Yun Ma

Ningyuan Li, Haiyang Shen, Mugeng Liu et al.

Recent advances in large language models and tool-using agents have expanded the range of benchmarked web tasks. Yet an important class of specialized retrieval tasks remains undercharacterized. On many specialized data-retrieval websites, answer-bearing evidence becomes accessible only after establishing the correct site-specific retrieval state through filters, views, hierarchies, or scopes. We term this capability state-gated retrieval (SGR). We introduce SGR-Bench, a benchmark for this setting containing 100 expert-curated tasks spanning six source families and 12 public data ecosystems. Each task requires discovering the appropriate website and configuring its site-specific retrieval state to produce a structured answer. SGR-Bench pairs constraint-guided and goal-oriented formulations of the same underlying problems, enabling controlled comparisons between explicit and implicit guidance for state-gated retrieval. We evaluate eight CLI-based agentic LLM systems and three commercial search-agent products. On SGR-Bench, the strongest system reaches only 66.18% item-level F1, while row-level F1 remains much lower. A manual audit of 156 analyzable failed CLI trajectories shows why: agents often reach a relevant web source, but establish the wrong site-specific retrieval state. Retrieval-scope drift (37.2%) and criterion mismatch (27.6%) dominate, whereas final answer composition accounts for only 10.3%. The dataset and single-case evaluation instructions are available at https://huggingface.co/datasets/PKUAIWeb/SGR-BENCH.

Haiyang Shen, Taian Guo, Xuanzhong Chen et al.

Although LLMs have made substantial progress in reasoning, systematically producing frontier-level reasoning data remains difficult. Existing synthesis methods often have limited visibility into the structural factors that govern problem difficulty, which can result in narrow diversity and unstable difficulty control. In this work, we view the difficulty of a reasoning problem as arising from the accumulation of atomic knowledge-reasoning transformations, which we term thought modes. Building on this perspective, we propose MindLoom, a framework for synthesizing frontier-level reasoning data through compositional thought mode engineering. Given a collection of hard problems with verified solutions, MindLoom first decomposes those solutions into thought mode chains that reveal each problem's construction logic. It then trains a retrieval model that matches problem states to compatible thought modes, providing guidance on which reasoning challenges to introduce during synthesis. New problems are composed by iteratively applying retrieved thought modes to seed questions, with distribution-aligned sampling to encourage diverse reasoning coverage. Finally, a rollout-based judging stage labels generated questions by difficulty and supplies judged-correct responses for supervised fine-tuning. We evaluate MindLoom on nine benchmarks covering five STEM disciplines and four mathematical reasoning tasks across multiple model families and sizes. Models fine-tuned on MindLoom-generated data achieves favorable performances over base models, distillation, and external-data baselines across the reported benchmarks. Ablation studies indicate the contribution of each component, and further analysis suggests that MindLoom covers a broad range of reasoning patterns while maintaining useful difficulty control. We have open-sourced our implementation at https://github.com/EachSheep/MindLoom.

Haiyang Shen, Jiuzheng Wang, Taian Guo et al.

As AI becomes part of everyday learning, many courses teach students to use it mainly as a productivity tool: how to prompt, search, summarize, write, code, and use tools more efficiently. We argue that AI education also needs a setting in which students learn to test AI and understand their own role in judging machine-produced knowledge. To this end, we introduce a course-based practice that teaches AI through benchmark construction, using deep research systems as a concrete example of AI-era knowledge work. Students turn disciplinary knowledge into verifiable expert-level questions, review one another's designs for ambiguity and shortcuts, and evaluate AI systems on the resulting tasks. This activity gives students direct exposure to a powerful tool while asking them to specify what a trustworthy answer would require. The produced benchmark, QuestBench, consists of 256 questions across 14 humanities and social-science domains. Evaluation on QuestBench shows that student-designed tasks reveal hidden failures in current deep research systems: across thirteen evaluated systems, the mean question-level pass rate is only 16.85%, and the best-performing system, GPT-5.5, reaches a 57.58% pass rate. The failures are educationally useful because they show how fluent, source-backed answers can still miss the right query, source, term, or evidence standard. Reflections from five student contributors suggest that benchmark construction can help students see professional knowledge not only as content AI may retrieve, but as the basis for judging AI outputs. We present QuestBench as a benchmark artifact and as a reusable classroom setting for a larger educational question: how students can remain responsible knowledge actors as AI enters learning and professional work. The dataset is available at https://huggingface.co/datasets/PKUAIWeb/QuestBench/tree/main.

Taian Guo, Haiyang Shen, Junyu Luo et al.

LLMs have demonstrated significant potential in quantitative finance by processing vast unstructured data to emulate human-like analytical workflows. However, current LLM-based methods primarily follow either an Asset-Centric paradigm focused on individual stock prediction or a Market-Centric approach for portfolio allocation, often remaining agnostic to the underlying reasoning that drives market movements. In this paper, we propose a Logic-Oriented perspective, modeling the financial market as a dynamic, evolutionary ecosystem of competing investment narratives, termed Modes of Thought. To operationalize this view, we introduce MEME (Modeling the Evolutionary Modes of Financial Markets), designed to reconstruct market dynamics through the lens of evolving logics. MEME employs a multi-agent extraction module to transform noisy data into high-fidelity Investment Arguments and utilizes Gaussian Mixture Modeling to uncover latent consensus within a semantic space. To model semantic drift among different market conditions, we also implement a temporal evaluation and alignment mechanism to track the lifecycle and historical profitability of these modes. By prioritizing enduring market wisdom over transient anomalies, MEME ensures that portfolio construction is guided by robust reasoning. Extensive experiments on three heterogeneous Chinese stock pools from 2023 to 2025 demonstrate that MEME consistently outperforms seven SOTA baselines. Further ablation studies, sensitivity analysis, lifecycle case study and cost analysis validate MEME's capacity to identify and adapt to the evolving consensus of financial markets. Our implementation can be found at https://github.com/gta0804/MEME.

Taian Guo, Haiyang Shen, Junyu Luo et al.

Extracting signals through alpha factor mining is a fundamental challenge in quantitative finance. Existing automated methods primarily follow two paradigms: Decoupled Factor Generation, which treats factor discovery as isolated events, and Iterative Factor Evolution, which focuses on local parent-child refinements. However, both paradigms lack a global structural view, often treating factor pools as unstructured collections or fragmented chains, which leads to redundant search and limited diversity. To address these limitations, we introduce AlphaPROBE (Alpha Mining via Principled Retrieval and On-graph Biased Evolution), a framework that reframes alpha mining as the strategic navigation of a Directed Acyclic Graph (DAG). By modeling factors as nodes and evolutionary links as edges, AlphaPROBE treats the factor pool as a dynamic, interconnected ecosystem. The framework consists of two core components: a Bayesian Factor Retriever that identifies high-potential seeds by balancing exploitation and exploration through a posterior probability model, and a DAG-aware Factor Generator that leverages the full ancestral trace of factors to produce context-aware, nonredundant optimizations. Extensive experiments on three major Chinese stock market datasets against 8 competitive baselines demonstrate that AlphaPROBE significantly gains enhanced performance in predictive accuracy, return stability and training efficiency. Our results confirm that leveraging global evolutionary topology is essential for efficient and robust automated alpha discovery. We have open-sourced our implementation at https://github.com/gta0804/AlphaPROBE.

Haifeng Wang, Hua Wu, Tian Wu et al.

In this report, we introduce ERNIE 5.0, a natively autoregressive foundation model desinged for unified multimodal understanding and generation across text, image, video, and audio. All modalities are trained from scratch under a unified next-group-of-tokens prediction objective, based on an ultra-sparse mixture-of-experts (MoE) architecture with modality-agnostic expert routing. To address practical challenges in large-scale deployment under diverse resource constraints, ERNIE 5.0 adopts a novel elastic training paradigm. Within a single pre-training run, the model learns a family of sub-models with varying depths, expert capacities, and routing sparsity, enabling flexible trade-offs among performance, model size, and inference latency in memory- or time-constrained scenarios. Moreover, we systematically address the challenges of scaling reinforcement learning to unified foundation models, thereby guaranteeing efficient and stable post-training under ultra-sparse MoE architectures and diverse multimodal settings. Extensive experiments demonstrate that ERNIE 5.0 achieves strong and balanced performance across multiple modalities. To the best of our knowledge, among publicly disclosed models, ERNIE 5.0 represents the first production-scale realization of a trillion-parameter unified autoregressive model that supports both multimodal understanding and generation. To facilitate further research, we present detailed visualizations of modality-agnostic expert routing in the unified model, alongside comprehensive empirical analysis of elastic training, aiming to offer profound insights to the community.

Zhuofan Shi, Peilun Jia, Baoqin Sun et al.

Recent deep research systems have improved the ability of large language models to produce long, grounded reports through iterative retrieval and reasoning. However, most text-centered systems rely mainly on textual evidence, while multimodal systems often retrieve images only weakly or generate charts themselves, leaving source figures underused as evidence. We present ViDR, a multimodal deep research framework that grounds long-form reports in source figures. ViDR treats source figures as retrievable, interpretable, routable, and verifiable evidence objects, while still generating analytical charts when needed. It builds an evidence-indexed outline linking claims to textual and visual evidence, refines noisy web images into source-figure evidence atoms through context-aware filtering, outline-aware reranking, and VLM-based visual analysis, and generates each section with section-specific evidence. ViDR further validates visual references to reduce hallucinated or misplaced figures. We also introduce MMR Bench+, a benchmark for evaluating visual evidence use in deep research reports, covering source-figure retrieval, placement, interpretation, verifiability, and analytical chart generation. Experiments show that ViDR improves overall report quality, source-figure integration, and verifiability over strong commercial and open-source baselines. These results suggest that source visual evidence is important for multimodal deep research, as it strengthens evidential grounding, visual support, and report verifiability.

Jialiang Han, Yudong Han, Gang Huang et al.

Federated learning (FL) is an emerging promising paradigm of privacy-preserving machine learning (ML). An important type of FL is cross-silo FL, which enables a small scale of organizations to cooperatively train a shared model by keeping confidential data locally and aggregating weights on a central parameter server. However, the central server may be vulnerable to malicious attacks or software failures in practice. To address this issue, in this paper, we propose DeFL, a novel decentralized weight aggregation framework for cross-silo FL. DeFL eliminates the central server by aggregating weights on each participating node and weights of only the current training round are maintained and synchronized among all nodes. We use Multi-Krum to enable aggregating correct weights from honest nodes and use HotStuff to ensure the consistency of the training round number and weights among all nodes. Besides, we theoretically analyze the Byzantine fault tolerance, convergence, and complexity of DeFL. We conduct extensive experiments over two widely-adopted public datasets, i.e. CIFAR-10 and Sentiment140, to evaluate the performance of DeFL. Results show that DeFL defends against common threat models with minimal accuracy loss, and achieves up to 100x reduction in storage overhead and up to 12x reduction in network overhead, compared to state-of-the-art decentralized FL approaches.

Haiyang Shen, Xuanzhong Chen, Wendong Xu et al.

Coding agents are increasingly used as iterative development partners, but most benchmarks still evaluate one specification followed by one final assessment. This leaves out a basic question: can an agent keep its own codebase working as requirements change? We introduce EvoCode-Bench, a benchmark of 26 stateful coding tasks and 227 evaluated rounds. Each task preserves the agent's workspace for 5-15 rounds, states requirements through observable behavior, and uses cumulative executable tests to check new requirements and still-active prior ones. We evaluate 13 coding agents with two metrics: MT@4, a four-attempt fail-stop multi-round score, and SR, a single-round score from a reference-completed prior state. For most agents, SR exceeds MT@4 by 22-40 points. The gap also changes rankings: the highest-SR agent (78.9) ranks only third in persistent execution (44.0 MT@4). Even the strongest agents achieve only about 50% success on multi-turn metrics, and aggregate pass rate drops below half of round-1 performance by round 5. Failure analysis shows tier-dependent behavior: weaker agents fail early, while stronger agents survive long enough to expose specification-tracking and regression failures. We release the benchmark data and Harbor multi-turn infrastructure.

Sixiong Xie, Zhuofan Shi, Haiyang Shen et al.

Deep research, in which an agent searches the open web, collects evidence, and derives an answer through extended reasoning, is a prominent use case for frontier language models. Frontier deep research products score high on existing benchmarks, making it difficult to distinguish their capabilities from current evaluation data alone. We introduce DeepWeb-Bench, a deep research benchmark that is substantially harder than existing benchmarks for the current frontier. Difficulty comes from three properties of the data itself: each task requires massive evidence collection, cross-source reconciliation, and long-horizon multi-step derivation. We represent these three sources of difficulty as four capability families (Retrieval, Derivation, Reasoning, and Calibration) and report results sliced by family. Every reference answer is accompanied by a source-provenance record with four disclosure levels and cross-source checks where available, making scores easier to audit against the underlying evidence. We evaluate DeepWeb-Bench on nine frontier models and report three findings: (1) retrieval is not the bottleneck, as retrieval failures account for only 12-14% of errors while derivation and calibration failures account for over 70%; (2) strong and weak models fail in qualitatively different ways, with strong models' errors dominated by incomplete derivation and weak models' by hallucinated precision; and (3) models exhibit genuine specialization across domains, with cross-model agreement of only rho = 0.61 and per-case disagreement reaching 18.8 percentage points. The public benchmark release includes the data, rubrics, and evaluation code.

Jinghe Zhang, Daliang Xu, Chenghua Wang et al.

Large language models (LLMs) are increasingly deployed on mobile devices, where Neural Processing Units (NPUs) necessitate fully static quantization for optimal inference efficiency. However, existing post-training quantization (PTQ) methods predominantly rely on dynamic activation quantization, rendering them incompatible with NPU hardware constraints. To bridge the gap between high-fidelity PTQ and NPU-constrained inference, we propose Quant.npu, a integer-only fully static quantization framework. It incorporates learnable quantization parameters and rotation matrices, enabling low-bit activation-weight quantization without runtime quantization parameters re-computation. Crucially, we identify that initialization and selective optimization of quantization parameters is pivotal for optimization stability, as improper initialization and naive joint optimization induce gradient instability that disrupts the optimization of rotation matrices. To address this, we propose a rotation-and-bit-width-aware initialization tailored to diverse activation profiles and a distribution-aware selective optimization (two-stage quantization pipeline) tailored to rotated and unrotated tensors. Furthermore, we introduce a sensitivity-guided adaptive mixed-precision scheme to balance accuracy with inference efficiency. Extensive experiments on real-world mobile NPUs demonstrate that Quant.npu achieves comparable accuracy to state-of-the-art methods, while reducing inference latency by up to 15.1%.

Kaibo Liu, Zhenpeng Chen, Yiyang Liu et al. · pku

Detecting tricky bugs in plausible programs, those that pass existing test suites yet still contain bugs, remains a significant challenge in software testing. To address this problem, we propose TrickCatcher, an LLM-powered approach to generating test cases for uncovering bugs in plausible programs. TrickCatcher operates in three stages: First, it uses an LLM to generate program variants based on the program under test (PUT) and its specification. Second, it employs an LLM to construct an input generator from the specification for producing test inputs. Finally, these inputs are executed on both the PUT and its program variants to detect inconsistencies in their outputs. We evaluate TrickCatcher on two datasets, TrickyBugs and EvalPlus, which include 366 human-written and 151 AI-generated plausible programs with tricky bugs. TrickCatcher achieves recall, precision, and F1 scores that are 1.80x, 2.65x, and 1.66x those of the state-of-the-art baselines, respectively. Code and data used are available at https://github.com/RinCloud/TrickCatcher.

Taian Guo, Haiyang Shen, JinSheng Huang et al. · pku

The application of LLM-based agents in financial investment has shown significant promise, yet existing approaches often require intermediate steps like predicting individual stock movements or rely on predefined, static workflows. These limitations restrict their adaptability and effectiveness in constructing optimal portfolios. In this paper, we introduce the Multi-Agent Scaling Simulation (MASS), a novel framework that leverages multi-agent simulation for direct, end-to-end portfolio construction. At its core, MASS employs a backward optimization process to dynamically learn the optimal distribution of heterogeneous agents, enabling the system to adapt to evolving market regimes. A key finding enabled by our framework is the exploration of the scaling effect for portfolio construction: we demonstrate that as the number of agents increases exponentially (up to 512), the aggregated decisions yield progressively higher excess returns. Extensive experiments on a challenging, self-collected dataset from the 2023 Chinese A-share market show that MASS consistently outperforms seven state-of-the-art baselines. Further backtesting, stability analyses and the experiment on data leakage concerns validate its enhanced profitability and robustness. We have open-sourced our code, dataset, and training snapshots at https://github.com/gta0804/MASS/ to foster further research.

Yaoqi Guo, Zhenpeng Chen, Jie M. Zhang et al.

Code generation, the automatic creation of source code from natural language descriptions, has garnered significant attention due to its potential to streamline software development. Inspired by research that links task-personality alignment with improved development outcomes, we conduct an empirical study on personality-guided code generation using large language models (LLMs). Specifically, we investigate how emulating personality traits appropriate to the coding tasks affects LLM performance. We extensively evaluate this approach using seven widely adopted LLMs across four representative datasets. Our results show that personality guidance significantly enhances code generation accuracy, with improved pass rates in 23 out of 28 LLM-dataset combinations. Notably, in 11 cases, the improvement exceeds 5%, and in 5 instances, it surpasses 10%, with the highest gain reaching 12.9%. Additionally, personality guidance can be easily integrated with other prompting strategies to further boost performance. We open-source our code and data at https://github.com/IanWalls/Persona-Code.

Zhuofan Shi, Zijie Guo, Xinjian Ma et al.

The rapid growth of multi-source, heterogeneous, and multimodal scientific data has increasingly exposed the limitations of traditional data management. Most existing DeepResearch (DR) efforts focus primarily on web search while overlooking local private data. Consequently, these frameworks exhibit low retrieval efficiency for private data and fail to comply with the FAIR principles, ultimately resulting in inefficiency and limited reusability. To this end, we propose IoDResearch (Internet of Data Research), a private data-centric Deep Research framework that operationalizes the Internet of Data paradigm. IoDResearch encapsulates heterogeneous resources as FAIR-compliant digital objects, and further refines them into atomic knowledge units and knowledge graphs, forming a heterogeneous graph index for multi-granularity retrieval. On top of this representation, a multi-agent system supports both reliable question answering and structured scientific report generation. Furthermore, we establish the IoD DeepResearch Benchmark to systematically evaluate both data representation and Deep Research capabilities in IoD scenarios. Experimental results on retrieval, QA, and report-writing tasks show that IoDResearch consistently surpasses representative RAG and Deep Research baselines. Overall, IoDResearch demonstrates the feasibility of private-data-centric Deep Research under the IoD paradigm, paving the way toward more trustworthy, reusable, and automated scientific discovery.

Yudong Han, Chao Xu, Xiaodan Ye et al.

We present WebSplatter, an end-to-end GPU rendering pipeline for the heterogeneous web ecosystem. Unlike naive ports, WebSplatter introduces a wait-free hierarchical radix sort that circumvents the lack of global atomics in WebGPU, ensuring deterministic execution across diverse hardware. Furthermore, we propose an opacity-aware geometry culling stage that dynamically prunes splats before rasterization, significantly reducing overdraw and peak memory footprint. Evaluation demonstrates that WebSplatter consistently achieves 1.2$\times$ to 4.5$\times$ speedups over state-of-the-art web viewers.

Zijian Shao, Haiyang Shen, Mugeng Liu et al.

Effective disease prediction in modern healthcare demands the twin goals of high accuracy and transparent, clinically meaningful explanations. Existing machine learning and large language model (LLM) based approaches often struggle to balance these goals. Many models yield accurate but unclear statistical outputs, while others generate fluent but statistically unsupported narratives, often undermining both the validity of the explanation and the predictive accuracy itself. This shortcoming comes from a shallow interaction with the data, preventing the development of a deep, detailed understanding similar to a human expert's. We argue that high accuracy and high-quality explanations are not separate objectives but are mutually reinforcing outcomes of a model that develops a deep, direct understanding of the data. To achieve this, we propose the Reflective Cognitive Architecture (RCA), a novel framework that coordinates multiple LLMs to learn from direct experience. RCA features an iterative rule refinement mechanism that improves its logic from prediction errors and a distribution-aware rules check mechanism that bases its reasoning in the dataset's global statistics. By using predictive accuracy as a signal to drive deeper comprehension, RCA builds a strong internal model of the data. We evaluated RCA on one private and two public datasets against 22 baselines. The results demonstrate that RCA not only achieves state-of-the-art accuracy and robustness with a relative improvement of up to 40\% over the baseline but, more importantly, leverages this deep understanding to excel in generating explanations that are clear, logical, evidence-based, and balanced, highlighting its potential for creating genuinely trustworthy clinical decision support systems. The code is available at \https://github.com/ssssszj/RCA.

Haiyang Shen, Yue Li, Desong Meng et al.

Recent advancements in integrating large language models (LLMs) with application programming interfaces (APIs) have gained significant interest in both academia and industry. Recent work demonstrates that these API-based agents exhibit relatively strong autonomy and planning capabilities. However, their ability to handle multi-dimensional difficulty levels, diverse task types, and real-world demands remains unknown. In this paper, we introduce \textsc{ShortcutsBench}, a large-scale benchmark for the comprehensive evaluation of API-based agents in solving real-world complex tasks. \textsc{ShortcutsBench} includes a wealth of real APIs from Apple Inc., refined user queries, human-annotated high-quality action sequences, detailed parameter filling values, and parameters requesting necessary input from the system or user. We revealed how existing benchmarks~/~datasets struggle to accommodate the advanced reasoning capabilities of existing more intelligent LLMs. Moreover, our extensive evaluation of agents built with $5$ leading open-source (size $\geq$ 57B) and $5$ closed-source LLMs (e.g. Gemini-1.5-Pro and GPT-4o-mini) with varying intelligence level reveals significant limitations of existing API-based agents in the whole process of handling complex queries related to API selection, parameter filling, and requesting necessary input from the system and the user. These findings highlight the great challenges that API-based agents face in effectively fulfilling real and complex user queries. All datasets, code, experimental logs, and results are available at \url{https://github.com/EachSheep/ShortcutsBench}.

Xudong Mao, Yun Ma, Zhenguo Yang et al.

We study the problem of unsupervised domain adaptation which aims to adapt models trained on a labeled source domain to a completely unlabeled target domain. Recently, the cluster assumption has been applied to unsupervised domain adaptation and achieved strong performance. One critical factor in successful training of the cluster assumption is to impose the locally-Lipschitz constraint to the model. Existing methods only impose the locally-Lipschitz constraint around the training points while miss the other areas, such as the points in-between training data. In this paper, we address this issue by encouraging the model to behave linearly in-between training points. We propose a new regularization method called Virtual Mixup Training (VMT), which is able to incorporate the locally-Lipschitz constraint to the areas in-between training data. Unlike the traditional mixup model, our method constructs the combination samples without using the label information, allowing it to apply to unsupervised domain adaptation. The proposed method is generic and can be combined with most existing models such as the recent state-of-the-art model called VADA. Extensive experiments demonstrate that VMT significantly improves the performance of VADA on six domain adaptation benchmark datasets. For the challenging task of adapting MNIST to SVHN, VMT can improve the accuracy of VADA by over 30\%. Code is available at \url{https://github.com/xudonmao/VMT}.

Sixiong Xie, Zhuofan Shi, Haiyang Shen et al.

As the capabilities of large language model (LLM) agents continue to advance, their advanced social behaviors, such as cooperation, deception, and collusion, call for systematic evaluation. However, existing benchmarks often emphasize a single capability dimension or rely solely on behavioral outcomes, overlooking rich process information from agents' decision reasoning and communicative interactions. To address this gap, we propose M3-Bench, a multi-stage benchmark for mixed-motive games, together with a process-aware evaluation framework that conducts synergistic analysis across three modules: BTA (Behavioral Trajectory Analysis), RPA (Reasoning Process Analysis), and CCA (Communication Content Analysis). Furthermore, we integrate the Big Five personality model and Social Exchange Theory to aggregate multi-dimensional evidence into interpretable social behavior portraits, thereby characterizing agents' personality traits and capability profiles beyond simple task scores or outcome-based metrics. Experimental results show that M3-Bench can reliably distinguish diverse social behavior competencies across models, and it reveals that some models achieve seemingly reasonable behavioral outcomes while exhibiting pronounced inconsistencies in their reasoning and communication.

Yun Ma, Yihong Wu, Pengkun Yang

We consider the problem of approximating a general Gaussian location mixture by finite mixtures. The minimum order of finite mixtures that achieve a prescribed accuracy (measured by various $f$-divergences) is determined within constant factors for the family of mixing distributions with compactly support or appropriate assumptions on the tail probability including subgaussian and subexponential. While the upper bound is achieved using the technique of local moment matching, the lower bound is established by relating the best approximation error to the low-rank approximation of certain trigonometric moment matrices, followed by a refined spectral analysis of their minimum eigenvalue. In the case of Gaussian mixing distributions, this result corrects a previous lower bound in [Allerton Conference 48 (2010) 620-628].

Qipeng Wang, Shiqi Jiang, Zhenpeng Chen et al.

Web applications have increasingly adopted Deep Learning (DL) through in-browser inference, wherein DL inference performs directly within Web browsers. The actual performance of in-browser inference and its impacts on the quality of experience (QoE) remain unexplored, and urgently require new QoE measurements beyond traditional ones, e.g., mainly focusing on page load time. To bridge this gap, we make the first comprehensive performance measurement of in-browser inference to date. Our approach proposes new metrics to measure in-browser inference: responsiveness, smoothness, and inference accuracy. Our extensive analysis involves 9 representative DL models across Web browsers of 50 popular PC devices and 20 mobile devices. The results reveal that in-browser inference exhibits a substantial latency gap, averaging 16.9 times slower on CPU and 4.9 times slower on GPU compared to native inference on PC devices. The gap on mobile CPU and mobile GPU is 15.8 times and 7.8 times, respectively. Furthermore, we identify contributing factors to such latency gap, including underutilized hardware instruction sets, inherent overhead in the runtime environment, resource contention within the browser, and inefficiencies in software libraries and GPU abstractions. Additionally, in-browser inference imposes significant memory demands, at times exceeding 334.6 times the size of the DL models themselves, partly attributable to suboptimal memory management. We also observe that in-browser inference leads to a significant 67.2% increase in the time it takes for GUI components to render within Web browsers, significantly affecting the overall user QoE of Web applications reliant on this technology

Zhiyang Chen, Daliang Xu, Yinyuan Zhang et al.

Large language models typically employ vocabularies of over 100k tokens, which creates a major computational bottleneck at the final linear projection layer when performing speculative decoding. Current methods for vocabulary pruning depend on either fixed or coarse-grained sub-vocabularies, requiring around 30k active tokens to preserve the quality of the draft model. We introduce MicroSpec, a training-free technique that overcomes this limitation by building a compact, context-sensitive active vocabulary on the fly for every decoding step. Exploiting the natural temporal locality found in language generation, MicroSpec attains high token coverage while reducing the average vocabulary size by more than 40x (down to under 3k tokens), all without any additional trained parameters. To translate this high sparsity into actual speedups on contemporary hardware, we present a co-designed system and algorithm that mitigates the overhead of sparse memory accesses via asynchronous gathering and GPU-resident state management. Acting as a plug-and-play enhancement, MicroSpec reduces draft inference latency by 51.6% on average, achieving an end-to-end speedup of 1.12-1.32x relative to the leading speculative decoding approach EAGLE-2 on various benchmarks, while also surpassing more sophisticated training-based pruning baselines.

Zhiyang Chen, Daliang Xu, Haiyang Shen et al.

Enhancing on-device large language models (LLMs) with contextual information from local data enables personalized and task-aware generation, powering use cases such as intelligent assistants and UI agents. While recent developments in neural processors have substantially improved the efficiency of prefill on mobile devices, the token-by-token generation process still suffers from high latency and limited hardware utilization due to its inherently memory-bound characteristics. This work presents sd.npu, a mobile inference framework that integrates speculative decoding with dynamic hardware scheduling to accelerate context-aware text generation on mobile devices. The framework introduces three synergistic components: (1) adaptive execution scheduling, which dynamically balances compute graphs between prefill and decoding phases; (2) context-aligned drafting, which improves speculative efficiency through lightweight online calibration to current tasks; and (3) hardware-efficient draft extension, which reuses and expands intermediate sequences to improve processing parallelism and reduce verification cost. Experiments on multiple smartphones and representative workloads show consistent improvements of up to 3.8x in generation speed and 4.7x in energy efficiency compared with existing mobile inference solutions. Component-level analysis further validates the contribution of each optimization.

Qi Yang, Weichen Bi, Haiyang Shen et al.

Graphical User Interface (GUI) datasets are crucial for various downstream tasks. However, GUI datasets often generate annotation information through automatic labeling, which commonly results in inaccurate GUI element BBox annotations, including missing, duplicate, or meaningless BBoxes. These issues can degrade the performance of models trained on these datasets, limiting their effectiveness in real-world applications. Additionally, existing GUI datasets only provide BBox annotations visually, which restricts the development of visually related GUI downstream tasks. To address these issues, we introduce PixelWeb, a large-scale GUI dataset containing over 100,000 annotated web pages. PixelWeb is constructed using a novel automatic annotation approach that integrates visual feature extraction and Document Object Model (DOM) structure analysis through two core modules: channel derivation and layer analysis. Channel derivation ensures accurate localization of GUI elements in cases of occlusion and overlapping elements by extracting BGRA four-channel bitmap annotations. Layer analysis uses the DOM to determine the visibility and stacking order of elements, providing precise BBox annotations. Additionally, PixelWeb includes comprehensive metadata such as element images, contours, and mask annotations. Manual verification by three independent annotators confirms the high quality and accuracy of PixelWeb annotations. Experimental results on GUI element detection tasks show that PixelWeb achieves performance on the mAP95 metric that is 3-7 times better than existing datasets. We believe that PixelWeb has great potential for performance improvement in downstream tasks such as GUI generation and automated user interaction.

Qiyang Zhang, Xiang Li, Xiangying Che et al.

Deploying deep learning (DL) on mobile devices has been a notable trend in recent years. To support fast inference of on-device DL, DL libraries play a critical role as algorithms and hardware do. Unfortunately, no prior work ever dives deep into the ecosystem of modern DL libs and provides quantitative results on their performance. In this paper, we first build a comprehensive benchmark that includes 6 representative DL libs and 15 diversified DL models. We then perform extensive experiments on 10 mobile devices, which help reveal a complete landscape of the current mobile DL libs ecosystem. For example, we find that the best-performing DL lib is severely fragmented across different models and hardware, and the gap between those DL libs can be rather huge. In fact, the impacts of DL libs can overwhelm the optimizations from algorithms or hardware, e.g., model quantization and GPU/DSP-based heterogeneous computing. Finally, atop the observations, we summarize practical implications to different roles in the DL lib ecosystem.

Jialiang Han, Yun Ma, Yudong Han

Federated learning (FL) is an emerging promising privacy-preserving machine learning paradigm and has raised more and more attention from researchers and developers. FL keeps users' private data on devices and exchanges the gradients of local models to cooperatively train a shared Deep Learning (DL) model on central custodians. However, the security and fault tolerance of FL have been increasingly discussed, because its central custodian mechanism or star-shaped architecture can be vulnerable to malicious attacks or software failures. To address these problems, Swarm Learning (SL) introduces a permissioned blockchain to securely onboard members and dynamically elect the leader, which allows performing DL in an extremely decentralized manner. Compared with tremendous attention to SL, there are few empirical studies on SL or blockchain-based decentralized FL, which provide comprehensive knowledge of best practices and precautions of deploying SL in real-world scenarios. Therefore, we conduct the first comprehensive study of SL to date, to fill the knowledge gap between SL deployment and developers, as far as we are concerned. In this paper, we conduct various experiments on 3 public datasets of 5 research questions, present interesting findings, quantitatively analyze the reasons behind these findings, and provide developers and researchers with practical suggestions. The findings have evidenced that SL is supposed to be suitable for most application scenarios, no matter whether the dataset is balanced, polluted, or biased over irrelevant features.

Xuanzhe Liu, Diandian Gu, Zhenpeng Chen et al.

Deep learning (DL) has become a key component of modern software. In the "big model" era, the rich features of DL-based software substantially rely on powerful DL models, e.g., BERT, GPT-3, and the recently emerging GPT-4, which are trained on the powerful cloud with large datasets. Hence, training effective DL models has become a vital stage in the whole software lifecycle. When training deep learning models, especially those big models, developers need to parallelize and distribute the computation and memory resources amongst multiple devices in the training process, which is known as distributed deep learning training, or distributed training for short. However, the unique challenges that developers encounter in distributed training process have not been studied in the software engineering community. Given the increasingly heavy dependence of current DL-based software on distributed training, this paper aims to fill in the knowledge gap and presents the first comprehensive study on developers' issues in distributed training. To this end, we analyze 1,131 real-world developers' issues about using these frameworks reported on Stack Overflow and GitHub. We construct a fine-grained taxonomy consisting of 30 categories regarding the fault symptoms and summarize common fix patterns for different symptoms. Based on the results, we suggest actionable implications on research avenues that can potentially facilitate the distributed training to develop DL-based software, such as focusing on the frequent and common fix patterns when designing testing or debugging tools, developing efficient testing and debugging techniques for communication configuration along with the synthesis of network configuration analysis, designing new multi-device checkpoint-and-replay techniques to help reproduction, and designing serverless APIs for cloud platforms.

Yanchao Tan, Carl Yang, Xiangyu Wei et al.

Implicit feedback is widely explored by modern recommender systems. Since the feedback is often sparse and imbalanced, it poses great challenges to the learning of complex interactions among users and items. Metric learning has been proposed to capture user-item interactions from implicit feedback, but existing methods only represent users and items in a single metric space, ignoring the fact that users can have multiple preferences and items can have multiple properties, which leads to potential conflicts limiting their performance in recommendation. To capture the multiple facets of user preferences and item properties while resolving their potential conflicts, we propose the novel framework of Multi-fAcet Recommender networks with Spherical optimization (MARS). By designing a cross-facet similarity measurement, we project users and items into multiple metric spaces for fine-grained representation learning, and compare them only in the proper spaces. Furthermore, we devise a spherical optimization strategy to enhance the effectiveness and robustness of the multi-facet recommendation framework. Extensive experiments on six real-world benchmark datasets show drastic performance gains brought by MARS, which constantly achieves up to 40\% improvements over the state-of-the-art baselines regarding both HR and nDCG metrics.

Jialiang Han, Yun Ma

Mobile devices enable users to retrieve information at any time and any place. Considering the occasional requirements and fragmentation usage pattern of mobile users, temporal recommendation techniques are proposed to improve the efficiency of information retrieval on mobile devices by means of accurately recommending items via learning temporal interests with short-term user interaction behaviors. However, the enforcement of privacy-preserving laws and regulations, such as GDPR, may overshadow the successful practice of temporal recommendation. The reason is that state-of-the-art recommendation systems require to gather and process the user data in centralized servers but the interaction behaviors data used for temporal recommendation are usually non-transactional data that are not allowed to gather without the explicit permission of users according to GDPR. As a result, if users do not permit services to gather their interaction behaviors data, the temporal recommendation fails to work. To realize the temporal recommendation in the post-GDPR era, this paper proposes $C^3DRec$, a cloud-client cooperative deep learning framework of mining interaction behaviors for recommendation while preserving user privacy. $C^3DRec$ constructs a global recommendation model on centralized servers using data collected before GDPR and fine-tunes the model directly on individual local devices using data collected after GDPR. We design two modes to accomplish the recommendation, i.e. pull mode where candidate items are pulled down onto the devices and fed into the local model to get recommended items, and push mode where the output of the local model is pushed onto the server and combined with candidate items to get recommended ones. Evaluation results show that $C^3DRec$ achieves comparable recommendation accuracy to the centralized approaches, with minimal privacy concern.

Jinfeng Wen, Yi Liu, Zhenpeng Chen et al.

Serverless computing has become a new trending paradigm in cloud computing, allowing developers to focus on the development of core application logic and rapidly construct the prototype via the composition of independent functions. With the development and prosperity of serverless computing, major cloud vendors have successively rolled out their commodity serverless computing platforms. However, the characteristics of these platforms have not been systematically studied. Measuring these characteristics can help developers to select the most adequate serverless computing platform and develop their serverless-based applications in the right way. To fill this knowledge gap, we present a comprehensive study on characterizing mainstream commodity serverless computing platforms, including AWS Lambda, Google Cloud Functions, Azure Functions, and Alibaba Cloud Function Compute. Specifically, we conduct both qualitative analysis and quantitative analysis. In qualitative analysis, we compare these platforms from three aspects (i.e., development, deployment, and runtime) based on their official documentation to construct a taxonomy of characteristics. In quantitative analysis, we analyze the runtime performance of these platforms from multiple dimensions with well-designed benchmarks. First, we analyze three key factors that can influence the startup latency of serverless-based applications. Second, we compare the resource efficiency of different platforms with 16 representative benchmarks. Finally, we measure their performance difference when dealing with different concurrent requests, and explore the potential causes in a black-box fashion. Based on the results of both qualitative and quantitative analysis, we derive a series of findings and provide insightful implications for both developers and cloud vendors.

Yangbin Chen, Yun Ma, Tom Ko et al.

Model-Agnostic Meta-Learning (MAML) and its variants are popular few-shot classification methods. They train an initializer across a variety of sampled learning tasks (also known as episodes) such that the initialized model can adapt quickly to new tasks. However, current MAML-based algorithms have limitations in forming generalizable decision boundaries. In this paper, we propose an approach called MetaMix. It generates virtual feature-target pairs within each episode to regularize the backbone models. MetaMix can be integrated with any of the MAML-based algorithms and learn the decision boundaries generalizing better to new tasks. Experiments on the mini-ImageNet, CUB, and FC100 datasets show that MetaMix improves the performance of MAML-based algorithms and achieves state-of-the-art result when integrated with Meta-Transfer Learning.

Kaidong Wu, Yun Ma, Gang Huang et al.

With the increasing popularity of blockchain technologies in recent years, blockchain-based decentralized applications (DApps for short in this paper) have been rapidly developed and widely adopted in many areas, being a hot topic in both academia and industry. Despite of the importance of DApps, we still have quite little understanding of DApps along with its ecosystem. To bridge the knowledge gap, this paper presents the first comprehensive empirical study of blockchain-based DApps to date, based on an extensive dataset of 995 Ethereum DApps and 29,846,075 transaction logs over them. We make a descriptive analysis of the popularity of DApps, summarize the patterns of how DApps use smart contracts to access the underlying blockchain, and explore the worth-addressing issues of deploying and operating DApps. Based on the findings, we propose some implications for DApp users to select proper DApps, for DApp developers to improve the efficiency of DApps, and for blockchain vendors to enhance the support of DApps.

Yun Ma, Dongwei Xiang, Shuyu Zheng et al.

Recently, several JavaScript-based deep learning frameworks have emerged, making it possible to perform deep learning tasks directly in browsers. However, little is known on what and how well we can do with these frameworks for deep learning in browsers. To bridge the knowledge gap, in this paper, we conduct the first empirical study of deep learning in browsers. We survey 7 most popular JavaScript-based deep learning frameworks, investigating to what extent deep learning tasks have been supported in browsers so far. Then we measure the performance of different frameworks when running different deep learning tasks. Finally, we dig out the performance gap between deep learning in browsers and on native platforms by comparing the performance of TensorFlow.js and TensorFlow in Python. Our findings could help application developers, deep-learning framework vendors and browser vendors to improve the efficiency of deep learning in browsers.

Yun Ma, Ziniu Hu, Dian Yang et al.

Unlike the Web where each web page has a global URL to reach, a specific "content page" inside a mobile app cannot be opened unless the user explores the app with several operations from the landing page. Recently, deep links have been advocated by major companies to enable targeting and opening a specific page of an app externally with an accessible uniform resource identifier (URI). To empirically investigate the state of the practice on adopting deep links, in this article, we present the largest empirical study of deep links over 20,000 Android apps, and find that deep links do not get wide adoption among current Android apps, and non-trivial manual efforts are required for app developers to support deep links. To address such an issue, we propose the Aladdin approach and supporting tool to release deep links to access arbitrary location of existing apps. Aladdin instantiates our novel cooperative framework to synergically combine static analysis and dynamic analysis while minimally engaging developers to provide inputs to the framework for automation, without requiring any coding efforts or additional deployment efforts. We evaluate Aladdin with popular apps and demonstrate its effectiveness and performance.

Yun Ma, Shuailiang Dong

The web browser is one of the major channels to access the Internet on mobile devices. Based on the smartphone usage logs from millions of real-world Android users, it is interesting to find that about 38% users have more than one browser on their devices. However, it is unclear whether the quality of browsing experiences are different when visiting the same webpage on different browsers. In this paper, we collect 3-week consecutive traces of 337 popular webpages on three popular mobile browsers: Chrome, Firefox, and Opera. We first use a list of metrics and conduct an empirical study to measure the differences of these metrics on different browsers. Then, we explore the variety of loading time and cache performance of different browsers when visiting the same webpage, which has a great impact on the browsing experience. Furthermore, we try to find which metrics have significant effect on the differences, investigating the possible causes. Finally, according to our findings, we give some recommendations to web developers, browser vendors, and end users.

Ziniu Hu, Yun Ma, Yangyang Huang

Generating test cases through automatic app exploration is very useful for analyzing and testing Android apps. However, test cases generated by current app-exploration tools are not reproducible, i.e. when the generated test case is re-executed, the app cannot reach the same state as the explored one. As a result, app developers are not able to reproduce the failure or crash reported during the exploration, to conduct regression test after fixing the bug, or to execute the same test in different environments. In this paper, we present DroidWalker, a dynamic-analysis tool to generate reproducible test cases for Android apps. The key design of our tool is a dynamic-adaptive model that can abstract the app state in a proper granularity so every state in the model can be reached afterwards. Given an app under test, DroidWalker first explores the app to build the model. Then developers can select the state in the model to be reproduced. Finally, DroidWalker executes all the generated test cases and the app could reach exactly the same state as the explored one. We apply DroidWalker in three real usage scenarios to demonstrate its practical usage. The video of our tool is at https://youtu.be/ndUD8Gxs800.

Ziniu Hu, Yun Ma, Qiaozhu Mei et al.

Mobile applications (a.k.a., apps), which facilitate a large variety of tasks on mobile devices, have become indispensable in our everyday lives. Accomplishing a task may require the user to navigate among various apps. Unlike Web pages that are inherently interconnected through hyperlinks, mobile apps are usually isolated building blocks, and the lack of direct links between apps has largely compromised the efficiency of task completion. In this paper, we present the first in-depth empirical study of inter-app navigation behaviors of smartphone users based on a comprehensive dataset collected through a sizable user study over three months. We propose a model to distinguish informational pages and transitional pages, based on which a large number of inter-app navigation are identified. We reveal that developing 'tunnels' between of isolated apps has a huge potential to reduce the cost of navigation. Our analysis provides various practical implications on how to improve app-navigation experiences from both the operating system's perspective and the developer's perspective.

Yun Ma, Xuanzhe Liu, Ruogu Du et al.

The mobile application (app) has become the main entrance to access the Internet on handheld devices. Unlike the Web where each webpage has a global URL to reach directly, a specific "content page" of an app can be opened only by exploring the app with several operations from the landing page. The interoperability between apps is quite fixed and thus limits the value-added "linked data" between apps. Recently, deep link has been proposed to enable targeting and opening a specific page of an app externally with an accessible uniform resource identifier (URI). However, implementing deep link for mobile apps requires a lot of manual efforts by app developers, which can be very error-prone and time-consuming. In this paper, we propose DroidLink to automatically generating deep links for existing Android apps. We design a deep link model suitable for automatic generation. Then we explore the transition of pages and build a navigation graph based on static and dynamic analysis of Android apps. Next, we realize an updating mechanism that keeps on revisiting the target app and discover new pages, and thus generates deep links for every single page of the app. Finally, we repackage the app with deep link supports, but requires no additional deployment requirements. We generate deep links for some popular apps and demonstrate the feasibility of DroidLink.

Xuanzhe Liu, Yun Ma, Shuailiang Dong et al.

Redundant transfer of resources is a critical issue for compromising the performance of mobile Web applications (a.k.a., apps) in terms of data traffic, load time, and even energy consumption. Evidence shows that the current cache mechanisms are far from satisfactory. With lessons learned from how native apps manage their resources, in this paper, we propose the ReWAP approach to fundamentally reducing redundant transfers by restructuring the resource loading of mobile Web apps. ReWAP is based on an efficient mechanism of resource packaging where stable resources are encapsulated and maintained into a package, and such a package shall be loaded always from the local storage and updated by explicitly refreshing. By retrieving and analyzing the update of resources, ReWAP maintains resource packages that can accurately identify which resources can be loaded from the local storage for a considerably long period. ReWAP also provides a wrapper for mobile Web apps to enable loading and updating resource packages in the local storage as well as loading resources from resource packages. ReWAP can be easily and seamlessly deployed into existing mobile Web architectures with minimal modifications, and is transparent to end-users. We evaluate ReWAP based on continuous 15-day access traces of 50 mobile Web apps that suffer heavily from the problem of redundant transfers. Compared to the original mobile Web apps with cache enabled, ReWAP can significantly reduce the data traffic, with the median saving up to 51%. In addition, ReWAP can incur only very minor runtime overhead of the client-side browsers.