Ju Ren

Shizuo Tian, Hao Wen, Yuxuan Chen et al.

The rapid development of mobile GUI agents has stimulated growing research interest in long-horizon task automation. However, building agents for these tasks faces a critical bottleneck: the reliance on ever-expanding interaction history incurs substantial context overhead. Existing context management and compression techniques often fail to preserve vital semantic information, leading to degraded task performance. We propose AgentProg, a program-guided approach for agent context management that reframes the interaction history as a program with variables and control flow. By organizing information according to the structure of program, this structure provides a principled mechanism to determine which information should be retained and which can be discarded. We further integrate a global belief state mechanism inspired by Belief MDP framework to handle partial observability and adapt to unexpected environmental changes. Experiments on AndroidWorld and our extended long-horizon task suite demonstrate that AgentProg has achieved the state-of-the-art success rates on these benchmarks. More importantly, it maintains robust performance on long-horizon tasks while baseline methods experience catastrophic degradation. Our system is open-sourced at https://github.com/MobileLLM/AgentProg.

Yizhe Zhao, Yongjian Fu, Zihao Feng et al.

Mobile advertising dominates app monetization but introduces risks ranging from intrusive user experience to malware delivery. Existing detection methods rely either on static analysis, which misses runtime behaviors, or on heuristic UI exploration, which struggles with sparse and obfuscated ads. In this paper, we present MANA, the first agentic multimodal reasoning framework for mobile ad detection. MANA integrates static, visual, temporal, and experiential signals into a reasoning-guided navigation strategy that determines not only how to traverse interfaces but also where to focus, enabling efficient and robust exploration. We implement and evaluate MANA on commercial smartphones over 200 apps, achieving state-of-the-art accuracy and efficiency. Compared to baselines, it improves detection accuracy by 30.5%-56.3% and reduces exploration steps by 29.7%-63.3%. Case studies further demonstrate its ability to uncover obfuscated and malicious ads, underscoring its practicality for mobile ad auditing and its potential for broader runtime UI analysis (e.g., permission abuse). Code and dataset are available at https://github.com/MANA-2026/MANA.

Sheng Yue, Guanbo Wang, Wei Shao et al.

This work aims to tackle a major challenge in offline Inverse Reinforcement Learning (IRL), namely the reward extrapolation error, where the learned reward function may fail to explain the task correctly and misguide the agent in unseen environments due to the intrinsic covariate shift. Leveraging both expert data and lower-quality diverse data, we devise a principled algorithm (namely CLARE) that solves offline IRL efficiently via integrating "conservatism" into a learned reward function and utilizing an estimated dynamics model. Our theoretical analysis provides an upper bound on the return gap between the learned policy and the expert policy, based on which we characterize the impact of covariate shift by examining subtle two-tier tradeoffs between the exploitation (on both expert and diverse data) and exploration (on the estimated dynamics model). We show that CLARE can provably alleviate the reward extrapolation error by striking the right exploitation-exploration balance therein. Extensive experiments corroborate the significant performance gains of CLARE over existing state-of-the-art algorithms on MuJoCo continuous control tasks (especially with a small offline dataset), and the learned reward is highly instructive for further learning.

Liekang Zeng, Shengyuan Ye, Xu Chen et al.

Recent years have witnessed a thriving growth of computing facilities connected at the network edge, cultivating edge networks as a fundamental infrastructure for supporting miscellaneous intelligent services.Meanwhile, Artificial Intelligence (AI) frontiers have extrapolated to the graph domain and promoted Graph Intelligence (GI). Given the inherent relation between graphs and networks, the interdiscipline of graph learning and edge networks, i.e., Edge GI or EGI, has revealed a novel interplay between them -- GI aids in optimizing edge networks, while edge networks facilitate GI model deployment. Driven by this delicate closed-loop, EGI is recognized as a promising solution to fully unleash the potential of edge computing power and is garnering growing attention. Nevertheless, research on EGI remains nascent, and there is a soaring demand within both the communications and AI communities for a dedicated venue to share recent advancements. To this end, this paper promotes the concept of EGI, explores its scope and core principles, and conducts a comprehensive survey concerning recent research efforts on this emerging field. Specifically, this paper introduces and discusses: 1) fundamentals of edge computing and graph learning,2) emerging techniques centering on the closed loop between graph intelligence and edge networks, and 3) open challenges and research opportunities of future EGI. By bridging the gap across communication, networking, and graph learning areas, we believe that this survey can garner increased attention, foster meaningful discussions, and inspire further research ideas in EGI.

Fucheng Jia, Shiqi Jiang, Ting Cao et al.

Web is increasingly becoming the primary platform to deliver AI services onto edge devices, making in-browser deep learning (DL) inference more prominent. Nevertheless, the heterogeneity of edge devices, combined with the underdeveloped state of Web hardware acceleration practices, hinders current in-browser inference from achieving its full performance potential on target devices. To address this issue, this paper presents the pioneering inbrowser inference system, nnJIT, which enables just-in-time (JIT) auto-generation of optimized computing kernels for edge devices. nnJIT is built upon two novel techniques that significantly reduce kernel search and compilation overhead while improving performance firmly: Tensor-Web Compiling Co-Design lowers compiling costs by around 100X through eliminating redundant and ineffective compiling passes; Web-Specific Lite Kernel Optimization Space reduces kernel tuning costs by focusing on Web programming requirements and efficient device resource utilization, pruning the optimization space from millions to only dozens. nnJIT is evaluated for modern models, e.g., BART, T5, and Llama 2, on a range of edge devices including laptops and smartphones using different browsers and hardware from ARM, Intel, AMD and Nvidia. The results show that nnJIT can achieve up to 8.2X faster within 30 seconds compared to the existing baselines.

Xingyuan Hua, Sheng Yue, Ju Ren

Recent advancements in agentic test-time scaling allow models to gather environmental feedback before committing to final actions. A key limitation of existing methods is that they typically employ undifferentiated exploration strategies, lacking the ability to adaptively distinguish when exploration is truly required. In this paper, we propose an exploration-aware reinforcement learning framework that enables LLM agents to adaptively explore only when uncertainty is high. Our method introduces a fine-grained reward function via variational inference that explicitly evaluates exploratory actions by estimating their potential to improve future decision-making, together with an exploration-aware grouping mechanism that separates exploratory actions from task-completion actions during optimization. By targeting informational gaps, this design allows agents to explore selectively and transition to execution as soon as the task context is clear. Empirically, we demonstrate that our approach achieves consistent improvements across a range of challenging text-based and GUI-based agent benchmarks. Code is available at \url{https://github.com/HansenHua/EAPO-ICML26} and models are available at https://huggingface.co/hansenhua/EAPO-ICML26.

Tuowei Wang, Fengzu Li, Yanfan Sun et al.

Large language models (LLMs) are increasingly needed for interactive mobile applications, but high-quality models exceed the limited DRAM available on smartphones. Flash storage can hold larger models, yet flash-backed inference is slow because autoregressive decoding repeatedly invokes the target model and incurs costly I/O. We observe that speculative decoding is a natural fit for this setting: a small draft model can remain in DRAM, while a larger flash-resident target model verifies multiple candidate tokens per invocation. However, existing methods assume server-class accelerators and fail to account for prolonged I/O latency, limited computation parallelism, and irregular speculation execution. We present Lever, an end-to-end system for efficient flash-backed LLM inference on smartphones. Lever jointly optimizes the three stages of speculative decoding under mobile constraints. For drafting, it builds token trees using an I/O- and compute-aware gain-cost objective. For verification, it prunes low-value branches through early-exit prediction to reduce target-model computation. For execution, it maps speculation efficiently across mobile CPU-NPU hardware to improve utilization. Comprehensive evaluations show that Lever reduces inference latency by an average of 2.93x over baseline flash-offloaded inference and 1.50x over conventional speculative decoding, narrowing the latency gap between flash-backed and memory-resident LLM inference.

Nan Qiao, Shuning Wang, Sijing Duan et al.

Photovoltaic (PV) power forecasting in edge-enabled grids requires balancing forecasting accuracy, robustness under weather-driven distribution shifts, and strict latency constraints. Existing models work well under normal conditions but often struggle with rare ramp events and unexpected weather changes. Relying solely on cloud-based large models often leads to significant communication delays, which can hinder timely and efficient forecasting in practical grid environments. To address these issues, we propose a condition-adaptive cloud-edge collaborative framework *CAPE* for PV forecasting. *CAPE* consists of three main modules: a site-specific expert model for routine predictions, a lightweight edge-side model for enhanced local inference, and a cloud-based large retrieval model that provides relevant historical cases when needed. These modules are coordinated by a screening module that evaluates uncertainty, out-of-distribution risk, weather mutations, and model disagreement. Furthermore, we employ a Lyapunov-guided routing strategy to dynamically determine when to escalate inference to more powerful models under long-term system constraints. The final forecast is produced through adaptive fusion of the selected model outputs. Experiments on two real-world PV datasets demonstrate that *CAPE* achieves superior performance in terms of forecasting accuracy, robustness, routing quality, and system efficiency.

Xingyuan Hua, Sheng Yue, Xinyi Li et al.

Multi-Agent Discussion (MAD) has garnered increasing attention very recently, where multiple LLM instances collaboratively solve problems via structured discussion. However, we find that current MAD methods easily suffer from discussion inconsistency, LLMs fail to reach a coherent solution, due to the misalignment between their individual contexts.In this paper, we introduce a multi-LLM context learning method (M2CL) that learns a context generator for each agent, capable of dynamically generating context instructions per discussion round via automatic information organization and refinement. Specifically, inspired by our theoretical insights on the context instruction, M2CL train the generators to control context coherence and output discrepancies via a carefully crafted self-adaptive mechanism.It enables LLMs to avoid premature convergence on majority noise and progressively reach the correct consensus. We evaluate M2CL on challenging tasks, including academic reasoning, embodied tasks, and mobile control. The results show that the performance of M2CL significantly surpasses existing methods by 20%--50%, while enjoying favorable transferability and computational efficiency.

Tuowei Wang, He Zhou, Chengru Song et al.

Large vision-language models (VLMs) are enabling interactive video reasoning, giving rise to streaming long-video understanding. In this setting, frames arrive continuously, while the system preserves long-term context and generates responses under strict latency constraints. A central challenge is KVCache management: as video streams grow, KVCache expands rapidly, increasing computation and memory overhead. Existing retrieval-based approaches exploit attention sparsity and offload inactive KVCache from GPU to CPU memory, but their token-level design causes high management overhead and fragmented data movement. We present Mosaic, the first cluster-driven VLM inference system for streaming long-video understanding. Our key insight is that VLM KVCache exhibits an implicit cross-modal clustering structure: retrieved KV states form groups jointly shaped by visual coherence and semantic relevance. Based on this observation, Mosaic uses cross-modal clusters as the basic unit of KVCache organization, maintenance, and retrieval. Evaluations show that Mosaic outperforms state-of-the-art baselines, achieving up to 1.38x speedup.

Zechen Li, Lanqing Yang, Yiheng Bian et al.

This paper presents an innovative frequency-embedded 3D Gaussian splatting (3DGS) algorithm for wideband radio-frequency (RF) radiance field modeling, offering an advancement over the existing works limited to single-frequency modeling. Grounded in fundamental physics, we uncover the complex relationship between EM wave propagation behaviors and RF frequencies. Inspired by this, we design an EM feature network with attenuation and radiance modules to learn the complex relationships between RF frequencies and the key properties of each 3D Gaussian, specifically the attenuation factor and RF signal intensity. By training the frequency-embedded 3DGS model, we can efficiently reconstruct RF radiance fields at arbitrary unknown frequencies within a given 3D environment. Finally, we propose a large-scale power angular spectrum (PAS) dataset containing 50000 samples ranging from 1 to 100 GHz in 6 indoor environments, and conduct extensive experiments to verify the effectiveness of our method. Our approach achieves an average Structural Similarity Index Measure (SSIM) up to 0.72, and a significant improvement up to 17.8% compared to the current state-of-the-art (SOTA) methods trained on individual test frequencies. Additionally, our method achieves an SSIM of 0.70 without prior training on these frequencies, which represents only a 2.8% performance drop compared to models trained with full PAS data. This demonstrates our model's capability to estimate PAS at unknown frequencies. For related code and datasets, please refer to https://github.com/sim-2-real/Wideband3DGS.

Tuowei Wang, Liyun Chu, Ruwen Fan et al.

The key-value (KV) cache has become the dominant contributor to memory consumption in large language model (LLM) inference. Although offloading KVCache from GPU high-bandwidth memory (HBM) to CPU DRAM alleviates device memory pressure, DRAM remains capacity-limited and costly for large, persistent workloads. Solid-state drives (SSDs) provide a cost-effective alternative, but naive SSD-based paging is fundamentally bandwidth-bound due to limited PCIe throughput and per-device bandwidth constraints. In this paper, we observe that KVCache activations in real-world workloads exhibit strong and stable correlations. We term this phenomenon KVCache Co-Activation, where accessing a KV entry is often accompanied by a stable and recurring set of other KV entries. Leveraging this property, we present Swarm, an SSD-based KVCache offloading system that converts bandwidth-bound single-device access into parallel I/O across multiple SSDs. Specifically, Swarm clusters co-activated KV entries offline and distributes the resulting clusters across SSDs using graph-based placement with selective replication to maximize parallel I/O bandwidth. At runtime, Swarm performs load-balanced cluster retrieval and dynamically adapts clustering and caching decisions to sustain high bandwidth utilization under evolving access patterns. Evaluations show that Swarm reduces I/O time by 2.41x and improves effective bandwidth utilization by 2.72x.

Zerui Qin, Sheng Yue, Xingyuan Hua et al.

Modern GUI agents typically rely on a model-centric and step-wise interaction paradigm, where LLMs must re-interpret the UI and re-decide actions at every screen, which is fragile in long-horizon tasks. In this paper, we propose Executable Agentic Memory (EAM), a structured Knowledge Graph (KG) that shifts GUI planning from free-form generation to a robust retrieval-and-execution process. Our approach includes a sample-efficient memory construction pipeline using state-aware DFS and action-group mining to compress multi-step routines. To ensure efficient planning, we introduce a value-guided graph search where a lightweight Q-function model steers Monte Carlo Tree Search (MCTS) over the KG. We theoretically establish bias-consistency for the Q-model and derive sample complexity bounds for path recovery. Empirically, EAM outperforms state-of-the-art baselines like UI-TARS-7B by up to $19.6\%$ on AndroidWorld, while reducing token costs $6\times$ relative to GPT-4o. With a $2.8$s average latency, EAM enables reliable, quick, and long-horizon GUI automation.

Zhengyi Zhong, Wenzheng Jiang, Weidong Bao et al.

Conventional federated learning (FL) assumes a closed world with a fixed total number of clients. In contrast, new clients continuously join the FL process in real-world scenarios, introducing new knowledge. This raises two critical demands: detecting new knowledge, i.e., knowledge discovery, and integrating it into the global model, i.e., knowledge adaptation. Existing research focuses on coarse-grained knowledge discovery, and often sacrifices source domain performance and adaptation efficiency. To this end, we propose a fine-grained federated domain adaptation approach in open set (Gains). Gains splits the model into an encoder and a classifier, empirically revealing features extracted by the encoder are sensitive to domain shifts while classifier parameters are sensitive to class increments. Based on this, we develop fine-grained knowledge discovery and contribution-driven aggregation techniques to identify and incorporate new knowledge. Additionally, an anti-forgetting mechanism is designed to preserve source domain performance, ensuring balanced adaptation. Experimental results on multi-domain datasets across three typical data-shift scenarios demonstrate that Gains significantly outperforms other baselines in performance for both source-domain and target-domain clients. Code is available at: https://github.com/Zhong-Zhengyi/Gains.

Deyu Zhang, Xicheng Zhang, Jiahao Li et al.

We introduce SRDrone, a novel system designed for self-refinement task planning in industrial-grade embodied drones. SRDrone incorporates two key technical contributions: First, it employs a continuous state evaluation methodology to robustly and accurately determine task outcomes and provide explanatory feedback. This approach supersedes conventional reliance on single-frame final-state assessment for continuous, dynamic drone operations. Second, SRDrone implements a hierarchical Behavior Tree (BT) modification model. This model integrates multi-level BT plan analysis with a constrained strategy space to enable structured reflective learning from experience. Experimental results demonstrate that SRDrone achieves a 44.87% improvement in Success Rate (SR) over baseline methods. Furthermore, real-world deployment utilizing an experience base optimized through iterative self-refinement attains a 96.25% SR. By embedding adaptive task refinement capabilities within an industrial-grade BT planning framework, SRDrone effectively integrates the general reasoning intelligence of Large Language Models (LLMs) with the stringent physical execution constraints inherent to embodied drones. Code is available at https://github.com/ZXiiiC/SRDrone.

Deyu Zhang, Tingting Long, Jinrui Zhang et al.

Enabling efficient text-video retrieval on edge-end devices is critical for real-world applications. Yet, existing methods face a critical challenge in balancing accuracy and computational efficiency: uniform frame sampling methods ensure content coverage but incur prohibitive computational costs, while salient-frame sampling methods reduce overhead but suffer from query-agnostic frame selection that biases retrieval results. To address this, we propose ProCLIP, a user-centric framework that achieves state-of-the-art accuracy with significantly improved efficiency. We design a prompt-aware frame sampling strategy that dynamically guides lightweight feature extractors using textual prompts to select semantically relevant frames, overcoming the limitations of existing salient-frame sampling methods which rely on static, query-agnostic selection criteria. Moreover, we adopt a two-stage candidate pruning strategy that combines rapid coarse filtering via a lightweight module with CLIP-powered fine-grained re-ranking, enhancing retrieval efficiency while preserving accuracy. Experiments across benchmarks show ProCLIP achieves 75.3% latency reduction versus baselines while maintaining competitive accuracy, i.e., R@1=49.0 in MSR-VTT dataset. Code is available at https://github.com/tiffylong/ProCLIP.

Nan Qiao, Sheng Yue, Ju Ren et al.

Offline Federated Reinforcement Learning (FRL), a marriage of federated learning and offline reinforcement learning, has attracted increasing interest recently. Albeit with some advancement, we find that the performance of most existing offline FRL methods drops dramatically when provided with mixed-quality data, that is, the logging behaviors (offline data) are collected by policies with varying qualities across clients. To overcome this limitation, this paper introduces a new vote-based offline FRL framework, named FOVA. It exploits a \emph{vote mechanism} to identify high-return actions during local policy evaluation, alleviating the negative effect of low-quality behaviors from diverse local learning policies. Besides, building on advantage-weighted regression (AWR), we construct consistent local and global training objectives, significantly enhancing the efficiency and stability of FOVA. Further, we conduct an extensive theoretical analysis and rigorously show that the policy learned by FOVA enjoys strict policy improvement over the behavioral policy. Extensive experiments corroborate the significant performance gains of our proposed algorithm over existing baselines on widely used benchmarks.

Nan Qiao, Sheng Yue, Shuning Wang et al.

Offline reinforcement learning (RL) can fail spectacularly when bootstrapped temporal-difference (TD) updates amplify their own errors, driving the critic toward extreme and unusable Q-values. A key counterintuitive insight of this work is that collapse is not only a property of the backup rule or network architecture: optimizer dynamics themselves can directly trigger or suppress instability. From a control-theoretic viewpoint, we model offline TD learning as a feedback system and analyze Adam-based critic updates. This yields a necessary and sufficient condition for stability of the induced local update dynamics: within the regime we analyze, these dynamics are stable if and only if the spectral radius of the corresponding update operator is strictly below one. Further analysis suggests that standard Adam updates can inadvertently distort the parameter geometry, motivating explicit orthogonality constraints to prevent TD error amplification. To this end, we propose AdamO, an Adam-based optimizer with a decoupled orthogonality correction regulated by a strict task-alignment budget. We prove that this design theoretically guarantees worst-case task safety and preserves Adam's continuous-time dissipative dynamics. Empirically, AdamO is broadly compatible with diverse offline RL baselines, improving stability and returns across a broad suite of benchmarks.

Ziqing Qiao, Yongheng Deng, Jiali Zeng et al.

Large Reasoning Models (LRMs) perform strongly in complex reasoning tasks via Chain-of-Thought (CoT) prompting, but often suffer from verbose outputs, increasing computational overhead. Existing fine-tuning-based compression methods either operate post-hoc pruning, risking disruption to reasoning coherence, or rely on sampling-based selection, which fails to remove redundant content thoroughly. To address these limitations, this work begins by framing two key patterns of redundant reflection in LRMs--Confidence Deficit, wherein the model reflects on correct intermediate steps, and Termination Delay, where reflection continues after a verified, confident answer--through a confidence-guided perspective. Based on this, we introduce ConCISE (Confidence-guided Compression In Step-by-step Efficient Reasoning), a framework designed to generate concise reasoning chains, integrating Confidence Injection to boost reasoning confidence, and Early Stopping to terminate reasoning when confidence is sufficient. Extensive experiments demonstrate that compared to baseline methods, fine-tuning LRMs on ConCISE-generated data yields a better balance between compression and task performance, reducing length by up to approximately 50% under SimPO, while maintaining high task accuracy.

Yang Xu, Yunlin Tan, Cheng Zhang et al.

Embedding watermarks into models has been widely used to protect model ownership in federated learning (FL). However, existing methods are inadequate for protecting the ownership of personalized models acquired by clients in personalized FL (PFL). This is due to the aggregation of the global model in PFL, resulting in conflicts over clients' private watermarks. Moreover, malicious clients may tamper with embedded watermarks to facilitate model leakage and evade accountability. This paper presents a robust watermark embedding scheme, named RobWE, to protect the ownership of personalized models in PFL. We first decouple the watermark embedding of personalized models into two parts: head layer embedding and representation layer embedding. The head layer belongs to clients' private part without participating in model aggregation, while the representation layer is the shared part for aggregation. For representation layer embedding, we employ a watermark slice embedding operation, which avoids watermark embedding conflicts. Furthermore, we design a malicious watermark detection scheme enabling the server to verify the correctness of watermarks before aggregating local models. We conduct an exhaustive experimental evaluation of RobWE. The results demonstrate that RobWE significantly outperforms the state-of-the-art watermark embedding schemes in FL in terms of fidelity, reliability, and robustness.

Sheng Yue, Zerui Qin, Yongheng Deng et al.

Federated Learning (FL) has garnered widespread interest in recent years. However, owing to strict privacy policies or limited storage capacities of training participants such as IoT devices, its effective deployment is often impeded by the scarcity of training data in practical decentralized learning environments. In this paper, we study enhancing FL with the aid of (large) pre-trained models (PMs), that encapsulate wealthy general/domain-agnostic knowledge, to alleviate the data requirement in conducting FL from scratch. Specifically, we consider a networked FL system formed by a central server and distributed clients. First, we formulate the PM-aided personalized FL as a regularization-based federated meta-learning problem, where clients join forces to learn a meta-model with knowledge transferred from a private PM stored at the server. Then, we develop an inexact-ADMM-based algorithm, AugFL, to optimize the problem with no need to expose the PM or incur additional computational costs to local clients. Further, we establish theoretical guarantees for AugFL in terms of communication complexity, adaptation performance, and the benefit of knowledge transfer in general non-convex cases. Extensive experiments corroborate the efficacy and superiority of AugFL over existing baselines.

Tuowei Wang, Kun Li, Zixu Hao et al.

The adaptation of pre-trained large language models (LLMs) to diverse downstream tasks via fine-tuning is critical for numerous applications. However, the inefficiency of parameter-efficient fine-tuning (PEFT) techniques presents significant challenges in terms of time investments and operational costs. In this paper, we first introduce a nuanced form of sparsity, termed Shadowy Sparsity, which is distinctive in fine-tuning and has not been adequately addressed for acceleration. Under Shadowy Sparsity, we propose Long Exposure, an efficient system to accelerate PEFT for LLMs. Long Exposure comprises three key components: Shadowy-sparsity Exposer employs a prolonged sensing range to capture more sparsity details under shadowy sparsity; Sequence-oriented Predictor provides efficient yet accurate predictions to handle large sequence inputs and constantly-evolving parameters; and Dynamic-aware Operator facilitates more structured computational patterns and coalesced memory accesses, addressing dynamic sparse operations. Extensive evaluations show that Long Exposure outperforms state-of-the-arts with up to a $2.49\times$ speedup in end-to-end fine-tuning, offering promising advancements in accelerating PEFT for LLMs.

Tuowei Wang, Ruwen Fan, Minxing Huang et al.

Large Language Models (LLMs) have achieved remarkable success across various domains, yet deploying them on mobile devices remains an arduous challenge due to their extensive computational and memory demands. While lightweight LLMs have been developed to fit mobile environments, they suffer from degraded model accuracy. In contrast, sparsity-based techniques minimize DRAM usage by selectively transferring only relevant neurons to DRAM while retaining the full model in external storage, such as flash. However, such approaches are critically limited by numerous I/O operations, particularly on smartphones with severe IOPS constraints. In this paper, we propose Neuralink, a novel approach that accelerates LLM inference on smartphones by optimizing neuron placement in flash memory. Neuralink leverages the concept of Neuron Co-Activation, where neurons frequently activated together are linked to facilitate continuous read access and optimize I/O efficiency. Our approach incorporates a two-stage solution: an offline stage that reorganizes neuron placement based on co-activation patterns, and an online stage that employs tailored data access and caching strategies to align well with hardware characteristics. Evaluations conducted on a variety of smartphones and LLMs demonstrate that Neuralink achieves on average $1.49\times$ improvements in end-to-end latency compared to the state-of-the-art. As the first solution to optimize storage placement under sparsity, Neuralink explores a new optimization space at the intersection of sparsity-driven algorithm and storage-level system co-design for LLM inference.

Fucheng Jia, Zewen Wu, Shiqi Jiang et al. · microsoft-research

Large language models (LLMs) are increasingly being deployed on mobile devices, but the limited DRAM capacity constrains the deployable model size. This paper introduces ActiveFlow, the first LLM inference framework that can achieve adaptive DRAM usage for modern LLMs (not ReLU-based), enabling the scaling up of deployable model sizes. The framework is based on the novel concept of active weight DRAM-flash swapping and incorporates three novel techniques: (1) Cross-layer active weights preloading. It uses the activations from the current layer to predict the active weights of several subsequent layers, enabling computation and data loading to overlap, as well as facilitating large I/O transfers. (2) Sparsity-aware self-distillation. It adjusts the active weights to align with the dense-model output distribution, compensating for approximations introduced by contextual sparsity. (3) Active weight DRAM-flash swapping pipeline. It orchestrates the DRAM space allocation among the hot weight cache, preloaded active weights, and computation-involved weights based on available memory. Results show ActiveFlow achieves the performance-cost Pareto frontier compared to existing efficiency optimization methods.

Wangyu Xue, Chen Qian, Jiayi Wu et al.

Existing works on human-centric video understanding typically focus on analyzing specific moment or entire videos. However, many applications require higher precision at the frame level. In this work, we propose a novel task, BestShot, which aims to locate highlight frames within human-centric videos via language queries. This task demands not only a deep semantic comprehension of human actions but also precise temporal localization. To support this task, we introduce the BestShot Benchmark. %The benchmark is meticulously constructed by combining human detection and tracking, potential frame selection based on human judgment, and detailed textual descriptions crafted by human input to ensure precision. The benchmark is meticulously constructed by combining human-annotated highlight frames, detailed textual descriptions and duration labeling. These descriptions encompass three critical elements: (1) Visual content; (2) Fine-grained action; and (3) Human Pose Description. Together, these elements provide the necessary precision to identify the exact highlight frames in videos. To tackle this problem, we have collected two distinct datasets: (i) ShotGPT4o Dataset, which is algorithmically generated by GPT-4o and (ii) Image-SMPLText Dataset, a dataset with large-scale and accurate per-frame pose description leveraging PoseScript and existing pose estimation datasets. Based on these datasets, we present a strong baseline model, ShotVL, fine-tuned from InternVL, specifically for BestShot. We highlight the impressive zero-shot capabilities of our model and offer comparative analyses with existing SOTA models. ShotVL demonstrates a significant 52% improvement over InternVL on the BestShot Benchmark and a notable 57% improvement on the THUMOS14 Benchmark, all while maintaining the SOTA performance in general image classification and retrieval.

Nan Qiao, Sheng Yue, Shuning Wang et al.

A fundamental challenge in offline reinforcement learning is distributional shift. Scarce data or datasets dominated by out-of-distribution (OOD) areas exacerbate this issue. Our theoretical analysis and experiments show that the standard squared error objective induces a harmful TD cross covariance. This effect amplifies in OOD areas, biasing optimization and degrading policy learning. To counteract this mechanism, we develop two complementary strategies: partitioned buffer sampling that restricts updates to localized replay partitions, attenuates irregular covariance effects, and aligns update directions, yielding a scheme that is easy to integrate with existing implementations, namely Clustered Cross-Covariance Control for TD (C^4). We also introduce an explicit gradient-based corrective penalty that cancels the covariance induced bias within each update. We prove that buffer partitioning preserves the lower bound property of the maximization objective, and that these constraints mitigate excessive conservatism in extreme OOD areas without altering the core behavior of policy constrained offline reinforcement learning. Empirically, our method showcases higher stability and up to 30% improvement in returns over prior methods, especially with small datasets and splits that emphasize OOD areas.

Tuowei Wang, Minxing Huang, Fengzu Li et al.

As the demand for human-like reasoning, multi-turn dialogues, and long-form responses grows, large language models (LLMs) are increasingly expected to support efficient and effective long-sequence decoding. However, due to limited DRAM capacity, long-seuqence LLM decoding on smartphones is constrained by the key-value cache (KVCache), whose memory footprint increases linearly with sequence length. Retrieval-based methods mitigate DRAM pressure by offloading KVCache to flash and retrieving query-relevant entries through cluster-based indexing. Unfortunately, as decoding progresses, KVCache distribution shifts render static or local cluster updates progressively misaligned, excluding essential entries or fetching redundant ones. These issues are further exacerbated by smartphone-specific limitations in bandwidth, IOPS, and memory capacity. We propose DynaKV, the first adaptive KVCache management approach that jointly addresses accuracy and efficiency for long-sequence decoding on smartphones. DynaKV integrates three key techniques: (1) Migration-Free Cluster Adaptation, which adaptively splits clusters during retrieval without incurring additional transfers; (2) Continuity-Centric Flash Management, which co-locates correlated entries and clusters and employs a dual-head layout for efficient updates; and (3) Memory-Efficient Cache Design, which virtualizes cache space across DRAM and flash and extends replacement policies to align with cluster-level access patterns. Evaluations demonstrate that DynaKV improves retrieval accuracy and reduces end-to-end latency compared to state-of-the-art solutions, achieving average gains of $1.38\times$ in accuracy and $1.47\times$ speedups. Furthermore, the insights of DynaKV naturally extend to other long-context workloads and multi-tier memory hierarchies, underscoring its broader applicability.

Zixu Hao, Jianyu Wei, Tuowei Wang et al. · microsoft-research

Deploying Large Language Models (LLMs) on mobile devices faces the challenge of insufficient performance in smaller models and excessive resource consumption in larger ones. This paper highlights that mobile Neural Processing Units (NPUs) have underutilized computational resources, particularly their matrix multiplication units, during typical LLM inference. To leverage this wasted compute capacity, we propose applying parallel test-time scaling techniques on mobile NPUs to enhance the performance of smaller LLMs. However, this approach confronts inherent NPU challenges, including inadequate hardware support for fine-grained quantization and low efficiency in general-purpose computations. To overcome these, we introduce two key techniques: a hardware-aware tile quantization scheme that aligns group quantization with NPU memory access patterns, and efficient LUT-based replacements for complex operations such as Softmax and dequantization. We design and implement an end-to-end inference system that leverages the NPU's compute capability to support test-time scaling on Qualcomm Snapdragon platforms. Experiments show our approach brings significant speedups: up to 19.0 for mixed-precision GEMM and 2.2 for Softmax. More importantly, we demonstrate that smaller models using test-time scaling can match or exceed the accuracy of larger models, achieving a new performance-cost Pareto frontier.

Tuowei Wang, Xingyu Chen, Kun Li et al.

The escalating demand for long-context applications has intensified the necessity of extending the LLM context windows. Despite recent fine-tuning approaches successfully expanding context lengths, their high memory footprints, especially for activations, present a critical practical limitation. Current parameter-efficient fine-tuning methods prioritize reducing parameter update overhead over addressing activation memory constraints. Similarly, existing sparsity mechanisms improve computational efficiency but overlook activation memory optimization due to the phenomenon of Shadowy Activation. In this paper, we propose LeMo, the first LLM fine-tuning system that explores and exploits a new token-level sparsity mechanism inherent in long-context scenarios, termed Contextual Token Sparsity. LeMo minimizes redundant token involvement by assessing the informativeness of token embeddings while preserving model accuracy. Specifically, LeMo introduces three key techniques: (1) Token Elimination, dynamically identifying and excluding redundant tokens across varying inputs and layers. (2) Pattern Prediction, utilizing well-trained predictors to approximate token sparsity patterns with minimal overhead. (3) Kernel Optimization, employing permutation-free and segment-based strategies to boost system performance. We implement LeMo as an end-to-end fine-tuning system compatible with various LLM architectures and other optimization techniques. Comprehensive evaluations demonstrate that LeMo reduces memory consumption by up to 1.93x and achieves up to 1.36x speedups, outperforming state-of-the-art fine-tuning systems.

Jinrui Zhang, Deyu Zhang, Tingting Long et al.

We present MobiFuse, a high-precision depth perception system on mobile devices that combines dual RGB and Time-of-Flight (ToF) cameras. To achieve this, we leverage physical principles from various environmental factors to propose the Depth Error Indication (DEI) modality, characterizing the depth error of ToF and stereo-matching. Furthermore, we employ a progressive fusion strategy, merging geometric features from ToF and stereo depth maps with depth error features from the DEI modality to create precise depth maps. Additionally, we create a new ToF-Stereo depth dataset, RealToF, to train and validate our model. Our experiments demonstrate that MobiFuse excels over baselines by significantly reducing depth measurement errors by up to 77.7%. It also showcases strong generalization across diverse datasets and proves effectiveness in two downstream tasks: 3D reconstruction and 3D segmentation. The demo video of MobiFuse in real-life scenarios is available at the de-identified YouTube link(https://youtu.be/jy-Sp7T1LVs).

Yongheng Deng, Ziqing Qiao, Ju Ren et al.

While large language models (LLMs) are empowered with broad knowledge, their task-specific performance is often suboptimal. It necessitates fine-tuning LLMs with task-specific data, but such data may be inaccessible due to privacy concerns. In this paper, we propose a novel approach to enhance LLMs with smaller language models (SLMs) that are trained on clients using their private task-specific data. To enable mutual enhancement between LLMs and SLMs, we propose CrossLM, where the SLMs promote the LLM to generate task-specific high-quality data, and both the LLM and SLMs are enhanced with the generated data. We evaluate CrossLM using publicly accessible language models across a range of benchmark tasks. The results demonstrate that CrossLM significantly enhances the task-specific performance of SLMs on clients and the LLM on the cloud server simultaneously while preserving the LLM's generalization capability.

Sheng Yue, Ju Ren, Jiang Xin et al.

Federated meta-learning (FML) has emerged as a promising paradigm to cope with the data limitation and heterogeneity challenges in today's edge learning arena. However, its performance is often limited by slow convergence and corresponding low communication efficiency. In addition, since the available radio spectrum and IoT devices' energy capacity are usually insufficient, it is crucial to control the resource allocation and energy consumption when deploying FML in practical wireless networks. To overcome the challenges, in this paper, we rigorously analyze each device's contribution to the global loss reduction in each round and develop an FML algorithm (called NUFM) with a non-uniform device selection scheme to accelerate the convergence. After that, we formulate a resource allocation problem integrating NUFM in multi-access wireless systems to jointly improve the convergence rate and minimize the wall-clock time along with energy cost. By deconstructing the original problem step by step, we devise a joint device selection and resource allocation strategy to solve the problem with theoretical guarantees. Further, we show that the computational complexity of NUFM can be reduced from $O(d^2)$ to $O(d)$ (with the model dimension $d$) via combining two first-order approximation techniques. Extensive simulation results demonstrate the effectiveness and superiority of the proposed methods in comparison with existing baselines.

Sheng Yue, Ju Ren, Jiang Xin et al.

In order to meet the requirements for performance, safety, and latency in many IoT applications, intelligent decisions must be made right here right now at the network edge. However, the constrained resources and limited local data amount pose significant challenges to the development of edge AI. To overcome these challenges, we explore continual edge learning capable of leveraging the knowledge transfer from previous tasks. Aiming to achieve fast and continual edge learning, we propose a platform-aided federated meta-learning architecture where edge nodes collaboratively learn a meta-model, aided by the knowledge transfer from prior tasks. The edge learning problem is cast as a regularized optimization problem, where the valuable knowledge learned from previous tasks is extracted as regularization. Then, we devise an ADMM based federated meta-learning algorithm, namely ADMM-FedMeta, where ADMM offers a natural mechanism to decompose the original problem into many subproblems which can be solved in parallel across edge nodes and the platform. Further, a variant of inexact-ADMM method is employed where the subproblems are `solved' via linear approximation as well as Hessian estimation to reduce the computational cost per round to $\mathcal{O}(n)$. We provide a comprehensive analysis of ADMM-FedMeta, in terms of the convergence properties, the rapid adaptation performance, and the forgetting effect of prior knowledge transfer, for the general non-convex case. Extensive experimental studies demonstrate the effectiveness and efficiency of ADMM-FedMeta, and showcase that it substantially outperforms the existing baselines.