arXiv subjects

Xiangyi Li

Publications and source records attributed to Xiangyi Li.

16 recordsLinked to original sources

arXiv2026

BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure

LM-agent benchmarks increasingly function as interactive evaluation infrastructure. Agents observe state, call tools, modify workspaces, submit artifacts, and receive rewards from outcome procedures. This interactivity makes evaluations vulnerable to reward hacking: an agent improves its measured score by exploiting the reward-relevant trajectory instead of solving the intended task. Existing defenses rely largely on task-specific patches, prompt instructions, or post-hoc detectors. They do not provide reusable evidence that a concrete run remained within its intended evaluation boundary. This paper presents BenchShield, a model-backed instrumentation layer for reward integrity in LLM-agent evaluation. BenchShield grounds detection in a finite lifecycle model of an evaluation's reward-relevant events. Within the benchmark infrastructure, two complementary analyses operate over this model. A static, phase-aware taint analysis exposes reward-hacking paths before a run. Its runtime counterpart uses infrastructure-side evidence to attribute concrete agent use and emit evidence-backed claims. We construct BenchShield Trajectories, a human-labeled corpus of 456 adjudicated trajectories from more than 31,000 public agent runs across three benchmarks. Compared with an agentic hackability scanner baseline on the same tasks and model, BenchShield improves full-chain recall from 23-94% to 77-100%, same-vector coverage from 16-56% to 43-78%, and reduces per-task cost by up to 65%. Its runtime analysis achieves 96% accuracy in detecting reward hacking from infrastructure-side evidence.

cs.CR↗

arXiv2026

Learnware for CSI Feedback: Scene-specific Small Models Can Do Big

Intelligent channel state information (CSI) feedback is essential for realizing the high capacity and spectral efficiency goals of future 6G systems, yet existing deep learning solutions face a trade-off between model generalization and scenario-specific performance. Large neural networks generalize well but incur high computational and tuning costs, while small models excel in particular environments but require repetitive costly end-to-end training for each base station (BS). To address these challenges, we introduce a model repository-based deployment framework in which a centralized AI data center maintains a catalog of scene-specific CSI models. The repository is enhanced with a Learnware-based framework, where each model is associated with a specification including semantic part (network architecture parameters) and statistical part (codeboo-fingerprint embeddings of training-data distributions). A BS submits only its local statistical specifications to retrieve the most relevant pre-trained model, enhancing data privacy by avoiding raw CSI transmission and drastically reducing retrieval latency and communication overhead. We further develop a data-driven search strategy that matches codebook fingerprints to model performance, achieving over 90% selection accuracy. In simulations, our scheme yields 18.8% and 57.7% performance improvements over the General Model in LOS and NLOS scenarios, respectively while reducing local fine-tuning by up to 1000 samples and 100 epochs. This Learnware-based approach minimizes redundant training, maximizes model reuse, and supports rapid,privacy-enhancing deployment of CSI feedback models.

cs.IT↗

arXiv2026

OpenThoughts-Agent: Data Recipes for Agentic Models

Agentic language models dramatically expand the applications of AI yet little is publicly known about how to curate training data for broadly capable agents. Existing open efforts such as SWE-Smith, SERA, and Nemotron-Terminal typically target a single benchmark, leaving open the question of how to train models that generalize across diverse agentic tasks. The OpenThoughts-Agent (OT-Agent) project addresses this gap with a fully open data curation pipeline for training agentic models. We conduct more than 100 controlled ablation experiments to systematically investigate each stage of the pipeline, yielding insights on the importance of task sources and diversity. We then assemble a training set of 100K examples from our pipeline and fine-tune Qwen3-32B on this dataset, which yields an average accuracy of 44.8% across seven agentic benchmarks and a 3.9 percentage point improvement over the strongest existing open data agentic model (Nemotron-Terminal-32B, 40.9%). Moreover, our training data exhibits strong scaling properties, outperforming alternative open datasets at every training set size in compute-controlled comparisons. We publicly release our training sets, data pipeline, experimental data, and models at openthoughts.ai to support future open research on agentic model training.

cs.AI↗

arXiv2026

SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?

AI agents are increasingly expected to complete long-horizon workflows that require sustained progress over hours, millions of tokens, and complex environments. Yet current agent benchmarks largely evaluate short-form tasks, such as single pull requests, small tickets, or 5-10 minute exercises, limiting our ability to measure agents' capabilities in planning, long-context understanding, and memory use. We introduce SWE-Marathon, a benchmark of 20 long-horizon tasks spanning software engineering and adjacent technical domains. Each task consists of a unique executable environment, a human-written reference solution, and a multi-layer verification suite. Logged agent attempts average 27.2M total tokens, making SWE-Marathon substantially longer-horizon than existing SWE and command-line agent benchmarks. Current frontier coding agents solve fewer than 30% of tasks. Failures often arise from poor self-verification, self-reported infeasibility, and premature termination. We also observe reward-hacking behavior in 13.8% of rollouts, where agents attempt to exploit the environment or verifier to bypass the intended workflow. SWE-Marathon includes adversarial review of test suites and execution environments, as well as multi-layer checks designed to prevent shortcut solutions. We release SWE-Marathon, evaluation code, and agent trajectories at https://swe-marathon.org/.

cs.SE↗

arXiv2026

AI-Empowered Low-Altitude Economy: Cooperative Sensing With Fixed Wireless Access

The rapid growth of the low-altitude economy has intensified safety concerns arising from unauthorized unmanned aerial vehicles (UAVs), positioning UAV supervision as a key use case in 3GPP. To precisely sense such UAVs with wide coverage and low cost, we leverage fixed wireless access (FWA) customer premises equipment (CPEs), static, densely deployed devices that serve as wireless cameras for the radio environment. We develop an artificial intelligence-empowered two-stage cooperative sensing pipeline that exploits uplink channel state information (CSI) from multiple base station-CPE pairs for UAV detection and localization. In cooperative detection, lightweight CSI features are first individually extracted by neural network, and then adaptively integrated through an attention-based scheme to declare UAV presence. The learned attention scores effectively identify the critical pairs during detection, while facilitating UAV-affected pair selection for subsequent localization. For cooperative localization, neural network initially generates individual estimates and extract CSI features from selected pairs. These estimates, together with features and pair indexes, are fused using a Transformer to produce a precise cooperative estimate. Simulations show that cooperative schemes significantly reduce the missed detection probability to 0.63% and realize a 95%-confidence positioning error of 6.50 m, satisfying 3GPP requirements and showing the potential of FWA-assisted cooperative sensing. Dataset and codes are available on GitHub.

eess.SP↗

arXiv2026

ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces

Large language model (LLM) agents are increasingly deployed to automate productivity tasks (e.g., email, scheduling, document management), but evaluating them on live services is risky due to potentially irreversible changes. Existing benchmarks rely on simplified environments and fail to capture realistic, stateful, multi-service workflows. We introduce ClawsBench, a benchmark for evaluating and improving LLM agents in realistic productivity settings. It includes five high-fidelity mock services (Gmail, Slack, Google Calendar, Google Docs, Google Drive) with full state management and deterministic snapshot/restore, along with 44 structured tasks covering single-service, cross-service, and safety-critical scenarios. We decompose agent scaffolding into two independent levers (domain skills that inject API knowledge via progressive disclosure, and a meta prompt that coordinates behavior across services) and vary both to measure their separate and combined effects. Experiments across 6 models, 4 agent harnesses, and 33 conditions show that with full scaffolding, agents achieve task success rates of 39-64% but exhibit unsafe action rates of 7-33%. On OpenClaw, the top five models fall within a 10 percentage-point band on task success (53-63%), with unsafe action rates from 7% to 23% and no consistent ordering between the two metrics. We identify eight recurring patterns of unsafe behavior, including multi-step sandbox escalation and silent contract modification. We release the trajectories and future dataset at https://clawsbench.com.

cs.AI↗

arXiv2026

SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

Agent Skills are structured packages of procedural knowledge that augment large language model (LLM) agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark whose current inventory contains 87 tasks across 8 domains paired with curated Skills and deterministic verifiers. Our latest aggregate evaluation runs the 87-task benchmark under matched no-Skills and curated-Skills conditions for 18 model-harness configurations. Curated Skills raise the average pass rate from 33.9% to 50.5% (+16.6 percentage points; 25.5% normalized gain), with configuration-level gains ranging from +4.1 to +25.7 pp. Focused Skills with at most three modules outperform larger or exhaustive bundles, and smaller models with Skills can match larger models without them. SkillsBench establishes paired evaluation as the foundation for rigorous measurement of Skill efficacy on agentic, expertise-heavy work.

cs.AI↗

arXiv2026

FPNet: Joint Wi-Fi Beamforming Matrix Feedback and Anomaly-Aware Indoor Positioning

Channel State Information (CSI) provides a detailed description of the wireless channel and has been widely adopted for Wi-Fi sensing, particularly for high-precision indoor positioning. However, complete CSI is rarely available in real-world deployments due to hardware constraints and the high communication overhead required for feedback. Moreover, existing positioning models lack mechanisms to detect when users move outside their trained regions, leading to unreliable estimates in dynamic environments. In this paper, we present FPNet, a unified deep learning framework that jointly addresses channel feedback compression, accurate indoor positioning, and robust anomaly detection (AD). FPNet leverages the beamforming feedback matrix (BFM), a compressed CSI representation natively supported by IEEE 802.11ac/ax/be protocols, to minimize feedback overhead while preserving critical positioning features. To enhance reliability, we integrate ADBlock, a lightweight AD module trained on normal BFM samples, which identifies out-of-distribution scenarios when users exit predefined spatial regions. Experimental results using standard 2.4 GHz Wi-Fi hardware show that FPNet achieves positioning accuracy above 97% with only 100 feedback bits, boosts net throughput by up to 22.92%, and attains AD accuracy over 99% with a false alarm rate below 1.5%. These results demonstrate FPNet's ability to deliver efficient, accurate, and reliable indoor positioning on commodity Wi-Fi devices.

cs.IT↗

arXiv2026

Widefield two-photon random illumination microscopy (2P-RIM)

Biological and biomedical samples are routinely examined using focused two-photon (2P) fluorescence microscopy due to its intrinsic axial sectioning and reduced out-of-focus bleaching. However, 2P imaging often requires excitation intensities that can damage samples through ionization and radical formation. Additionally, the lateral resolution of 2P microscopy is lower compared to linear one-photon (1P) fluorescence microscopy. Widefield 2P microscopy, using cameras, holds promise for reducing photo-toxicity while maintaining high image acquisition rates. Widefield imaging trades the high power and short integration times of sequential single point scanning for the low power and extended integration times of parallel detection across millions of pixels. However, generating effective axial sectioning over arbitrarily large fields of view (FOVs) has remained a challenge. In this work, we introduce 2P Random Illumination Microscopy (2P-RIM), an easy-to-implement 2P widefield technique, that achieves low photo-damage, fast imaging, micrometric axial sectioning, and enhanced lateral resolution for arbitrarily large FOVs. By using widefield speckled illuminations in conjunction with an image standard deviation matching algorithm, 2P-RIM demonstrated multicolor imaging over FOVs greater than 200 um, lateral resolution 220 nm, axial sectioning 2 um, and peak excitation powers about 10 times lower than those used in focused laser scanning microscopy.

physics.optics↗

arXiv2025

In2x at WMT25 Translation Task

This paper presents the open-system submission by the In2x research team for the WMT25 General Machine Translation Shared Task. Our submission focuses on Japanese-related translation tasks, aiming to explore a generalizable paradigm for extending large language models (LLMs) to other languages. This paradigm encompasses aspects such as data construction methods and reward model design. The ultimate goal is to enable large language model systems to achieve exceptional performance in low-resource or less commonly spoken languages.

cs.CL↗

arXiv2025

AI for CSI Prediction in 5G-Advanced and Beyond

Artificial intelligence (AI) is pivotal in advancing fifth-generation (5G)-Advanced and sixth-generation systems, capturing substantial research interest. Both the 3rd Generation Partnership Project (3GPP) and leading corporations champion AI's standardization in wireless communication. This piece delves into AI's role in channel state information (CSI) prediction, a sub-use case acknowledged in 5G-Advanced by the 3GPP. We offer an exhaustive survey of AI-driven CSI prediction, highlighting crucial elements like accuracy, generalization, and complexity. Further, we touch on the practical side of model management, encompassing training, monitoring, and data gathering. Moreover, we explore prospects for CSI prediction in future wireless communication systems, entailing integrated design with feedback, multitasking synergy, and predictions in rapid scenarios. This article seeks to be a touchstone for subsequent research in this burgeoning domain.

eess.SP↗

arXiv2025

Efficient Deployment of Deep MIMO Detection Using Learngene

Deep learning (DL) has introduced a new paradigm in multiple-input multiple-output (MIMO) detection, balancing performance and complexity. However, the practical deployment of DL-based detectors is hindered by poor generalization, necessitating costly retraining for different devices and scenarios. To address this challenge, this paper presents a novel knowledge transfer technique, termed learngene, for the design of a DL-based MIMO detector and proposes an efficient deployment framework. The proposed detector, SDNet, leverages zero-forcing detection outputs and least squares-estimated channel state information (CSI) as inputs. It is further optimized through a collective-individual paradigm to enhance knowledge transfer. In this paradigm, learngene, a reusable neural network (NN) segment, encapsulates detection meta-knowledge acquired from large-scale collective models trained by manufacturers. This segment can then be distributed to device-specific teams. By integrating learngene into different lightweight individual models, detection meta-knowledge is efficiently transferred across heterogeneous NNs, enabling adaptation to diverse devices and scenarios. Simulation results demonstrate that the proposed scheme enhances performance, enables rapid adaptation, and ensures high scalability, with transferred parameters comprising only 10.8% of the total model size.

eess.SP↗

arXiv2024

Deep Learning-based CSI Feedback in Wi-Fi Systems

In Wi-Fi systems, channel state information (CSI) plays a crucial role in enabling access points to execute beamforming operations. However, the feedback overhead associated with CSI significantly hampers the throughput improvements. Recent advancements in deep learning (DL) have transformed the approach to CSI feedback in cellular systems. Drawing inspiration from the successes witnessed in the realm of mobile communications, this paper introduces a DL-based CSI feedback framework, named EFNet, tailored for Wi-Fi systems. The proposed framework leverages an autoencoder to achieve precise feedback with minimal overhead. The process involves the station utilizing the encoder to compress and quantize a series of matrices into codeword bit streams, which are then fed back to the access point. Subsequently, the decoder installed at the AP reconstructs beamforming matrices from these bit streams. We implement the EFNet system using standard Wi-Fi equipment operating in the 2.4 GHz band. Experimental findings in an office environment reveal a remarkable 80.77% reduction in feedback overhead compared to the 802.11ac standard, alongside a significant boost in net throughput of up to 30.72%.

eess.SP↗

arXiv2023

Auto-CsiNet: Scenario-customized Automatic Neural Network Architecture Generation for Massive MIMO CSI Feedback

Deep learning has revolutionized the design of the channel state information (CSI) feedback module in wireless communications. However, designing the optimal neural network (NN) architecture for CSI feedback can be a laborious and time-consuming process. Manual design can be prohibitively expensive for customizing NNs to different scenarios. This paper proposes using neural architecture search (NAS) to automate the generation of scenario-customized CSI feedback NN architectures, thereby maximizing the potential of deep learning in exclusive environments. By employing automated machine learning and gradient-descent-based NAS, an efficient and cost-effective architecture design process is achieved. The proposed approach leverages implicit scene knowledge, integrating it into the scenario customization process in a data-driven manner, and fully exploits the potential of deep learning for each specific scenario. To address the issue of excessive search, early stopping and elastic selection mechanisms are employed, enhancing the efficiency of the proposed scheme. The experimental results demonstrate that the automatically generated architecture, known as Auto-CsiNet, outperforms manually-designed models in both reconstruction performance (achieving approximately a 14% improvement) and complexity (reducing it by approximately 50%). Furthermore, the paper analyzes the impact of the scenario on the NN architecture and its capacity.

cs.IT↗

arXiv2023

Multi-task Learning-based CSI Feedback Design in Multiple Scenarios

For frequency division duplex systems, the essential downlink channel state information (CSI) feedback includes the links of compression, feedback, decompression and reconstruction to reduce the feedback overhead. One efficient CSI feedback method is the Auto-Encoder (AE) structure based on deep learning, yet facing problems in actual deployments, such as selecting the deployment mode when deploying in a cell with multiple complex scenarios. Rather than designing an AE network with huge complexity to deal with CSI of all scenarios, a more realistic mode is to divide the CSI dataset by region/scenario and use multiple relatively simple AE networks to handle subregions' CSI. However, both require high memory capacity for user equipment (UE) and are not suitable for low-level devices. In this paper, we propose a new user-friendly-designed framework based on the latter multi-tasking mode. Via Multi-Task Learning, our framework, Single-encoder-to-Multiple-decoders (S-to-M), designs the multiple independent AEs into a joint architecture: a shared encoder corresponds to multiple task-specific decoders. We also complete our framework with GateNet as a classifier to enable the base station autonomously select the right task-specific decoder corresponding to the subregion. Experiments on the simulating multi-scenario CSI dataset demonstrate our proposed S-to-M's advantages over the other benchmark modes, i.e., significantly reducing the model complexity and the UE's memory consumption

eess.SP↗

arXiv2019

Spatio-Temporal Representation with Deep Neural Recurrent Network in MIMO CSI Feedback

In multiple-input multiple-output (MIMO) systems, it is crucial of utilizing the available channel state information (CSI) at the transmitter for precoding to improve the performance of frequency division duplex (FDD) networks. One of the mainchallenges is to compress a large amount of CSI in CSI feedback transmission in massive MIMO systems. In this paper, we propose a deep learning (DL)-based approach that uses a deep recurrent neural network (RNN) to learn temporal correlation and adopts depthwise separable convolution to shrink the model. The feature extraction module is also elaborately devised by studyingdecoupled spatio-temporal feature representations in different structures. Experimental results demonstrate that the proposed approach outperforms existing DL-based methods in terms of recovery quality and accuracy, which can also achieve remarkable robustness at low compression ratio (CR).

eess.SP↗