arXiv subjects

Longxiang Wang

Publications and source records attributed to Longxiang Wang.

10 recordsLinked to original sources

arXiv2026

Agentic Environment Engineering for Large Language Models: A Survey of Environment Modeling, Synthesis, Evaluation, and Application

Environments serve as interactive systems for large language model (LLM) based agents across diverse scenarios and play a crucial role in driving the continual evolution of model capabilities. Despite this importance, existing work lacks a systematic categorization and deep analysis. This paper systematically studies current researches on agentic environments from the perspective of the environment engineering lifecycle, covering their modeling, synthesis, evaluation and application. Specifically, the paper first introduces representative environments from the perspectives of eight attributes and eight domains, providing detailed analyses of their development paths and highlighting their core capabilities. Second, for automated environment synthesis, two paradigms are introduced, such as symbolic synthesis and neural synthesis. This paper also shows different environment evaluation methods in each paradigm. Thirdly, the corresponding environment applications from the perspective of agent-environment co-evolution are discussed. In specific, the paper characterizes the primary pathways for agent evolution in dynamic environments from four complementary perspectives: memory-centric experience evolution, orchestration-centric workflow evolution, trajectory-centric offline evolution, and exploration-centric online evolution. And three paradigms of environment evolution are identified, namely neural-driven, difficulty-driven, and scaling-driven approaches. At last, several promising future directions are discussed, including Environment-as-a-Service, Multi-agent Environments, and Neural-Symbolic Environments.

cs.CL↗

arXiv2026

HELLoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Models

Low-Rank Adaptation (LoRA) dominates parameter-efficient fine-tuning of large language models, yet most variants target dense architectures. Mixture-of-Experts (MoE) models scale parameters at near-constant per-token compute, and their sparse activation patterns create untapped opportunities for more efficient adaptation. We propose Hot-Experts Layer-level Low-Rank Adaptation (HELLoRA), which attaches LoRA modules only to the most frequently activated experts at each layer. This simple mechanism reduces trainable parameters and adapter-induced FLOPs while improving downstream performance, an effect we attribute to a form of structured regularization that preserves pretrained expert specialization. To stress-test HELLoRA under extreme parameter budgets, we further compose it with LoRI to form HELLoRI, which freezes the up-projection and sparsifies the down-projection. Across three MoE backbones, namely OlMoE-1B-7B, Mixtral-8x7B, and DeepSeekMoE, and three task families covering mathematical reasoning, code generation, and safety alignment, HELLoRA consistently outperforms strong PEFT baselines. Relative to vanilla LoRA on OlMoE, HELLoRA uses 15.7% of the trainable parameters, reduces adapter FLOPs by 38.7%, achieves 1.9x the training throughput, and improves accuracy by 9.2%. On DeepSeekMoE, HELLoRA outperforms LoRA while using only 23.2% of its trainable parameters. These results demonstrate that activation-aware adapter placement is an effective and practical route to scaling PEFT for MoE language models.

cs.LG↗

arXiv2026

STAR: Mitigating Cascading Errors in Spatial Reasoning via Turn-point Alignment and Segment-level DPO

Structured spatial navigation is a core benchmark for Large Language Models (LLMs) spatial reasoning. Existing paradigms like Visualization-of-Thought (VoT) are prone to cascading errors in complex topologies. To solve this, we propose STAR, a two-stage framework grounded on topological anchors, and introduce the RedMaze-23K dataset with human-inspired turnpoint annotations. The first stage uses supervised fine-tuning to help models internalize spatial semantics and prune redundant paths. The second adopts Spatial-aware Segment-level Direct Preference Optimization (SDPO) to refine self-correction in long-horizon navigation. Experiments show STAR achieves state-of-the-art performance among open-source models: its 32B variant outperforms DeepSeek-V3 (29.27% vs. 25.00%) and reaches 82.4% of GPT-4's performance.

cs.CV↗

arXiv2026

OptiLeak: Efficient Prompt Reconstruction via Reinforcement Learning in Multi-tenant LLM Services

Multi-tenant LLM serving frameworks widely adopt shared Key-Value caches to enhance efficiency. However, this creates side-channel vulnerabilities enabling prompt leakage attacks. Prior studies identified these attack surfaces yet focused on expanding attack vectors rather than optimizing attack performance, reporting impractically high attack costs that underestimate the true privacy risk. We propose OptiLeak, a reinforcement learning-enhanced framework that maximizes prompt reconstruction efficiency through two-stage fine-tuning. Our key insight is that domain-specific ``hard tokens'' -- terms difficult to predict yet carrying sensitive information -- can be automatically identified via likelihood ranking and used to construct preference pairs for Direct Preference Optimization, eliminating manual annotation. This enables effective preference alignment while avoiding the overfitting issues of extended supervised fine-tuning. Evaluated on three benchmarks spanning medical and financial domains, OptiLeak achieves up to $12.48\times$ reduction in average requests per token compared to baseline approaches, with consistent improvements across model scales from 3B to 14B parameters. Our findings demonstrate that cache-based prompt leakage poses a more severe threat than previously reported, underscoring the need for robust cache isolation in production deployments.

cs.CR↗

arXiv2025

Taiji: A DPU Memory Elasticity Solution for In-production Cloud Environments

The growth of cloud computing drives data centers toward higher density and efficiency. Data processing units (DPUs) enhance server network and storage performance but face challenges such as long hardware upgrade cycles and limited resources. To address these, we propose Taiji, a resource-elasticity architecture for DPUs. Combining hybrid virtualization with parallel memory swapping, Taiji switches the DPU's operating system (OS) into a guest OS and inserts a lightweight virtualization layer, making nearly all DPU memory swappable. It achieves memory overcommitment for the switched guest OS via high-performance memory elasticity, fully transparent to upper-layer applications, and supports hot-switch and hot-upgrade to meet in-production cloud requirements. Experiments show that Taiji expands DPU memory resources by over 50%, maintains virtualization overhead around 5%, and ensures 90% of swap-ins complete within 10 microseconds. Taiji delivers an efficient, reliable, low-overhead elasticity solution for DPUs and is deployed in large-scale production systems across more than 30,000 servers.

cs.OS↗

arXiv2025

Vmem: A Lightweight Hot-Upgradable Memory Management for In-production Cloud Environment

Traditional memory management suffers from metadata overhead, architectural complexity, and stability degradation, problems intensified in cloud environments. Existing software/hardware optimizations are insufficient for cloud computing's dual demands of flexibility and low overhead. This paper presents Vmem, a memory management architecture for in-production cloud environments that enables flexible, efficient cloud server memory utilization through lightweight reserved memory management. Vmem is the first such architecture to support online upgrades, meeting cloud requirements for high stability and rapid iterative evolution. Experiments show Vmem increases sellable memory rate by about 2%, delivers extreme elasticity and performance, achieves over 3x faster boot time for VFIO-based virtual machines (VMs), and improves network performance by about 10% for DPU-accelerated VMs. Vmem has been deployed at large scale for seven years, demonstrating efficiency and stability on over 300,000 cloud servers supporting hundreds of millions of VMs.

cs.OS↗

arXiv2025

EvoEmpirBench: Dynamic Spatial Reasoning with Agent-ExpVer

Most existing spatial reasoning benchmarks focus on static or globally observable environments, failing to capture the challenges of long-horizon reasoning and memory utilization under partial observability and dynamic changes. We introduce two dynamic spatial benchmarks, locally observable maze navigation and match-2 elimination that systematically evaluate models' abilities in spatial understanding and adaptive planning when local perception, environment feedback, and global objectives are tightly coupled. Each action triggers structural changes in the environment, requiring continuous update of cognition and strategy. We further propose a subjective experience-based memory mechanism for cross-task experience transfer and validation. Experiments show that our benchmarks reveal key limitations of mainstream models in dynamic spatial reasoning and long-term memory, providing a comprehensive platform for future methodological advances. Our code and data are available at https://anonymous.4open.science/r/EvoEmpirBench-143C/.

cs.CV↗

arXiv2025

Memory-Anchored Multimodal Reasoning for Explainable Video Forensics

We address multimodal deepfake detection requiring both robustness and interpretability by proposing FakeHunter, a unified framework that combines memory guided retrieval, a structured Observation-Thought-Action reasoning loop, and adaptive forensic tool invocation. Visual representations from a Contrastive Language-Image Pretraining (CLIP) model and audio representations from a Contrastive Language-Audio Pretraining (CLAP) model retrieve semantically aligned authentic exemplars from a large scale memory, providing contextual anchors that guide iterative localization and explanation of suspected manipulations. Under low internal confidence the framework selectively triggers fine grained analyses such as spatial region zoom and mel spectrogram inspection to gather discriminative evidence instead of relying on opaque marginal scores. We also release X-AVFake, a comprehensive audio visual forgery benchmark with fine grained annotations of manipulation type, affected region or entity, reasoning category, and explanatory justification, designed to stress contextual grounding and explanation fidelity. Extensive experiments show that FakeHunter surpasses strong multimodal baselines, and ablation studies confirm that both contextual retrieval and selective tool activation are indispensable for improved robustness and explanatory precision.

cs.MM↗

arXiv2025

CALM: Curiosity-Driven Auditing for Large Language Models

Auditing Large Language Models (LLMs) is a crucial and challenging task. In this study, we focus on auditing black-box LLMs without access to their parameters, only to the provided service. We treat this type of auditing as a black-box optimization problem where the goal is to automatically uncover input-output pairs of the target LLMs that exhibit illegal, immoral, or unsafe behaviors. For instance, we may seek a non-toxic input that the target LLM responds to with a toxic output or an input that induces the hallucinative response from the target LLM containing politically sensitive individuals. This black-box optimization is challenging due to the scarcity of feasible points, the discrete nature of the prompt space, and the large search space. To address these challenges, we propose Curiosity-Driven Auditing for Large Language Models (CALM), which uses intrinsically motivated reinforcement learning to finetune an LLM as the auditor agent to uncover potential harmful and biased input-output pairs of the target LLM. CALM successfully identifies derogatory completions involving celebrities and uncovers inputs that elicit specific names under the black-box setting. This work offers a promising direction for auditing black-box LLMs. Our code is available at https://github.com/x-zheng16/CALM.git.

cs.AI↗

arXiv2024

Dual-pronged deep learning preprocessing on heterogeneous platforms with CPU, Accelerator and CSD

For image-related deep learning tasks, the first step often involves reading data from external storage and performing preprocessing on the CPU. As accelerator speed increases and the number of single compute node accelerators increases, the computing and data transfer capabilities gap between accelerators and CPUs gradually increases. Data reading and preprocessing become progressively the bottleneck of these tasks. Our work, DDLP, addresses the data computing and transfer bottleneck of deep learning preprocessing using Computable Storage Devices (CSDs). DDLP allows the CPU and CSD to efficiently parallelize preprocessing from both ends of the datasets, respectively. To this end, we propose two adaptive dynamic selection strategies to make DDLP control the accelerator to automatically read data from different sources. The two strategies trade-off between consistency and efficiency. DDLP achieves sufficient computational overlap between CSD data preprocessing and CPU preprocessing, accelerator computation, and accelerator data reading. In addition, DDLP leverages direct storage technology to enable efficient SSD-to-accelerator data transfer. In addition, DDLP reduces the use of expensive CPU and DRAM resources with more energy-efficient CSDs, alleviating preprocessing bottlenecks while significantly reducing power consumption. Extensive experimental results show that DDLP can improve learning speed by up to 23.5% on ImageNet Dataset while reducing energy consumption by 19.7% and CPU and DRAM usage by 37.6%. DDLP also improves the learning speed by up to 27.6% on the Cifar-10 dataset.

cs.DC↗