agent Graph-CoT system co-designed with an optimized LLM serving architecture. GLM decomposes reasoning into specialized agents for classification, reasoning, action generation, and graph retrieval, enabling branching and selective context sharing to reduce prompt length and reasoning iterations while preserving reasoning quality, thereby improving accuracy and reducing overall token consumption. To scale inference, we introduce a Graph-CoT-aware LLM inference mechanism with graph-specific KV-cache management, priority-based eviction, and pipelined execution to improve serving efficiency. Experiments demonstrate that GLM improves answer accuracy by up to 38%, reduces token cost by up to 95.7%, lowers inference latency by 90.3%, and achieves up to 15.1x higher throughput compared to state-of-the-art Graph-CoT baselines, enabling efficient adoption for complex real-world reasoning at scale.

Zhibin Wang, Shipeng Li, Yuhang Zhou et al.

User experience is a critical factor Large Language Model (LLM) serving systems must consider, where service level objectives (SLOs) considering the experience of individual requests and system level metrics (SLMs) considering the overall system performance are two key performance measures. However, we observe two notable issues in existing metrics: 1) manually delaying the delivery of some tokens can improve SLOs, and 2) actively abandoning requests that do not meet SLOs can improve SLMs, both of which are counterintuitive. In this paper, we revisit SLOs and SLMs in LLM serving, and propose a new SLO that aligns with user experience. Based on the SLO, we propose a comprehensive metric framework called smooth goodput, which integrates SLOs and SLMs to reflect the nature of user experience in LLM serving. Through this unified framework, we reassess the performance of different LLM serving systems under multiple workloads. Evaluation results show that our metric framework provides a more comprehensive view of token delivery and request processing, and effectively captures the optimal point of user experience and system performance with different serving strategies.

Chen Gong, Zhenzhe Zheng, Fan Wu et al.

In modern mobile applications, users frequently encounter various new contexts, necessitating on-device continual learning (CL) to ensure consistent model performance. While existing research predominantly focused on developing lightweight CL frameworks, we identify that data scarcity is a critical bottleneck for on-device CL. In this work, we explore the potential of leveraging abundant cloud-side data to enrich scarce on-device data, and propose a private, efficient and effective data enrichment framework Delta. Specifically, Delta first introduces a directory dataset to decompose the data enrichment problem into device-side and cloud-side sub-problems without sharing sensitive data. Next, Delta proposes a soft data matching strategy to effectively solve the device-side sub-problem with sparse user data, and an optimal data sampling scheme for cloud server to retrieve the most suitable dataset for enrichment with low computational complexity. Further, Delta refines the data sampling scheme by jointly considering the impact of enriched data on both new and past contexts, mitigating the catastrophic forgetting issue from a new aspect. Comprehensive experiments across four typical mobile computing tasks with varied data modalities demonstrate that Delta could enhance the overall model accuracy by an average of 15.1%, 12.4%, 1.1% and 5.6% for visual, IMU, audio and textual tasks compared with few-shot CL, and consistently reduce the communication costs by over 90% compared to federated CL.

Zhibin Wang, Rui Ning, Chao Fang et al.

Prefix-sharing among multiple prompts presents opportunities to combine the operations of the shared prefix, while attention computation in the decode stage, which becomes a critical bottleneck with increasing context lengths, is a memory-intensive process requiring heavy memory access on the key-value (KV) cache of the prefixes. Therefore, in this paper, we explore the potential of prefix-sharing in the attention computation of the decode stage. However, the tree structure of the prefix-sharing mechanism presents significant challenges for attention computation in efficiently processing shared KV cache access patterns while managing complex dependencies and balancing irregular workloads. To address the above challenges, we propose a dedicated attention kernel to combine the memory access of shared prefixes in the decoding stage, namely FlashForge. FlashForge delivers two key innovations: a novel shared-prefix attention kernel that optimizes memory hierarchy and exploits both intra-block and inter-block parallelism, and a comprehensive workload balancing mechanism that efficiently estimates cost, divides tasks, and schedules execution. Experimental results show that FlashForge achieves an average 1.9x speedup and 120.9x memory access reduction compared to the state-of-the-art FlashDecoding kernel regarding attention computation in the decode stage and 3.8x end-to-end time per output token compared to the vLLM.

Hang Zeng, Xiangyu Liu, Yong Hu et al.

Users interacting with large language models (LLMs) under their real identifiers often unknowingly risk disclosing private information. Automatically notifying users whether their queries leak privacy and which phrases leak what private information has therefore become a practical need. Existing privacy detection methods, however, were designed for different objectives and application domains, typically tagging personally identifiable information (PII) in anonymous content, which is insufficient in real-name interaction scenarios with LLMs. In this work, to support the development and evaluation of privacy detection models for LLM interactions that are deployable on local user devices, we construct a large-scale multilingual dataset with 249K user queries and 154K annotated privacy phrases. In particular, we build an automated privacy annotation pipeline with strong LLMs to automatically extract privacy phrases from dialogue datasets and annotate leaked information. We also design evaluation metrics at the levels of privacy leakage, extracted privacy phrase, and privacy information. We further establish baseline methods using light-weight LLMs with both tuning-free and tuning-based methods, and report a comprehensive evaluation of their performance. Evaluation results reveal a gap between current performance and the requirements of real-world LLM applications, motivating future research into more effective local privacy detection methods grounded in our dataset.

Shiju Zhao, Junhao Hu, Rongxiao Huang et al.

The context caching technique is employed to accelerate the Multimodal Large Language Model (MLLM) inference by prevailing serving platforms currently. However, this approach merely reuses the Key-Value (KV) cache of the initial sequence of prompt, resulting in full KV cache recomputation even if the prefix differs slightly. This becomes particularly inefficient in the context of interleaved text and images, as well as multimodal retrieval-augmented generation. This paper proposes position-independent caching as a more effective approach for multimodal information management. We have designed and implemented a caching system, named MPIC, to address both system-level and algorithm-level challenges. MPIC stores the KV cache on local disks when receiving multimodal data, and calculates and loads the KV cache in parallel during inference. To mitigate accuracy degradation, we have incorporated the integrated reuse and recompute mechanism within the system. The experimental results demonstrate that MPIC can achieve up to 54\% reduction in response time and 2$\times$ improvement in throughput compared to existing context caching systems, while maintaining negligible or no accuracy loss.

Yucheng Ding, Yangwenjian Tan, Xiangyu Liu et al.

In many practical natural language applications, user data are highly sensitive, requiring anonymous uploads of text data from mobile devices to the cloud without user identifiers. However, the absence of user identifiers restricts the ability of cloud-based language models to provide personalized services, which are essential for catering to diverse user needs. The trivial method of replacing an explicit user identifier with a static user embedding as model input still compromises data anonymization. In this work, we propose to let each mobile device maintain a user-specific distribution to dynamically generate user embeddings, thereby breaking the one-to-one mapping between an embedding and a specific user. We further theoretically demonstrate that to prevent the cloud from tracking users via uploaded embeddings, the local distributions of different users should either be derived from a linearly dependent space to avoid identifiability or be close to each other to prevent accurate attribution. Evaluation on both public and industrial datasets using different language models reveals a remarkable improvement in accuracy from incorporating anonymous user embeddings, while preserving real-time inference requirement.

Zewei Xin, Qinya Li, Chaoyue Niu et al.

Large text-to-image models demonstrate impressive generation capabilities; however, their substantial size necessitates expensive cloud servers for deployment. Conversely, light-weight models can be deployed on edge devices at lower cost but often with inferior generation quality for complex user prompts. To strike a balance between performance and cost, we propose a routing framework, called RouteT2I, which dynamically selects either the large cloud model or the light-weight edge model for each user prompt. Since generated image quality is challenging to measure and compare directly, RouteT2I establishes multi-dimensional quality metrics, particularly, by evaluating the similarity between the generated images and both positive and negative texts that describe each specific quality metric. RouteT2I then predicts the expected quality of the generated images by identifying key tokens in the prompt and comparing their impact on the quality. RouteT2I further introduces the Pareto relative superiority to compare the multi-metric quality of the generated images. Based on this comparison and predefined cost constraints, RouteT2I allocates prompts to either the edge or the cloud. Evaluation reveals that RouteT2I significantly reduces the number of requesting large cloud model while maintaining high-quality image generation.

Liang Mi, Weijun Wang, Wenming Tu et al.

Large Multimodal Models (LMMs) have shown significant progress in various complex vision tasks with the solid linguistic and reasoning capacity inherited from large language models (LMMs). Low-rank adaptation (LoRA) offers a promising method to integrate external knowledge into LMMs, compensating for their limitations on domain-specific tasks. However, the existing LoRA model serving is excessively computationally expensive and causes extremely high latency. In this paper, we present an end-to-end solution that empowers diverse vision tasks and enriches vision applications with LoRA LMMs. Our system, VaLoRA, enables accurate and efficient vision tasks by 1) an accuracy-aware LoRA adapter generation approach that generates LoRA adapters rich in domain-specific knowledge to meet application-specific accuracy requirements, 2) an adaptive-tiling LoRA adapters batching operator that efficiently computes concurrent heterogeneous LoRA adapters, and 3) a flexible LoRA adapter orchestration mechanism that manages application requests and LoRA adapters to achieve the lowest average response latency. We prototype VaLoRA on five popular vision tasks on three LMMs. Experiment results reveal that VaLoRA improves 24-62% of the accuracy compared to the original LMMs and reduces 20-89% of the latency compared to the state-of-the-art LoRA model serving systems.

Chaoyue Niu, Yucheng Ding, Junhui Lu et al.

The conventional cloud-based large model learning framework is increasingly constrained by latency, cost, personalization, and privacy concerns. In this survey, we explore an emerging paradigm: collaborative learning between on-device small model and cloud-based large model, which promises low-latency, cost-efficient, and personalized intelligent services while preserving user privacy. We provide a comprehensive review across hardware, system, algorithm, and application layers. At each layer, we summarize key problems and recent advances from both academia and industry. In particular, we categorize collaboration algorithms into data-based, feature-based, and parameter-based frameworks. We also review publicly available datasets and evaluation metrics with user-level or device-level consideration tailored to collaborative learning settings. We further highlight real-world deployments, ranging from recommender systems and mobile livestreaming to personal intelligent assistants. We finally point out open research directions to guide future development in this rapidly evolving field.

Bonan Zhang, Zhongqi Chen, Bowen Song et al.

Reinforcement learning (RL) has become a standard paradigm for refining large language models (LLMs) beyond pre-training and instruction tuning. A prominent line of work is RL with verifiable rewards (RLVR), which leverages automatically verifiable outcomes (e.g., correctness or executability) to generate reward signals. While efficient, this framework faces two key limitations: First, its binary feedback is too sparse to capture the quality of the reasoning process. Second, its coarse-grained rewards potentially lead to vanishing gradients. Inspired by observations from human learning, we introduce a RL technique that integrates verifiable outcomes with the model's own confidence estimates. This joint design enriches the reward signal, providing finer-grained feedback and implicitly supervising the reasoning process. Experimental results demonstrate that our proposed method enhances RL performance across multiple datasets and reduces token consumption during inference, while incurring negligible additional training cost. Moreover, it can be used as a plug-in module to enhance other state-of-the-art RL methods.

Maozhe Zhao, Shengzhong Liu, Fan Wu et al.

Mobile video analysis systems often encounter various deploying environments, where environment shifts present greater demands for responsiveness in adaptations of deployed "expert DNN models". Existing model adaptation frameworks primarily operate in a cloud-centric way, exhibiting degraded performance during adaptation and delayed reactions to environment shifts. Instead, this paper proposes MOCHA, a novel framework optimizing the responsiveness of continuous model adaptation through hierarchical collaborations between mobile and cloud resources. Specifically, MOCHA (1) reduces adaptation response delays by performing on-device model reuse and fast fine-tuning before requesting cloud model retrieval and end-to-end retraining; (2) accelerates history expert model retrieval by organizing them into a structured taxonomy utilizing domain semantics analyzed by a cloud foundation model as indices; (3) enables efficient local model reuse by maintaining onboard expert model caches for frequent scenes, which proactively prefetch model weights from the cloud model database. Extensive evaluations with real-world videos on three DNN tasks show MOCHA improves the model accuracy during adaptation by up to 6.8% while saving the response delay and retraining time by up to 35.5x and 3.0x respectively.

Hongxuan Zhang, Yao Zhao, Jiaqi Zheng et al.

The emergence of long-context text applications utilizing large language models (LLMs) has presented significant scalability challenges, particularly in memory footprint. The linear growth of the Key-Value (KV) cache responsible for storing attention keys and values to minimize redundant computations can lead to substantial increases in memory consumption, potentially causing models to fail to serve with limited memory resources. To address this issue, we propose a novel approach called Cache Sparse Representation (CSR), which converts the KV cache by transforming the dense Key-Value cache tensor into sparse indexes and weights, offering a more memory-efficient representation during LLM inference. Furthermore, we introduce NeuralDict, a novel neural network-based method for automatically generating the dictionary used in our sparse representation. Our extensive experiments demonstrate that CSR achieves performance comparable to state-of-the-art KV cache quantization algorithms while maintaining robust functionality in memory-constrained environments.

Shangyu Liu, Zhenzhe Zheng, Xiaoyao Huang et al.

Small language models (SLMs) support efficient deployments on resource-constrained edge devices, but their limited capacity compromises inference performance. Retrieval-augmented generation (RAG) is a promising solution to enhance model performance by integrating external databases, without requiring intensive on-device model retraining. However, large-scale public databases and user-specific private contextual documents are typically located on the cloud and the device separately, while existing RAG implementations are primarily centralized. To bridge this gap, we propose DRAGON, a distributed RAG framework to enhance on-device SLMs through both general and personal knowledge without the risk of leaking document privacy. Specifically, DRAGON decomposes multi-document RAG into multiple parallel token generation processes performed independently and locally on the cloud and the device, and employs a newly designed Speculative Aggregation, a dual-side speculative algorithm to avoid frequent output synchronization between the cloud and device. A new scheduling algorithm is further introduced to identify the optimal aggregation side based on real-time network conditions. Evaluations on real-world hardware testbed demonstrate a significant performance improvement of DRAGON-up to 1.9x greater gains over standalone SLM compared to the centralized RAG, substantial reduction in per-token latency, and negligible Time to First Token (TTFT) overhead.

Jiaming Liu, Linghe Kong, Guihai Chen

Segment anything model (SAM) has shown impressive general-purpose segmentation performance on natural images, but its performance on camouflaged object detection (COD) is unsatisfactory. In this paper, we propose SAM-COD that performs camouflaged object detection for RGB-D inputs. While keeping the SAM architecture intact, dual stream adapters are expanded on the image encoder to learn potential complementary information from RGB images and depth images, and fine-tune the mask decoder and its depth replica to perform dual-stream mask prediction. In practice, the dual stream adapters are embedded into the attention block of the image encoder in a parallel manner to facilitate the refinement and correction of the two types of image embeddings. To mitigate channel discrepancies arising from dual stream embeddings that do not directly interact with each other, we augment the association of dual stream embeddings using bidirectional knowledge distillation including a model distiller and a modal distiller. In addition, to predict the masks for RGB and depth attention maps, we hybridize the two types of image embeddings which are jointly learned with the prompt embeddings to update the initial prompt, and then feed them into the mask decoders to synchronize the consistency of image embeddings and prompt embeddings. Experimental results on four COD benchmarks show that our SAM-COD achieves excellent detection performance gains over SAM and achieves state-of-the-art results with a given fine-tuning paradigm.

Hang Zeng, Chaoyue Niu, Fan Wu et al.

Existing work on large language model (LLM) personalization assigned different responding roles to LLMs, but overlooked the diversity of queriers. In this work, we propose a new form of querier-aware LLM personalization, generating different responses even for the same query from different queriers. We design a dual-tower model architecture with a cross-querier general encoder and a querier-specific encoder. We further apply contrastive learning with multi-view augmentation, pulling close the dialogue representations of the same querier, while pulling apart those of different queriers. To mitigate the impact of query diversity on querier-contrastive learning, we cluster the dialogues based on query similarity and restrict the scope of contrastive learning within each cluster. To address the lack of datasets designed for querier-aware personalization, we also build a multi-querier dataset from English and Chinese scripts, as well as WeChat records, called MQDialog, containing 173 queriers and 12 responders. Extensive evaluations demonstrate that our design significantly improves the quality of personalized response generation, achieving relative improvement of 8.4% to 48.7% in ROUGE-L scores and winning rates ranging from 54% to 82% compared with various baseline methods.

Chenyu Mu, Guihai Chen, Xun Yang et al.

Medical image segmentation is crucial for clinical applications, but it is frequently disrupted by noisy annotations and ambiguous anatomical boundaries, limiting its application in real-world scenarios. Existing methods often directly adapt noisy label learning techniques designed for instance classification, overlooking the pixel-wise heterogeneity in medical segmentation with its spatially and anatomically varying difficulties. Consequently, global assumptions or simple confidence metrics fail to address these local variations, leaving boundary ambiguities unresolved. To address this issue, we propose MetaDCSeg, a robust framework that dynamically learns optimal pixel-wise weights to suppress the influence of noisy labels while preserving reliable annotations. By explicitly modeling boundary uncertainty through a Dynamic Center Distance (DCD) mechanism, our approach utilizes weighted feature distances for foreground, background, and boundary centers, directing the model's attention toward hard-to-segment pixels near ambiguous boundaries. This strategy enables more precise handling of structural boundaries, which are often overlooked by existing methods, and significantly enhances segmentation performance. Extensive experiments across four benchmark datasets with varying noise levels demonstrate that MetaDCSeg outperforms existing state-of-the-art methods.