Haibo Chen

Zhizhong Wang, Lei Zhao, Zhiwen Zuo et al.

Arbitrary style transfer (AST) transfers arbitrary artistic styles onto content images. Despite the recent rapid progress, existing AST methods are either incapable or too slow to run at ultra-resolutions (e.g., 4K) with limited resources, which heavily hinders their further applications. In this paper, we tackle this dilemma by learning a straightforward and lightweight model, dubbed MicroAST. The key insight is to completely abandon the use of cumbersome pre-trained Deep Convolutional Neural Networks (e.g., VGG) at inference. Instead, we design two micro encoders (content and style encoders) and one micro decoder for style transfer. The content encoder aims at extracting the main structure of the content image. The style encoder, coupled with a modulator, encodes the style image into learnable dual-modulation signals that modulate both intermediate features and convolutional filters of the decoder, thus injecting more sophisticated and flexible style signals to guide the stylizations. In addition, to boost the ability of the style encoder to extract more distinct and representative style signals, we also introduce a new style signal contrastive loss in our model. Compared to the state of the art, our MicroAST not only produces visually superior results but also is 5-73 times smaller and 6-18 times faster, for the first time enabling super-fast (about 0.5 seconds) AST at 4K ultra-resolutions. Code is available at https://github.com/EndyWon/MicroAST.

Haibo Chen, Lei Zhao, Jun Li et al.

Artistic style transfer aims to create new artistic images by rendering a given photograph with the target artistic style. Existing methods learn styles simply based on global statistics or local patches, lacking careful consideration of the drawing process in practice. Consequently, the stylization results either fail to capture abundant and diversified local style patterns, or contain undesired semantic information of the style image and deviate from the global style distribution. To address this issue, we imitate the drawing process of humans and propose a Two-Stage Statistics-Aware Transformation (TSSAT) module, which first builds the global style foundation by aligning the global statistics of content and style features and then further enriches local style details by swapping the local statistics (instead of local features) in a patch-wise manner, significantly improving the stylization effects. Moreover, to further enhance both content and style representations, we introduce two novel losses: an attention-based content loss and a patch-based style loss, where the former enables better content preservation by enforcing the semantic relation in the content image to be retained during stylization, and the latter focuses on increasing the local style similarity between the style and stylized images. Extensive qualitative and quantitative experiments verify the effectiveness of our method.

Hongyu Lin, Mingyu Li, Weichen Zhang et al.

Documentation has long guided computer system tuning by distilling expert knowledge into per-parameter recommendations. Yet such guides capture only what experts conclude, discarding how they reason. This fundamental gap manifests in three concrete deficiencies: documentation grows stale as software evolves, fails under heterogeneous workloads, and ignores inter-parameter dependencies. We propose shifting from static documentation to dynamic action for system tuning. We introduce PerfEvolve, which translates expert tuning methodologies into executable skills that equip LLM-based agents to perform version-consistency verification, workload-specific profiling, and multi-parameter joint optimization. Evaluated on PostgreSQL under TPC-C and TPC-H benchmarks, PerfEvolve outperforms state-of-the-art documentation-driven tuning baselines by up to 35.2%. The tool is available at https://github.com/ISCAS-OSLab/PerfEvolve.

Bin Xu, Pengfei Hu, Wenxin Zheng et al.

GPU-based simulation environments for embodied AI interleave physics simulation (CUDA) and photorealistic rendering (Vulkan) on a single device. We observe that two foundational scenarios -- simulation data generation and RL training -- can be naturally adapted to execute their simulation and rendering phases concurrently, presenting a significant opportunity to improve GPU utilization through spatial multiplexing. However, a fundamental obstacle we term execution isolation prevents this: CUDA and Vulkan create separate GPU contexts whose channels are bound to different scheduling groups, confining compute and graphics to mutually exclusive time slices. Existing spatial-sharing techniques are limited to the CUDA ecosystem, while temporal-sharing approaches underutilize available resources. This paper presents VUDA, a system that breaks execution isolation to enable spatial parallelism between CUDA compute and Vulkan graphics workloads. VUDA is built on two key observations: although CUDA and Vulkan expose different programming abstractions, their execution paths converge to a common channel primitive at the driver and hardware level; meanwhile, their virtual-address spaces are inherently disjoint, making safe page-table merging feasible without remapping. VUDA exposes a thin API for developers to annotate co-schedulable CUDA streams, and realizes spatial sharing through channel redirection into Vulkan's scheduling domain and page-table grafting to unify address spaces, eliminating all data copying on the critical path. Experiments on representative embodied-AI workloads show that VUDA delivers up to 85% higher throughput than temporal-sharing baselines, while improving GPU utilization and reducing end-to-end latency.

Yunpeng Dong, Jingkai He, Yuze Hou et al.

LLM-powered AI agents require high-frequency state exploration (e.g., test-time tree search and reinforcement learning), relying on rapid checkpoint and rollback (C/R) of the complete sandbox state, including files and process state (e.g., memory, contexts, etc.). Existing mechanisms duplicate the entire state, causing hundreds of milliseconds to seconds of latency per C/R, which severely bottlenecks deep search and large-scale fan-outs. This paper observes that subsequent checkpoints in AI agents are highly similar. Therefore, instead of full duplication, a sandbox should only duplicate the changes between consecutive checkpoints (Key Insight). However, it is non-trivial to realize the idea, mainly due to the missing OS supports. This paper proposes a new OS-level abstraction, DeltaState, to enable the change-based transactional C/R for AI agents with two co-designed OS mechanisms. First, DeltaFS enables change-based filesystem C/R by organizing the file states into layers and dynamically freezing the writable layer and inserting a new one during checkpoint, reducing file updates to copy-on-write, and making rollback a simple layer switch. Second, DeltaCR enables change-based process state C/R using incremental dumps, and accelerates rollback by bypassing traditional pipelines to directly fork() from a frozen template process. We then present DeltaBox, a novel agent sandbox achieving millisecond level C/R through the two new mechanisms. Evaluations on SWE-bench and RL micro-benchmarks show DeltaBox completes checkpoint and rollback in millisecond-level latency (14ms and 5ms, respectively), empowering agents to explore substantially more nodes under fixed time budgets.

Jun Li, Zikun Chen, Haibo Chen et al.

Novel object synthesis by integrating distinct textual concepts from diverse categories remains a significant challenge in Text-to-Image (T2I) generation. Existing methods often suffer from insufficient concept mixing, lack of rigorous evaluation, and suboptimal outputs-manifesting as conceptual imbalance, superficial combinations, or mere juxtapositions. To address these limitations, we propose Reinforcement Mixing Learning (RMLer), a framework that formulates cross-category concept fusion as a reinforcement learning problem: mixed features serve as states, mixing strategies as actions, and visual outcomes as rewards. Specifically, we design an MLP-policy network to predict dynamic coefficients for blending cross-category text embeddings. We further introduce visual rewards based on (1) semantic similarity and (2) compositional balance between the fused object and its constituent concepts, optimizing the policy via proximal policy optimization. At inference, a selection strategy leverages these rewards to curate the highest-quality fused objects. Extensive experiments demonstrate RMLer's superiority in synthesizing coherent, high-fidelity objects from diverse categories, outperforming existing methods. Our work provides a robust framework for generating novel visual concepts, with promising applications in film, gaming, and design.

Yiyang Wu, Xunjie Wang, Jinyu Gu et al.

Secure containers isolate each container with its own kernel, mitigating shared-kernel attacks prevalent in traditional container systems. However, existing designs still face a fundamental isolation--performance trade-off. Nested-cloud deployments amplify the cost of VM exits and page-table management, while emerging agentic workloads expose bursty memory demand that requires fine-grained elasticity. We attribute this trade-off to two root causes. First, existing designs lack lightweight intra-container isolation primitives for frequent container user--kernel transitions. Second, the host treats container memory management as opaque, forcing reactive secondary faults and coarse-grained huge page mappings to amortize their cost. This paper presents ParaCell, a paravirtualized secure container runtime built on two insights. First, intra-address-space hardware protection primitives can provide lightweight intra-container isolation. ParaCell uses MPK-based XGates to isolate the container user and container kernel within a single address space, turning frequent user--kernel transitions into direct domain switches. Second, container kernel allocators already encode memory-management intent. ParaCell introduces Pager to interpose on allocation and free events, batch proactive GPA to HPA bindings and unbindings, and avoid reactive shadow page-table faults while preserving fine-grained memory elasticity. ParaCell is implemented as a drop-in replacement for RunV. Our experiments demonstrate that, across traditional cloud and emerging agent applications, ParaCell reduces latency by up to 57% and 79% over PVM, and by up to 33% and 88% over RunV, in bare-metal and nested setups, respectively. On agent workloads, ParaCell saves up to 35.6% memory compared with the state-of-the-art VM memory reclamation technique, HyperAlloc.

Yifan Peng, Jiafei Fan, Xingda Wei et al.

Learned top-K search is a promising approach for serving vector queries with both high accuracy and performance. However, current models trained for a specific K value fail to generalize to real-world multi-K queries: they suffer from accuracy degradation (for larger Ks) and performance loss (for smaller Ks). Training the model to generalize on different Ks requires orders of magnitude more preprocessing time and is not suitable for serving vector queries in the wild. We present OMEGA, a K-generalizable learned top-K search method that simultaneously achieves high accuracy, high performance, and low preprocessing cost for multi-K vector queries. The key idea is that a base model properly trained on K=1 with our trajectory-based features can be used to accurately predict larger Ks with a dynamic refinement procedure and smaller Ks with minimal performance loss. To make our refinements efficient, we further leverage the statistical properties of top-K searches to reduce excessive model invocations. Extensive evaluations on multiple public and production datasets show that, under the same preprocessing budgets, OMEGA achieves 6-33% lower average latency compared to state-of-the-art learned search methods, while all systems achieve the same recall target. With only 16-30% of the preprocessing time, OMEGA attains 1.01-1.28x of the optimal average latency of these baselines.

Yue Yu, Haibo Chen, Shuo Chen et al.

Instilling creativity in text-to-image (T2I) generation presents a significant challenge, as it requires synthesized images to exhibit not only visual novelty and surprise, but also artistic value. Current T2I models, however, are largely optimized for literal text-image alignment with their data distribution, and their noise prediction networks constrain the generation to high-probability regions, consequently generating outputs that lack authentic creativity. To address this, we propose a Self-Creative Diffusion (SCDiff) model for meaningful T2I generations featuring two core modules: a learnable spatial weighting (LSW) module and a visual-semantic mixing loss (VSML). The LSW module designs a parametric Kaiser-Bessel window to reinforce central image features, fostering novel and surprising generation. The VSML module introduces a dual loss function: a similarity loss constrains that the new images align with its textual description, while a diversity loss maximizes its distinction from the original image, enhancing both semantic value and visual novelty. Extensive experiments demonstrate that our model substantially improves creativity, semantic alignment, and visual coherence, offering a simple yet powerful framework for generating creative objects.

Yanqi Zhang, Yuwei Hu, Runyuan Zhao et al.

Large language models (LLMs) demonstrate remarkable capabilities but face substantial serving costs due to their high memory demands, with the key-value (KV) cache being a primary bottleneck. State-of-the-art KV cache compression techniques, such as quantization and pruning, apply uniform treatment to both keys and values, and discard unimportant tokens entirely, overlooking the fine-grained distinctions in the significance of individual KV cache components. To address such limitations, we introduce \textit{DiffKV}, a novel framework for efficient KV cache compression that exploits three levels of differentiation in the KV cache: (1) the differing impact of keys and values on attention computation, (2) the varying importance of tokens, and (3) the diverse dynamic sparsity patterns across attention heads. These levels of differentiation introduce irregular memory usage patterns across different requests and attention heads, posing significant scalability challenges for memory management. To address these challenges, DiffKV proposes an on-GPU memory manager that compacts fragmented free memory list into contiguous regions in parallel, effectively translating sparsity in the KV cache into performance gains. We evaluate DiffKV on several mainstream LLMs, including the emerging thinking models that generate extended chains of thought. DiffKV is able to compress the KV cache by $2.7\times$ to $5.7\times$ with near-lossless accuracy on complex workloads requiring sophisticated reasoning and long-generation capabilities, and enhances throughput by $1.9\times$ to $5.4\times$. Source codes of DiffKV are available at https://github.com/zyqCSL/DiffKV.

Feiyang Chen, Yu Cheng, Lei Wang et al.

Transformers and large language models (LLMs) have revolutionized machine learning, with attention mechanisms at the core of their success. As the landscape of attention variants expands, so too do the challenges of optimizing their performance, particularly across different hardware platforms. Current optimization strategies are often narrowly focused, requiring extensive manual intervention to accommodate changes in model configurations or hardware environments. In this paper, we introduce AttentionEngine, a comprehensive framework designed to streamline the optimization of attention mechanisms across heterogeneous hardware backends. By decomposing attention computation into modular operations with customizable components, AttentionEngine enables flexible adaptation to diverse algorithmic requirements. The framework further automates kernel optimization through a combination of programmable templates and a robust cross-platform scheduling strategy. Empirical results reveal performance gains of up to 10x on configurations beyond the reach of existing methods. AttentionEngine offers a scalable, efficient foundation for developing and deploying attention mechanisms with minimal manual tuning. Our code has been open-sourced and is available at https://github.com/microsoft/AttentionEngine.

Yixin Song, Zeyu Mi, Haotong Xie et al.

This paper introduces PowerInfer, a high-speed Large Language Model (LLM) inference engine on a personal computer (PC) equipped with a single consumer-grade GPU. The key principle underlying the design of PowerInfer is exploiting the high locality inherent in LLM inference, characterized by a power-law distribution in neuron activation. This distribution indicates that a small subset of neurons, termed hot neurons, are consistently activated across inputs, while the majority, cold neurons, vary based on specific inputs. PowerInfer exploits such an insight to design a GPU-CPU hybrid inference engine: hot-activated neurons are preloaded onto the GPU for fast access, while cold-activated neurons are computed on the CPU, thus significantly reducing GPU memory demands and CPU-GPU data transfers. PowerInfer further integrates adaptive predictors and neuron-aware sparse operators, optimizing the efficiency of neuron activation and computational sparsity. The evaluation shows that PowerInfer significantly outperforms llama.cpp by up to 11.69x while retaining model accuracy across various LLMs (including OPT-175B) on a single NVIDIA RTX 4090 GPU. For the OPT-30B model, PowerInfer achieves performance comparable to that of a high-end server-grade A100 GPU, reaching 82% of its token generation rate on a single consumer-grade RTX 4090 GPU.

Haibo Chen, Xin Wang, Jiaheng Chao et al.

Leveraging Graph Neural Networks (GNNs) as graph encoders and aligning the resulting representations with Large Language Models (LLMs) through alignment instruction tuning has become a mainstream paradigm for constructing Graph Language Models (GLMs), combining the generalization ability of LLMs with the structural modeling capacity of GNNs. However, existing GLMs that adopt GNNs as graph encoders largely overlook the problem of aligning GNN-encoded representations across domains and tasks with the LLM token space to obtain unified graph tokens, thereby limiting their ability to generalize across diverse graph data. To bridge this gap, we aim to incorporate a multi-domain, multi-task GNN encoder into GLMs and align its representations with LLMs to enable multi-domain, multi-task graph alignment instruction tuning. This alignment problem remains underexplored and poses two key challenges: 1) learning GNN-encoded representations that are simultaneously generalizable across domains and tasks and well aligned with textual semantics is difficult, due to substantial variations in graph structures, feature distributions, and supervision signals, together with the lack of textual-semantic alignment guidance in task-specific GNN training; 2) diverse graph data and task-specific instructions can exhibit different degrees of compatibility with the LLM token space during instruction tuning, leading to varying alignment difficulty and rendering a fixed alignment strategy suboptimal. To tackle these challenges, we propose UniGraphLM, a Unified Graph Language Model that incorporates a multi-domain, multi-task GNN encoder to learn generalizable graph representations aligned with textual semantics, and then adaptively aligns these representations with the LLM.

Haoran Ding, Zhaoguo Wang, Haibo Chen

LLM-assisted software development has become increasingly prevalent, and can generate large-scale systems, such as compilers. It becomes crucial to strengthen the correctness of the generated code. However, automated reasoning for large-scale systems remains challenging due to code complexity. Hoare logic offers an approach to decomposing a large system into smaller components and reasoning about them separately (i.e., compositional reasoning). However, existing works still struggle to scale, because Hoare logic requires writing formal specifications for each function, imposing a heavy human burden. The problem is exacerbated when code is generated by LLMs, as developers lack a deep understanding of each function's expected behavior. This paper presents FM-Agent, the first framework that realizes automated compositional reasoning for large-scale systems. Leveraging LLMs, FM-Agent introduces a top-down paradigm to automatically generate function-level specifications. Specifically, FM-Agent derives the specification of a function from how its callers expect the function to behave, so the generated specifications can reflect the developer's intent of a function even if the implementation is buggy. Developers' intent is usually expressed in natural language, while existing verifiers only support formulas. Therefore, FM-Agent generalizes Hoare-style inference to reason about functions against natural-language specifications. Finally, to confirm bug existence and explain bug causes, FM-Agent automatically generates test cases to trigger potential bugs. In our evaluation, FM-Agent successfully reasons about large-scale systems within 2 days, each of which has up to 143k LoC. These systems have already been tested by their developers, but FM-Agent still finds 522 newly discovered bugs. These bugs can cause serious consequences, including system crashes and incorrect execution results.

Yixin Song, Haotong Xie, Zhengyan Zhang et al.

Exploiting activation sparsity is a promising approach to significantly accelerating the inference process of large language models (LLMs) without compromising performance. However, activation sparsity is determined by activation functions, and commonly used ones like SwiGLU and GeGLU exhibit limited sparsity. Simply replacing these functions with ReLU fails to achieve sufficient sparsity. Moreover, inadequate training data can further increase the risk of performance degradation. To address these challenges, we propose a novel dReLU function, which is designed to improve LLM activation sparsity, along with a high-quality training data mixture ratio to facilitate effective sparsification. Additionally, we leverage sparse activation patterns within the Feed-Forward Network (FFN) experts of Mixture-of-Experts (MoE) models to further boost efficiency. By applying our neuron sparsification method to the Mistral and Mixtral models, only 2.5 billion and 4.3 billion parameters are activated per inference iteration, respectively, while achieving even more powerful model performance. Evaluation results demonstrate that this sparsity achieves a 2-5x decoding speedup. Remarkably, on mobile phones, our TurboSparse-Mixtral-47B achieves an inference speed of 11 tokens per second. Our models are available at \url{https://huggingface.co/PowerInfer}

Jiahao Chen, Dingji Li, Zeyu Mi et al.

Today's mainstream virtualization systems comprise of two cooperative components: a kernel-resident driver that accesses virtualization hardware and a user-level helper process that provides VM management and I/O virtualization. However, this virtualization architecture has intrinsic issues in both security (a large attack surface) and performance. While there is a long thread of work trying to minimize the kernel-resident driver by offloading functions to user mode, they face a fundamental tradeoff between security and performance: more offloading may reduce the kernel attack surface, yet increase the runtime ring crossings between the helper process and the driver, and thus more performance cost. This paper explores a new design called delegated virtualization, which completely separates the control plane (the kernel driver) from the data plane (the helper process) and thus eliminates the kernel driver from runtime intervention. The resulting user-level hypervisor, called DuVisor, can handle all VM operations without trapping into the kernel once the kernel driver has done the initialization. DuVisor retrofits existing hardware virtualization support with a new delegated virtualization extension to directly handle VM exits, configure virtualization registers, manage the stage-2 page table and virtual devices in user mode. We have implemented the hardware extension on an open-source RISC-V CPU and built a Rust-based hypervisor atop the hardware. Evaluation on FireSim shows that DuVisor outperforms KVM by up to 47.96\% in a variety of real-world applications and significantly reduces the attack surface.

Zhizhong Wang, Lei Zhao, Haibo Chen et al.

In this paper, we present the texture reformer, a fast and universal neural-based framework for interactive texture transfer with user-specified guidance. The challenges lie in three aspects: 1) the diversity of tasks, 2) the simplicity of guidance maps, and 3) the execution efficiency. To address these challenges, our key idea is to use a novel feed-forward multi-view and multi-stage synthesis procedure consisting of I) a global view structure alignment stage, II) a local view texture refinement stage, and III) a holistic effect enhancement stage to synthesize high-quality results with coherent structures and fine texture details in a coarse-to-fine fashion. In addition, we also introduce a novel learning-free view-specific texture reformation (VSTR) operation with a new semantic map guidance strategy to achieve more accurate semantic-guided and structure-preserved texture transfer. The experimental results on a variety of application scenarios demonstrate the effectiveness and superiority of our framework. And compared with the state-of-the-art interactive texture transfer algorithms, it not only achieves higher quality results but, more remarkably, also is 2-5 orders of magnitude faster. Code is available at https://github.com/EndyWon/Texture-Reformer.

Dongdong Zhang, Dong Cao, Haibo Chen

Brain computer interface (BCI) has been popular as a key approach to monitor our brains recent year. Mental states monitoring is one of the most important BCI applications and becomes increasingly accessible. However, the mental state prediction accuracy and generality through encephalogram (EEG) are not good enough for everyday use. Here in this paper we present a deep learning-based EEG decoding method to read mental states. We propose a novel 1D convolutional neural network with different filter lengths to capture different frequency bands information. To improve the prediction accuracy, we also used a resnet-like structure to train a relatively deep convolutional neural network to promote feature extraction. Compared with traditional ways of predictions such as KNN and SVM, we achieved a significantly better result with an accuracy of 96.40%. Also, in contrast with some already published open source deep neural network structures, our methods achieved the state of art prediction accuracy on a mental state recognition dataset. Our results demonstrate using only 1D convolution could extract the features of EEG and the possibility of mental state prediction using portable EEG devices.

Xin Wang, Haibo Chen, Wenxuan Liu et al.

Foundation models (FMs) are increasingly deployed in open-world settings where distribution shift is the rule rather than the exception. The out-of-distribution (OOD) phenomena they face -- knowledge boundaries, capability ceilings, compositional shifts, and open-ended task variation -- differ in kind from the settings that have shaped prior OOD research, and are further complicated because the pretraining and post-training distributions of modern FMs are often only partially observed. Our position is that OOD for foundation models is a structurally distinct problem that cannot be solved within the prevailing model-centric paradigm, and that agentic systems constitute the missing paradigm required to address it. We defend this claim through four steps. First, we give a stage-aware formalization of OOD that accommodates partially observed multi-stage training distributions. Second, we prove a parameter coverage ceiling: there exist practically relevant inputs that no model-centric method (training-time or test-time) can handle within tolerance $\varepsilon$, for reasons intrinsic to parameter-based representation. Third, we characterize agentic OOD systems by four structural properties -- perception, strategy selection, external action, and closed-loop verification -- and show that they strictly extend the reachable set beyond the ceiling. Fourth, we respond to seven counterarguments, conceding two, and outline a research agenda. We do not claim that agentic methods subsume model-centric ones; we argue that the two are complementary, and that progress on FM-OOD requires explicit recognition of the agentic paradigm as a first-class research direction.

Xiaoniu Song, Zihang Zhong, Rong Chen et al.

The promising applications of large language models are often limited by the constrained GPU memory capacity available on edge devices. Mixture-of-Experts (MoE) models help address this issue by activating only a subset of the model's parameters during computation. This approach allows the unused parameters to be offloaded to host memory, thereby reducing the overall GPU memory demand. However, existing cache-based offloading solutions handle cache misses reactively, which significantly impacts system performance. In this paper, we introduce ProMoE, a novel proactive caching system that utilizes intermediate results to predict subsequent expert usage. By proactively fetching experts in advance, ProMoE eliminates passive cache misses, removes loading time from the critical path, and reduces the performance overhead associated with offloading. Our evaluations demonstrate that ProMoE achieves an average speedup of 2.20x (up to 3.21x) and 2.07x (up to 5.02x) in the prefill and decode stages, respectively, compared to existing offloading solutions.

Le Chen, Erhu Feng, Yubin Xia et al.

LLM agents increasingly adopt skills as a reusable unit of composition. While skills are shared across diverse agent platforms, current systems treat them as raw context, causing the same skill to behave inconsistently for different agents. This fragility undermines skill portability and execution efficiency. To address this challenge, we analyze 118,000 skills and draw inspiration from traditional compiler design. We treat skills as code and LLMs as heterogeneous processors. To make portability actionable, we decompose a skill's requirements into a set of primitive capabilities, and measure how well each model-harness pair supports them. Based on these capability profiles, we propose SkillRT, a compilation and runtime system designed for portable and efficient skill execution. At compile time, SkillRT performs capability-based compilation, environment binding, and concurrency extraction. At runtime, SkillRT applies JIT code solidification and adaptive recompilation for performance optimization. We evaluate SkillRT across eight LLMs of varying scales and three agent harnesses, covering SkillsBench and representative skill tasks. Results demonstrate that SkillRT significantly improves task completion rates across different models and environments while reducing token consumption by up to 40%. In terms of performance, SkillRT achieves up to 3.2x speedup with enhanced parallelism, and 19-50x latency reduction through code solidification.

Jingkai He, Tianjian Li, Erhu Feng et al.

With the rapid advancement of large language models (LLMs), reinforcement learning (RL) has emerged as a pivotal methodology for enhancing the reasoning capabilities of LLMs. Unlike traditional pre-training approaches, RL encompasses multiple stages: rollout, reward, and training, which necessitates collaboration among various worker types. However, current RL systems continue to grapple with substantial GPU underutilization, due to two primary factors: (1) The rollout stage dominates the overall RL process due to test-time scaling; (2) Imbalances in rollout lengths (within the same batch) result in GPU bubbles. While prior solutions like asynchronous execution and truncation offer partial relief, they may compromise training accuracy for efficiency. Our key insight stems from a previously overlooked observation: rollout responses exhibit remarkable similarity across adjacent training epochs. Based on the insight, we introduce RhymeRL, an LLM RL system designed to accelerate RL training with two key innovations. First, to enhance rollout generation, we present HistoSpec, a speculative decoding inference engine that utilizes the similarity of historical rollout token sequences to obtain accurate drafts. Second, to tackle rollout bubbles, we introduce HistoPipe, a two-tier scheduling strategy that leverages the similarity of historical rollout distributions to balance workload among rollout workers. We have evaluated RhymeRL within a real production environment, demonstrating scalability from dozens to thousands of GPUs. Experimental results demonstrate that RhymeRL achieves a 2.6x performance improvement over existing methods, without compromising accuracy or modifying the RL paradigm.

Jiahao Wang, Jinbo Han, Xingda Wei et al.

Serving large language models (LLMs) is important for cloud providers, and caching intermediate results (KV\$) after processing each request substantially improves serving throughput and latency. However, there is limited understanding of how LLM serving benefits from KV\$ caching, where system design decisions like cache eviction policies are highly workload-dependent. In this paper, we present the first systematic characterization of the KV\$ workload patterns from one of the leading LLM service providers. We draw observations that were not covered by previous studies focusing on synthetic workloads, including: KV\$ reuses are skewed across requests, where reuses between single-turn requests are equally important as multi-turn requests; the reuse time and probability are diverse considering all requests, but for a specific request category, the pattern tends to be predictable; and the overall cache size required for an ideal cache hit ratio is moderate. Based on the characterization, we further propose a workload-aware cache eviction policy that improves the serving performance under real-world traces, especially with limited cache capacity.

Le Chen, Dahu Feng, Erhu Feng et al.

With the rapid advancement of artificial intelligence technologies such as ChatGPT, AI agents, and video generation, contemporary mobile systems have begun integrating these AI capabilities on local devices to enhance privacy and reduce response latency. To meet the computational demands of AI tasks, current mobile SoCs are equipped with diverse AI accelerators, including GPUs and Neural Processing Units (NPUs). However, there has not been a comprehensive characterization of these heterogeneous processors, and existing designs typically only leverage a single AI accelerator for LLM inference, leading to suboptimal use of computational resources and memory bandwidth. In this paper, we first summarize key performance characteristics of heterogeneous processors, SoC memory bandwidth, etc. Drawing on these observations, we propose different heterogeneous parallel mechanisms to fully exploit both GPU and NPU computational power and memory bandwidth. We further design a fast synchronization mechanism between heterogeneous processors that leverages the unified memory architecture. By employing these techniques, we present HeteroInfer, the fastest LLM inference engine in mobile devices which supports GPU-NPU heterogeneous execution. Evaluation shows that HeteroInfer delivers a 1.34x to 6.02x end-to-end speedup over state-of-the-art GPU-only and NPU-only LLM engines, while maintaining negligible interference with other applications.

Erhu Feng, Wenbo Zhou, Zibin Liu et al.

AI agents, empowered by Large Language Models (LLMs) and communication protocols such as MCP and A2A, have rapidly evolved from simple chatbots to autonomous entities capable of executing complex, multi-step tasks, demonstrating great potential. However, the LLMs' inherent uncertainty and heavy computational resource requirements pose four significant challenges to the development of safe and efficient agents: reliability, privacy, cost and performance. Existing approaches, like model alignment, workflow constraints and on-device model deployment, can partially alleviate some issues but often with limitations, failing to fundamentally resolve these challenges. This paper proposes a new paradigm called AgentRR (Agent Record & Replay), which introduces the classical record-and-replay mechanism into AI agent frameworks. The core idea is to: 1. Record an agent's interaction trace with its environment and internal decision process during task execution, 2. Summarize this trace into a structured "experience" encapsulating the workflow and constraints, and 3. Replay these experiences in subsequent similar tasks to guide the agent's behavior. We detail a multi-level experience abstraction method and a check function mechanism in AgentRR: the former balances experience specificity and generality, while the latter serves as a trust anchor to ensure completeness and safety during replay. In addition, we explore multiple application modes of AgentRR, including user-recorded task demonstration, large-small model collaboration and privacy-aware agent execution, and envision an experience repository for sharing and reusing knowledge to further reduce deployment cost.

Tianyi Chu, Wei Xing, Jiafu Chen et al.

Existing generative adversarial network (GAN) based conditional image generative models typically produce fixed output for the same conditional input, which is unreasonable for highly subjective tasks, such as large-mask image inpainting or style transfer. On the other hand, GAN-based diverse image generative methods require retraining/fine-tuning the network or designing complex noise injection functions, which is computationally expensive, task-specific, or struggle to generate high-quality results. Given that many deterministic conditional image generative models have been able to produce high-quality yet fixed results, we raise an intriguing question: is it possible for pre-trained deterministic conditional image generative models to generate diverse results without changing network structures or parameters? To answer this question, we re-examine the conditional image generation tasks from the perspective of adversarial attack and propose a simple and efficient plug-in projected gradient descent (PGD) like method for diverse and controllable image generation. The key idea is attacking the pre-trained deterministic generative models by adding a micro perturbation to the input condition. In this way, diverse results can be generated without any adjustment of network structures or fine-tuning of the pre-trained models. In addition, we can also control the diverse results to be generated by specifying the attack direction according to a reference text or image. Our work opens the door to applying adversarial attack to low-level vision tasks, and experiments on various conditional image generation tasks demonstrate the effectiveness and superiority of the proposed method.

Rongxin Cheng, Yuxin Lai, Xingda Wei et al.

Serving LLMs with a cluster of GPUs is common nowadays, where the serving system must meet strict latency SLOs required by applications. However, the stateful nature of LLM serving requires maintaining huge states (i.e., KVCache) in limited GPU memory. Under spikes in real-world workloads, GPU memory can be easily throttled, leading to orders of magnitude higher response latency due to queuing introduced by waiting for KVCache to be reclaimed. Prior KVCache-centric approaches handle load throttling by dropping, migrating, or swapping KVCache. These methods fail to release sufficient memory quickly with requests still queued. This paper proposes the first parameter-centric approach to handling throttling by selectively dropping replicated parameters to instantly free memory for requests, based on an unnoticed observation that model parameters are commonly replicated across GPUs for serving LLMs. With additional memory, all requests can be served with a larger batch without queuing. To make the parameter-centric approach correct and efficient, we cooperatively execute requests on GPUs with a complete copy of parameters using pipeline parallelism, and derive an appropriate drop plan without unnecessary cooperation. We also design techniques to minimize the performance overhead due to pipeline parallelism with the execution patterns of requests under drop. Evaluations show that {\sys} reduces the tail TTFT of requests under throttling by up to 72.2 times compared to the state-of-the-art systems including Llumnix, vLLM and InferCept.

Jiafu Chen, Wei Xing, Jiakai Sun et al.

3D scene stylization refers to transform the appearance of a 3D scene to match a given style image, ensuring that images rendered from different viewpoints exhibit the same style as the given style image, while maintaining the 3D consistency of the stylized scene. Several existing methods have obtained impressive results in stylizing 3D scenes. However, the models proposed by these methods need to be re-trained when applied to a new scene. In other words, their models are coupled with a specific scene and cannot adapt to arbitrary other scenes. To address this issue, we propose a novel 3D scene stylization framework to transfer an arbitrary style to an arbitrary scene, without any style-related or scene-related re-training. Concretely, we first map the appearance of the 3D scene into a 2D style pattern space, which realizes complete disentanglement of the geometry and appearance of the 3D scene and makes our model be generalized to arbitrary 3D scenes. Then we stylize the appearance of the 3D scene in the 2D style pattern space via a prompt-based 2D stylization algorithm. Experimental results demonstrate that our proposed framework is superior to SOTA methods in both visual quality and generalization.

Cheng Zhang, Erhu Feng, Xi Zhao et al.

With the rapid advancement of Vision-Language Models (VLMs), GUI-based mobile agents have emerged as a key development direction for intelligent mobile systems. However, existing agent models continue to face significant challenges in real-world task execution, particularly in terms of accuracy and efficiency. To address these limitations, we propose MobiAgent, a comprehensive mobile agent system comprising three core components: the MobiMind-series agent models, the AgentRR acceleration framework, and the MobiFlow benchmarking suite. Furthermore, recognizing that the capabilities of current mobile agents are still limited by the availability of high-quality data, we have developed an AI-assisted agile data collection pipeline that significantly reduces the cost of manual annotation. Compared to both general-purpose LLMs and specialized GUI agent models, MobiAgent achieves state-of-the-art performance in real-world mobile scenarios.

Mingyu Wu, Zhaoguo Wang, Jiabin Wang et al.

Autonomous scientific research, capable of independently conducting complex experiments and serving non-specialists, represents a long-held aspiration. Achieving it requires a fundamental paradigm shift driven by artificial intelligence (AI). While autonomous experimental systems are emerging, they remain confined to areas featuring singular objectives and well-defined, simple experimental workflows, such as chemical synthesis and catalysis. We present an AI-native autonomous laboratory, targeting highly complex scientific experiments for applications like autonomous biomolecular engineering. This system autonomously manages instrumentation, formulates experiment-specific procedures and optimization heuristics, and concurrently serves multiple user requests. Founded on a co-design philosophy of models, experiments, and instruments, the platform supports the co-evolution of AI models and the automation system. This establishes an end-to-end, multi-user autonomous laboratory that handles complex, multi-objective experiments across diverse instrumentation. Our autonomous laboratory supports fundamental nucleic acid functions-including synthesis, transcription, amplification, and sequencing. It also enables applications in fields such as disease diagnostics, drug development, and information storage. Without human intervention, it autonomously optimizes experimental performance to match state-of-the-art results achieved by human scientists. In multi-user scenarios, the platform significantly improves instrument utilization and experimental efficiency. This platform paves the way for advanced biomaterials research to overcome dependencies on experts and resource barriers, establishing a blueprint for science-as-a-service at scale.

Xin Wang, Haoyang Li, Haibo Chen et al. · tsinghua

Large language models (LLMs) have substantially advanced machine learning research, including natural language processing, computer vision, data mining, etc., yet they still exhibit critical limitations in explainability, reliability, adaptability, and extensibility. In this paper, we overview a promising learning paradigm, i.e., Modular Machine Learning (MML), as an essential approach toward new-generation LLMs capable of addressing these issues. We begin by systematically and comprehensively surveying the existing literature on modular machine learning, with a particular focus on modular data representation and modular models. Then, we propose a unified MML framework for LLMs, which decomposes the complex structure of LLMs into three interdependent components: modular representation, modular model, and modular reasoning. Specifically, the MML paradigm discussed in this article is able to: i) clarify the internal working mechanism of LLMs through the disentanglement of semantic components; ii) allow for flexible and task-adaptive model design; iii) enable an interpretable and logic-driven decision-making process. We further elaborate a feasible implementation of MML-based LLMs via leveraging advanced techniques such as disentangled representation learning, neural architecture search and neuro-symbolic learning. Last but not least, we critically identify the remaining key challenges, such as the integration of continuous neural and discrete symbolic processes, joint optimization, and computational scalability, present promising future research directions that deserve further exploration. Ultimately, we believe the integration of the MML with LLMs has the potential to bridge the gap between statistical (deep) learning and formal (logical) reasoning, thereby paving the way for robust, adaptable, and trustworthy AI systems across a wide range of real-world applications.

Zhenliang Xue, Hanpeng Hu, Xing Chen et al.

Large multimodal models (LMMs) have demonstrated excellent capabilities in both understanding and generation tasks with various modalities. While these models can accept flexible combinations of input data, their training efficiency suffers from two major issues: pipeline stage imbalance caused by heterogeneous model architectures, and training data dynamicity stemming from the diversity of multimodal data. In this paper, we present PipeWeaver, a dynamic pipeline scheduling framework designed for LMM training. The core of PipeWeaver is dynamic interleaved pipeline, which searches for pipeline schedules dynamically tailored to current training batches. PipeWeaver addresses issues of LMM training with two techniques: adaptive modality-aware partitioning and efficient pipeline schedule search within a hierarchical schedule space. Meanwhile, PipeWeaver utilizes SEMU (Step Emulator), a training simulator for multimodal models, for accurate performance estimations, accelerated by spatial-temporal subgraph reuse to improve search efficiency. Experiments show that PipeWeaver can enhance LMM training efficiency by up to 97.3% compared to state-of-the-art systems, and demonstrate excellent adaptivity to LMM training's data dynamicity.

Gansen Hu, Zhaoguo Wang, Jinglin Wei et al.

Large language models (LLMs) demonstrate remarkable capabilities but face deployment challenges due to their massive parameter counts. While existing compression techniques like pruning can reduce model size, it leads to significant accuracy degradation under high compression ratios. We present a novel perspective inspired by constant folding in compiler optimization. Our approach enables parameter reduction by treating activation functions in LLMs as linear functions. However, recent LLMs use complex non-linear activations like GELU that prevent direct application of this technique. We propose TARDIS, which enables optimization of LLMs with non-linear activations by partially approximating them with linear functions in frequently occurring input ranges. For outlier inputs, TARDIS employs an online predictor to dynamically fall back to original computations. Our experiments demonstrate that TARDIS achieves 80% parameter reduction in feed-forward networks, while significantly outperforming state-of-the-art pruning methods Wanda and RIA with up to 65% higher accuracy. In practical deployments for a 7B model, TARDIS achieves 1.6x end-to-end inference speedup when integrated with the vLLM serving system, and 1.4x speedup with the widely adopted HuggingFace implementation, while incurring only a 10.9% accuracy trade-off.

Zibin Liu, Cheng Zhang, Xi Zhao et al.

Large Language Model (LLM) agents are increasingly deployed to automate complex workflows in mobile and desktop environments. However, current model-centric agent architectures struggle to self-evolve post-deployment: improving personalization, capability, and efficiency typically requires continuous model retraining/fine-tuning, which incurs prohibitive computational overheads and suffers from an inherent trade-off between model accuracy and inference efficiency. To enable iterative self-evolution without model retraining, we propose MOBIMEM, a memory-centric agent system. MOBIMEM first introduces three specialized memory primitives to decouple agent evolution from model weights: (1) Profile Memory uses a lightweight distance-graph (DisGraph) structure to align with user preferences, resolving the accuracy-latency trade-off in user profile retrieval; (2) Experience Memory employs multi-level templates to instantiate execution logic for new tasks, ensuring capability generalization; and (3) Action Memory records fine-grained interaction sequences, reducing the reliance on expensive model inference. Building upon this memory architecture, MOBIMEM further integrates a suite of OS-inspired services to orchestrate execution: a scheduler that coordinates parallel sub-task execution and memory operations; an agent record-and-replay (AgentRR) mechanism that enables safe and efficient action reuse; and a context-aware exception handling that ensures graceful recovery from user interruptions and runtime errors. Evaluation on AndroidWorld and top-50 apps shows that MOBIMEM achieves 83.1% profile alignment with 23.83 ms retrieval time (280x faster than GraphRAG baselines), improves task success rates by up to 50.3%, and reduces end-to-end latency by up to 9x on mobile devices.

Haoyang Li, Haibo Chen, Xin Wang et al.

Graphs are a fundamental data structure for representing relational information in domains such as social networks, molecular systems, and knowledge graphs. However, graph learning models often suffer from limited generalization when applied beyond their training distributions. In practice, distribution shifts may arise from changes in graph structure, domain semantics, available modalities, or task formulations. To address these challenges, graph foundation models (GFMs) have recently emerged, aiming to learn general-purpose representations through large-scale pretraining across diverse graphs and tasks. In this survey, we review recent progress on GFMs from the perspective of out-of-distribution (OOD) generalization. We first discuss the main challenges posed by distribution shifts in graph learning and outline a unified problem setting. We then organize existing approaches based on whether they are designed to operate under a fixed task specification or to support generalization across heterogeneous task formulations, and summarize the corresponding OOD handling strategies and pretraining objectives. Finally, we review common evaluation protocols and discuss open directions for future research. To the best of our knowledge, this paper is the first survey for OOD generalization in GFMs.

Rongxin Cheng, Kai Zhou, Xingda Wei et al.

Rollout dominates the training time in large language model (LLM) post-training, where the trained model is used to generate tokens given a batch of prompts. SpecActor achieves fast rollout with speculative decoding that deploys a fast path (e.g., a smaller model) to accelerate the unparallelizable generation, while the correctness is guaranteed by fast parallel verification of the outputs with the original model. SpecActor addresses two foundational challenges in speculative rollout by (1) a \emph{dynamic decoupled speculation} execution method that maximizes the GPU computational efficiency to realize speedup for large-batch execution -- a configuration common in training but unfriendly to speculative execution and (2) a \emph{dynamic Best-of-N speculation} method that selects and combines different drafting methods according to the rollout progress. It substantially improves the speculation accuracy even when the best drafting method is unknown a priori, meanwhile without requiring adding extra computation resources. {\sys} is {1.3--1.7}\,$\times$ faster than common post-training baselines, and is {1.3--1.5}\,$\times$ faster compared to naively adopting speculative decoding for rollout.

Jiahao Zhou, Chengliang Lin, Dingji Li et al.

Semantic top-K selection with cross-encoder rerankers underpins of on-device AI services, such as retrieval-augmented generation, agent memory, and personalized recommendation. However, its latency and memory demands dominate end-to-end budgets on edge hardware. Revisiting the objective of top-K selection, we reveal that only relative rankings matter, not exact per-candidate scores. We further observe sequence-level sparsity: relative rankings stabilize early in intermediate layers, allowing pruning opportunities prior to completing full inference. Building on this insight, we propose monolithic forwarding and develop a training-free inference system, GRATING. By maintaining a global view of all candidates, it reduces latency through progressive cluster pruning. It also bounds peak memory usage by strategically overlapping I/O with computation via dual-layer sliding window and chunked execution. We evaluate GRATING against state-of-the-art baselines on rerankers from 0.6B to 8B parameters across Apple M2 and RTX 5070. GRATING consistently reduces latency by up to 89.0% and peak memory by up to 94.9% in microbenchmarks, without any loss in precision. Across three real-world on-device AI applications, GRATING lowers latency by 11.6%-51.0% and peak memory by 18.6%-77.8%, demonstrating substantial improvements in efficiency and deployability.

Hongtao Lyu, Boyue Liu, Mingyu Wu et al.

Large language model (LLM) inference systems face a fundamental tension between minimizing Time-to-First-Token (TTFT) latency for new requests and maintaining a high, steady token generation rate (low Time-Per-Output-Token, or TPOT) for ongoing requests. Existing stall-free batching schedulers proposed by Sarathi, while effective at preventing decode stalls, introduce significant computational unfairness. They prioritize decode tasks excessively, simultaneously leading to underutilized decode slack and unnecessary prefill queuing delays, which collectively degrade the system's overall quality of service (QoS). This work identifies the root cause of this unfairness: the non-monotonic nature of Time-Between-Tokens (TBT) as a scheduling metric and the rigid decode-prioritizing policy that fails to adapt to dynamic workload bursts. We therefore propose FairBatching, a novel LLM inference scheduler that enforces fair resource allocation between prefill and decode tasks. It features an adaptive batch capacity determination mechanism, which dynamically adjusts the computational budget to improve the GPU utilization without triggering SLO violations. Its fair and dynamic batch formation algorithm breaks away from the decode-prioritizing paradigm, allowing computation resources to be reclaimed from bursting decode tasks to serve prefill surges, achieving global fairness. Furthermore, FairBatching provides a novel load estimation method, enabling more effective coordination with upper-level schedulers. Implemented and evaluated on realistic traces, FairBatching significantly reduces TTFT tail latency by up to 2.29x while robustly maintaining TPOT SLOs, achieving overall 20.0% improvement in single-node capacity and 54.3% improvement in cluster-level capacity.

Yuan Wang, Mingyu Li, Haibo Chen

Computer-use agents (CUAs) powered by large language models (LLMs) have emerged as a promising approach to automating computer tasks, yet they struggle with graphical user interfaces (GUIs). GUIs, designed for humans, force LLMs to decompose high-level goals into lengthy, error-prone sequences of fine-grained actions, resulting in low success rates and an excessive number of LLM calls. We propose Goal-Oriented Interface (GOI), a novel abstraction that transforms existing GUIs into three declarative primitives: access, state, and observation, which are better suited for LLMs. Our key idea is policy-mechanism separation: LLMs focus on high-level semantic planning (policy) while GOI handles low-level navigation and interaction (mechanism). GOI does not require modifying the application source code or relying on application programming interfaces (APIs). We evaluate GOI with Microsoft Office Suite (Word, PowerPoint, Excel) on Windows. Compared to a leading GUI-based agent baseline, GOI improves task success rates by 67% and reduces interaction steps by 43.5%. Notably, GOI completes over 61% of successful tasks with a single LLM call.

Yixin Song, Zhenliang Xue, Dongliang Wei et al.

While frontier large language models (LLMs) continue to push capability boundaries, their deployment remains confined to GPU-powered cloud infrastructure. We challenge this paradigm with SmallThinker, a family of LLMs natively designed - not adapted - for the unique constraints of local devices: weak computational power, limited memory, and slow storage. Unlike traditional approaches that mainly compress existing models built for clouds, we architect SmallThinker from the ground up to thrive within these limitations. Our innovation lies in a deployment-aware architecture that transforms constraints into design principles. First, We introduce a two-level sparse structure combining fine-grained Mixture-of-Experts (MoE) with sparse feed-forward networks, drastically reducing computational demands without sacrificing model capacity. Second, to conquer the I/O bottleneck of slow storage, we design a pre-attention router that enables our co-designed inference engine to prefetch expert parameters from storage while computing attention, effectively hiding storage latency that would otherwise cripple on-device inference. Third, for memory efficiency, we utilize NoPE-RoPE hybrid sparse attention mechanism to slash KV cache requirements. We release SmallThinker-4B-A0.6B and SmallThinker-21B-A3B, which achieve state-of-the-art performance scores and even outperform larger LLMs. Remarkably, our co-designed system mostly eliminates the need for expensive GPU hardware: with Q4_0 quantization, both models exceed 20 tokens/s on ordinary consumer CPUs, while consuming only 1GB and 8GB of memory respectively. SmallThinker is publicly available at hf.co/PowerInfer/SmallThinker-4BA0.6B-Instruct and hf.co/PowerInfer/SmallThinker-21BA3B-Instruct.

Xin Wang, Zeyang Zhang, Linxin Xiao et al. · tsinghua

Multi-modal graphs, which integrate diverse multi-modal features and relations, are ubiquitous in real-world applications. However, existing multi-modal graph learning methods are typically trained from scratch for specific graph data and tasks, failing to generalize across various multi-modal graph data and tasks. To bridge this gap, we explore the potential of Multi-modal Graph Large Language Models (MG-LLM) to unify and generalize across diverse multi-modal graph data and tasks. We propose a unified framework of multi-modal graph data, task, and model, discovering the inherent multi-granularity and multi-scale characteristics in multi-modal graphs. Specifically, we present five key desired characteristics for MG-LLM: 1) unified space for multi-modal structures and attributes, 2) capability of handling diverse multi-modal graph tasks, 3) multi-modal graph in-context learning, 4) multi-modal graph interaction with natural language, and 5) multi-modal graph reasoning. We then elaborate on the key challenges, review related works, and highlight promising future research directions towards realizing these ambitious characteristics. Finally, we summarize existing multi-modal graph datasets pertinent for model training. We believe this paper can contribute to the ongoing advancement of the research towards MG-LLM for generalization across multi-modal graph data and tasks.

Qingyuan Liu, Liyan Chen, Yanning Yang et al.

Large Language Models (LLMs) increasingly require processing long text sequences, but GPU memory limitations force difficult trade-offs between memory capacity and bandwidth. While HBM-based acceleration offers high bandwidth, its capacity remains constrained. Offloading data to host-side DIMMs improves capacity but introduces costly data swapping overhead. We identify that the critical memory bottleneck lies in the decoding phase of multi-head attention (MHA) exclusively, which demands substantial capacity for storing KV caches and high bandwidth for attention computation. Our key insight reveals this operation uniquely aligns with modern DIMM-based processing-in-memory (PIM) architectures, which offers scalability of both capacity and bandwidth. Based on this observation and insight, we propose L3, a hardware-software co-designed system integrating DIMM-PIM and GPU devices. L3 introduces three innovations: First, hardware redesigns resolve data layout mismatches and computational element mismatches in DIMM-PIM, enhancing LLM inference utilization. Second, communication optimization enables hiding the data transfer overhead with the computation. Third, an adaptive scheduler coordinates GPU-DIMM-PIM operations to maximize parallelism between devices. Evaluations using real-world traces show L3 achieves up to 6.1$\times$ speedup over state-of-the-art HBM-PIM solutions while significantly improving batch sizes.

Zhenliang Xue, Yixin Song, Zeyu Mi et al.

Large language models (LLMs) on smartphones enable real-time AI assistance and privacy-preserving, offline operation. However, resource constraints of smartphones limit current deployments to small language models (SLMs), significantly compromising their capabilities. This paper introduces PowerInfer-2, a smartphone-based framework that enables fast inference for LLMs exceeding the memory capacity. The key insight is decomposing matrix operations into neuron clusters as the basic processing unit, which enables flexible scheduling and efficient I/O-computation pipelining. PowerInfer-2 leverages this neuron-cluster-based design in both computation and storage. For computation, neuron clusters with dense activations are processed on NPU, while sparse clusters use CPU. The storage engine provides a fine-grained pipeline mechanism that coordinates cluster-level computation and I/O operations, enhanced by a segmented neuron cache to reduce I/O activities. PowerInfer-2 achieves up to a 27.8x speed increase compared to state-of-the-art frameworks. PowerInfer-2 is the first system to serve a 47B LLM on a smartphone, achieving 11.68 tokens/s. Notably, these performance improvements preserve model quality with negligible accuracy degradation.

Zhizhong Wang, Lei Zhao, Haibo Chen et al.

Gram-based and patch-based approaches are two important research lines of style transfer. Recent diversified Gram-based methods have been able to produce multiple and diverse stylized outputs for the same content and style images. However, as another widespread research interest, the diversity of patch-based methods remains challenging due to the stereotyped style swapping process based on nearest patch matching. To resolve this dilemma, in this paper, we dive into the crux of existing patch-based methods and propose a universal and efficient module, termed DivSwapper, for diversified patch-based arbitrary style transfer. The key insight is to use an essential intuition that neural patches with higher activation values could contribute more to diversity. Our DivSwapper is plug-and-play and can be easily integrated into existing patch-based and Gram-based methods to generate diverse results for arbitrary styles. We conduct theoretical analyses and extensive experiments to demonstrate the effectiveness of our method, and compared with state-of-the-art algorithms, it shows superiority in diversity, quality, and efficiency.

Zhiwen Zuo, Lei Zhao, Zhizhong Wang et al.

Multimodal image-to-image translation (I2IT) aims to learn a conditional distribution that explores multiple possible images in the target domain given an input image in the source domain. Conditional generative adversarial networks (cGANs) are often adopted for modeling such a conditional distribution. However, cGANs are prone to ignore the latent code and learn a unimodal distribution in conditional image synthesis, which is also known as the mode collapse issue of GANs. To solve the problem, we propose a simple yet effective method that explicitly estimates and maximizes the mutual information between the latent code and the output image in cGANs by using a deep mutual information neural estimator in this paper. Maximizing the mutual information strengthens the statistical dependency between the latent code and the output image, which prevents the generator from ignoring the latent code and encourages cGANs to fully utilize the latent code for synthesizing diverse results. Our method not only provides a new perspective from information theory to improve diversity for I2IT but also achieves disentanglement between the source domain content and the target domain style for free.

Rubing Huang, Weifeng Sun, Yinyin Xu et al.

Random testing (RT) is a well-studied testing method that has been widely applied to the testing of many applications, including embedded software systems, SQL database systems, and Android applications. Adaptive random testing (ART) aims to enhance RT's failure-detection ability by more evenly spreading the test cases over the input domain. Since its introduction in 2001, there have been many contributions to the development of ART, including various approaches, implementations, assessment and evaluation methods, and applications. This paper provides a comprehensive survey on ART, classifying techniques, summarizing application areas, and analyzing experimental evaluations. This paper also addresses some misconceptions about ART, and identifies open research challenges to be further investigated in the future work.

Zhizhong Wang, Lei Zhao, Haibo Chen et al.

Image style transfer is an underdetermined problem, where a large number of solutions can satisfy the same constraint (the content and style). Although there have been some efforts to improve the diversity of style transfer by introducing an alternative diversity loss, they have restricted generalization, limited diversity and poor scalability. In this paper, we tackle these limitations and propose a simple yet effective method for diversified arbitrary style transfer. The key idea of our method is an operation called deep feature perturbation (DFP), which uses an orthogonal random noise matrix to perturb the deep image feature maps while keeping the original style information unchanged. Our DFP operation can be easily integrated into many existing WCT (whitening and coloring transform)-based methods, and empower them to generate diverse results for arbitrary styles. Experimental results demonstrate that this learning-free and universal method can greatly increase the diversity while maintaining the quality of stylization.

Dong Cao, Lisha Xu, HaiBo Chen

With the rapid development of deep learning algorithms, action recognition in video has achieved many important research results. One issue in action recognition, Zero-Shot Action Recognition (ZSAR), has recently attracted considerable attention, which classify new categories without any positive examples. Another difficulty in action recognition is that untrimmed data may seriously affect model performance. We propose a composite two-stream framework with a pre-trained model. Our proposed framework includes a classifier branch and a composite feature branch. The graph network model is adopted in each of the two branches, which effectively improves the feature extraction and reasoning ability of the framework. In the composite feature branch, a 3-channel self-attention models are constructed to weight each frame in the video and give more attention to the key frames. Each self-attention models channel outputs a set of attention weights to focus on a particular aspect of the video, and a set of attention weights corresponds to a one-dimensional vector. The 3-channel self-attention models can evaluate key frames from multiple aspects, and the output sets of attention weight vectors form an attention matrix, which effectively enhances the attention of key frames with strong correlation of action. This model can implement action recognition under zero-shot conditions, and has good recognition performance for untrimmed video data. Experimental results on relevant data sets confirm the validity of our model.

Dong Cao, Dongdong Zhang, HaiBo Chen

Millions of people with severe speech disorders around the world may regain their communication capabilities through techniques of silent speech recognition (SSR). Using electroencephalography (EEG) as a biomarker for speech decoding has been popular for SSR. However, the lack of SSR text corpus has impeded the development of this technique. Here, we construct a novel task-oriented text corpus, which is utilized in the field of SSR. In the process of construction, we propose a task-oriented hybrid construction method based on natural language generation algorithm. The algorithm focuses on the strategy of data-to-text generation, and has two advantages including linguistic quality and high diversity. These two advantages use template-based method and deep neural networks respectively. In an SSR experiment with the generated text corpus, analysis results show that the performance of our hybrid construction method outperforms the pure method such as template-based natural language generation or neural natural language generation models.

Chuzhe Tang, Zhiyuan Dong, Minjie Wang et al.

The recent proposal of learned index structures opens up a new perspective on how traditional range indexes can be optimized. However, the current learned indexes assume the data distribution is relatively static and the access pattern is uniform, while real-world scenarios consist of skew query distribution and evolving data. In this paper, we demonstrate that the missing consideration of access patterns and dynamic data distribution notably hinders the applicability of learned indexes. To this end, we propose solutions for learned indexes for dynamic workloads (called Doraemon). To improve the latency for skew queries, Doraemon augments the training data with access frequencies. To address the slow model re-training when data distribution shifts, Doraemon caches the previously-trained models and incrementally fine-tunes them for similar access patterns and data distribution. Our preliminary result shows that, Doraemon improves the query latency by 45.1% and reduces the model re-training time to 1/20.