Wenhan Yu

Yilun Yao, Xinyu Tan, Chao-Hsuan Liu et al.

LLM agents are increasingly deployed as executable systems that use tools, modify workspaces, and produce concrete artifacts. In such workflows, performance depends not only on the base model, but also on the harness: the system layer that manages context, tools, state, constraints, permissions, tracing, and recovery. However, existing benchmarks typically abstract away execution, compare complete agent systems, or hold the harness fixed, making execution-layer variation difficult to study. We introduce Harness-Bench, a diagnostic benchmark for evaluating configuration-level harness effects in realistic agent workflows. Harness-Bench evaluates representative harness configurations across multiple model backends under shared task environments, budgets, and evaluation protocols, while preserving each harness's native execution behavior. The benchmark contains 106 sandboxed offline tasks constructed from practical agent-use patterns and manually reviewed for realism, solvability, oracle-checkability, and integrity. Each run records final artifacts, execution traces, usage statistics, and validator outputs, enabling analysis beyond final completion. Across 5,194 execution trajectories, we observe substantial variation in completion, process quality, efficiency, and failure behavior across model-harness pairings. These results suggest that agent capability should be reported at the model-harness configuration level rather than attributed to the base model alone. Our analysis further identifies recurring execution-alignment failures, where plausible reasoning becomes decoupled from tool feedback, workspace state, evidence, or verifiable output contracts. Harness-Bench provides a reproducible foundation for diagnosing and improving reliable, efficient, and auditable agent execution stacks.

Wenhan Yu, Terence Jie Chua, Jun Zhao

Technology advancements in wireless communications and high-performance Extended Reality (XR) have empowered the developments of the Metaverse. The demand for the Metaverse applications and hence, real-time digital twinning of real-world scenes is increasing. Nevertheless, the replication of 2D physical world images into 3D virtual objects is computationally intensive and requires computation offloading. The disparity in transmitted object dimension (2D as opposed to 3D) leads to asymmetric data sizes in uplink (UL) and downlink (DL). To ensure the reliability and low latency of the system, we consider an asynchronous joint UL-DL scenario where in the UL stage, the smaller data size of the physical world images captured by multiple extended reality users (XUs) will be uploaded to the Metaverse Console (MC) to be construed and rendered. In the DL stage, the larger-size 3D virtual objects need to be transmitted back to the XUs. We design a novel multi-agent reinforcement learning algorithm structure, namely Asynchronous Actors Hybrid Critic (AAHC), to optimize the decisions pertaining to computation offloading and channel assignment in the UL stage and optimize the DL transmission power in the DL stage. Extensive experiments demonstrate that compared to proposed baselines, AAHC obtains better solutions with satisfactory training time.

Terence Jie Chua, Wenhan Yu, Jun Zhao

Improving the interactivity and interconnectivity between people is one of the highlights of the Metaverse. The Metaverse relies on a core approach, digital twinning, which is a means to replicate physical world objects, people, actions and scenes onto the virtual world. Being able to access scenes and information associated with the physical world, in the Metaverse in real-time and under mobility, is essential in developing a highly accessible, interactive and interconnective experience for all users. This development allows users from other locations to access high-quality real-world and up-to-date information about events happening in another location, and socialize with others hyper-interactively. Nevertheless, receiving continual, smooth updates generated by others from the Metaverse is a challenging task due to the large data size of the virtual world graphics and the need for low latency transmission. With the development of Mobile Augmented Reality (MAR), users can interact via the Metaverse in a highly interactive manner, even under mobility. Hence in our work, we considered an environment with users in moving Internet of Vehicles (IoV), downloading real-time virtual world updates from Metaverse Service Provider Cell Stations (MSPCSs) via wireless communications. We design an environment with multiple cell stations, where there will be a handover of users' virtual world graphic download tasks between cell stations. As transmission latency is the primary concern in receiving virtual world updates under mobility, our work aims to allocate system resources to minimize the total time taken for users in vehicles to download their virtual world scenes from the cell stations. We utilize deep reinforcement learning and evaluate the performance of the algorithms under different environmental configurations. Our work provides a use case of the Metaverse over AI-enabled 6G communications.

Terence Jie Chua, Wenhan Yu, Jun Zhao

Real-time Digital Twinning of physical world scenes onto the Metaverse is necessary for a myriad of applications such as augmented-reality (AR) assisted driving. In AR assisted driving, physical environment scenes are first captured by Internet of Vehicles (IoVs) and are uploaded to the Metaverse. A central Metaverse Map Service Provider (MMSP) will aggregate information from all IoVs to develop a central Metaverse Map. Information from the Metaverse Map can then be downloaded into individual IoVs on demand and be delivered as AR scenes to the driver. However, the growing interest in developing AR assisted driving applications which relies on digital twinning invites adversaries. These adversaries may place physical adversarial patches on physical world objects such as cars, signboards, or on roads, seeking to contort the virtual world digital twin. Hence, there is a need to detect these physical world adversarial patches. Nevertheless, as real-time, accurate detection of adversarial patches is compute-intensive, these physical world scenes have to be offloaded to the Metaverse Map Base Stations (MMBS) for computation. Hence in our work, we considered an environment with moving Internet of Vehicles (IoV), uploading real-time physical world scenes to the MMBSs. We formulated a realistic joint variable optimization problem where the MMSPs' objective is to maximize adversarial patch detection mean average precision (mAP), while minimizing the computed AR scene up-link transmission latency and IoVs' up-link transmission idle count, through optimizing the IoV-MMBS allocation and IoV up-link scene resolution selection. We proposed a Heterogeneous Action Proximal Policy Optimization (HAPPO) (discrete-continuous) algorithm to tackle the proposed problem. Extensive experiments shows HAPPO outperforms baseline models when compared against key metrics.

Terence Jie Chua, Wenhan Yu, Jun Zhao

The Metaverse can be considered the extension of the present-day web, which integrates the physical and virtual worlds, delivering hyper-realistic user experiences. The inception of the Metaverse brings forth many ecosystem services such as content creation, social entertainment, in-world value transfer, intelligent traffic, healthcare. These services are compute-intensive and require computation offloading onto a Metaverse edge computing server (MECS). Existing Metaverse edge computing approaches do not efficiently and effectively handle resource allocation to ensure a fluid, seamless and hyper-realistic Metaverse experience required for Metaverse ecosystem services. Therefore, we introduce a new Metaverse-compatible, Unified, User and Task (UUT) centered artificial intelligence (AI)- based mobile edge computing (MEC) paradigm, which serves as a concept upon which future AI control algorithms could be built to develop a more user and task-focused MEC.

Peiyuan Si, Wenhan Yu, Jun Zhao et al.

Metaverse is expected to create a virtual world closely connected with reality to provide users with immersive experience with the support of 5G high data rate communication technique. A huge amount of data in physical world needs to be synchronized to the virtual world to provide immersive experience for users, and there will be higher requirements on coverage to include more users into Metaverse. However, 5G signal suffers severe attenuation, which makes it more expensive to maintain the same coverage. Unmanned aerial vehicle (UAV) is a promising candidate technique for future implementation of Metaverse as a low-cost and high-mobility platform for communication devices. In this paper, we propose a proximal policy optimization (PPO) based double-agent cooperative reinforcement learning method for channel allocation and trajectory control of UAV to collect and synchronize data from the physical world to the virtual world, and expand the coverage of Metaverse services economically. Simulation results show that our proposed method is able to achieve better performance compared to the benchmark approaches.

Zihang Li, Rui Zhou, Yingcheng Shi et al.

When a large language model under reinforcement learning commits a wrong reasoning step early in a trajectory, standard algorithms force it to keep generating until the maximum horizon, spending compute on tokens that never receive positive reward and polluting advantage estimates with post-failure noise. We propose ESPO (Early-Stopping Proximal Policy Optimization), which detects trajectory failure on-the-fly and terminates rollouts early. At each generation step, ESPO computes a surrogate regret using only the logits already computed during sampling, and terminates when the smoothed cumulative regret significantly exceeds its estimated values. Truncated trajectories are treated as absorbing failure states with a terminal reward, concentrating negative temporal-difference (TD) errors near the detected failure step without any additional reward model or human annotation. On DeepSeek-R1-Distill-Qwen-7B trained for mathematical reasoning, ESPO surpasses PPO on AIME~2024 (46.28% vs. 45.25%), AMC~2023 (85.83% vs. 82.94%), and MATH-500 (87.42% vs. 85.43%), while saving more than 20% rollout tokens cumulatively.

Terence Jie Chua, Wenhan Yu, Jun Zhao

Development of defenses against physical world attacks such as adversarial patches is gaining traction within the research community. We contribute to the field of adversarial patch detection by introducing an uncertainty-based adversarial patch localizer which localizes adversarial patch on an image, permitting post-processing patch-avoidance or patch-reconstruction. We quantify our prediction uncertainties with the development of \textit{\textbf{D}etection of \textbf{U}ncertainties in the \textbf{E}xceedance of \textbf{T}hreshold} (DUET) algorithm. This algorithm provides a framework to ascertain confidence in the adversarial patch localization, which is essential for safety-sensitive applications such as self-driving cars and medical imaging. We conducted experiments on localizing adversarial patches and found our proposed DUET model outperforms baseline models. We then conduct further analyses on our choice of model priors and the adoption of Bayesian Neural Networks in different layers within our model architecture. We found that isometric gaussian priors in Bayesian Neural Networks are suitable for patch localization tasks and the presence of Bayesian layers in the earlier neural network blocks facilitates top-end localization performance, while Bayesian layers added in the later neural network blocks contribute to better model generalization. We then propose two different well-performing models to tackle different use cases.

Wenrui Liu, Zixiang Liu, Elsie Dai et al.

Large Language Models (LLMs) are increasingly serving as autonomous agents, and their utilization of external tools via the Model Context Protocol (MCP) is considered a future trend. Current MCP evaluation sets suffer from issues such as reliance on external MCP services and a lack of difficulty awareness. To address these limitations, we propose MCPAgentBench, a benchmark based on real-world MCP definitions designed to evaluate the tool-use capabilities of agents. We construct a dataset containing authentic tasks and simulated MCP tools. The evaluation employs a dynamic sandbox environment that presents agents with candidate tool lists containing distractors, thereby testing their tool selection and discrimination abilities. Furthermore, we introduce comprehensive metrics to measure both task completion rates and execution efficiency. Experiments conducted on various latest mainstream Large Language Models reveal significant performance differences in handling complex, multi-step tool invocations. All code is open-source at Github.

Zhewen Tan, Yilun Yao, Huiyan Jin et al.

Large language model agents increasingly rely on persistent memory to store past interactions, retrieve relevant demonstrations, and improve long-horizon task execution. However, this memory mechanism also creates a practical security vulnerability: an adversarial user may inject malicious records into the agent's memory through ordinary interaction, and these records can later be retrieved to steer the agent's reasoning and actions. Existing defenses primarily focus on online intervention, such as prompt filtering or output blocking, but they do not address the post-hoc question of which stored memories are responsible after harmful behavior has already been observed. We propose \textbf{MemAudit}, a post-hoc causal memory auditing framework for memory-augmented LLM agents. The framework combines two complementary signals: (1) a counterfactual memory influence score that measures each memory's causal contribution to harmful outputs, and (2) a memory consistency graph that identifies structurally anomalous memories within the broader memory store. We evaluate MemAudit against MINJA, a query-only memory injection attack in which malicious records are generated and stored through normal agent interactions rather than direct memory-bank modification. Across both QA and reasoning-agent settings, MemAudit substantially reduces attack success rates under realistic post-hoc auditing scenarios. The results show that QA attack success is reduced from $70\%$ to $0\%$, while RAP attack success drops from $83.3\%$ to $0\%$.

Wenhan Yu, Terence Jie Chua, Jun Zhao

The Metaverse is emerging as maturing technologies are empowering the different facets. Virtual Reality (VR) technologies serve as the backbone of the virtual universe within the Metaverse to offer a highly immersive user experience. As mobility is emphasized in the Metaverse context, VR devices reduce their weights at the sacrifice of local computation abilities. In this paper, for a system consisting of a Metaverse server and multiple VR users, we consider two cases of (i) the server generating frames and transmitting them to users, and (ii) users generating frames locally and thus consuming device energy. Moreover, in our multi-user VR scenario for the Metaverse, users have different characteristics and demands for Frames Per Second (FPS). Then the channel access arrangement (including the decisions on frame generation location), and transmission powers for the downlink communications from the server to the users are jointly optimized to improve the utilities of users. This joint optimization is addressed by deep reinforcement learning (DRL) with heterogeneous actions. Our proposed user-centric DRL algorithm is called User-centric Critic with Heterogenous Actors (UCHA). Extensive experiments demonstrate that our UCHA algorithm leads to remarkable results under various requirements and constraints.

Wenhan Yu, Terence Jie Chua, Jun Zhao

The Metaverse and its promises are fast becoming reality as maturing technologies are empowering the different facets. One of the highlights of the Metaverse is that it offers the possibility for highly immersive and interactive socialization. Virtual reality (VR) technologies are the backbone for the virtual universe within the Metaverse as they enable a hyper-realistic and immersive experience, and especially so in the context of socialization. As the virtual world 3D scenes to be rendered are of high resolution and frame rate, these scenes will be offloaded to an edge server for computation. Besides, the metaverse is user-center by design, and human users are always the core. In this work, we introduce a multi-user VR computation offloading over wireless communication scenario. In addition, we devised a novel user-centered deep reinforcement learning approach to find a near-optimal solution. Extensive experiments demonstrate that our approach can lead to remarkable results under various requirements and constraints.

Wenhan Yu, Xinbo Lin, Lanxin Ni et al.

Large language models (LLMs) have demonstrated strong reasoning abilities across specialized domains, motivating research into their application to legal reasoning. However, existing legal benchmarks often conflate factual recall with genuine inference, fragment the reasoning process, and overlook the quality of reasoning. To address these limitations, we introduce MSLR, the first Chinese multi-step legal reasoning dataset grounded in real-world judicial decision making. MSLR adopts the IRAC framework (Issue, Rule, Application, Conclusion) to model structured expert reasoning from official legal documents. In addition, we design a scalable Human-LLM collaborative annotation pipeline that efficiently produces fine-grained step-level reasoning annotations and provides a reusable methodological framework for multi-step reasoning datasets. Evaluation of multiple LLMs on MSLR shows only moderate performance, highlighting the challenges of adapting to complex legal reasoning. Further experiments demonstrate that Self-Initiated Chain-of-Thought prompts generated by models autonomously improve reasoning coherence and quality, outperforming human-designed prompts. MSLR contributes to advancing LLM reasoning and Chain-of-Thought strategies and offers open resources for future research. The dataset and code are available at https://github.com/yuwenhan07/MSLR-Bench and https://law.sjtu.edu.cn/flszyjzx/index.html.

Terence Jie Chua, Wenhan Yu, Jun Zhao et al.

The emergence of foundation models, including language and vision models, has reshaped AI's landscape, offering capabilities across various applications. Deploying and fine-tuning these large models, like GPT-3 and BERT, presents challenges, especially in the current foundation model era. We introduce Emulator-Assisted Tuning (EAT) combined with Parameter-Efficient Fine-Tuning (PEFT) to form Parameter-Efficient Emulator-Assisted Tuning (PEAT). Further, we expand this into federated learning as Federated PEAT (FedPEAT). FedPEAT uses adapters, emulators, and PEFT for federated model tuning, enhancing model privacy and memory efficiency. Adapters adjust pre-trained models, while emulators give a compact representation of original models, addressing both privacy and efficiency. Adaptable to various neural networks, our approach also uses deep reinforcement learning for hyper-parameter optimization. We tested FedPEAT in a unique scenario with a server participating in collaborative federated tuning, showcasing its potential in tackling foundation model challenges.

Wenhan Yu, Terence Jie Chua, Jun Zhao

The efficient deployment and fine-tuning of foundation models are pivotal in contemporary artificial intelligence. In this study, we present a groundbreaking paradigm integrating Mobile Edge Computing (MEC) with foundation models, specifically designed to enhance local task performance on user equipment (UE). Central to our approach is the innovative Emulator-Adapter architecture, segmenting the foundation model into two cohesive modules. This design not only conserves computational resources but also ensures adaptability and fine-tuning efficiency for downstream tasks. Additionally, we introduce an advanced resource allocation mechanism that is fine-tuned to the needs of the Emulator-Adapter structure in decentralized settings. To address the challenges presented by this system, we employ a hybrid multi-agent Deep Reinforcement Learning (DRL) strategy, adept at handling mixed discrete-continuous action spaces, ensuring dynamic and optimal resource allocations. Our comprehensive simulations and validations underscore the practical viability of our approach, demonstrating its robustness, efficiency, and scalability. Collectively, this work offers a fresh perspective on deploying foundation models and balancing computational efficiency with task proficiency.

Wenhan Yu, Jun Zhao

Advanced video technologies are driving the development of the futuristic Metaverse, which aims to connect users from anywhere and anytime. As such, the use cases for users will be much more diverse, leading to a mix of 360-degree videos with two types: non-VR and VR 360-degree videos. This paper presents a novel Quality of Service model for heterogeneous 360-degree videos with different requirements for frame rates and cybersickness. We propose a frame-slotted structure and conduct frame-wise optimization using self-designed differentiated deep reinforcement learning algorithms. Specifically, we design two structures, Separate Input Differentiated Output (SIDO) and Merged Input Differentiated Output (MIDO), for this heterogeneous scenario. We also conduct comprehensive experiments to demonstrate their effectiveness.

Zihang Li, Yuhang Wang, Yikun Zong et al.

Chain-of-Thought (CoT) prompting has significantly enhanced the mathematical reasoning capabilities of Large Language Models. We find existing fine-tuning datasets frequently suffer from the "answer right but reasoning wrong" probelm, where correct final answers are derived from hallucinated, redundant, or logically invalid intermediate steps. This paper proposes EntroCoT, a unified framework for automatically identifying and refining low-quality CoT supervision traces. EntroCoT first proposes an entropy-based mechanism to segment the reasoning trace into multiple steps at uncertain junctures, and then introduces a Monte Carlo rollout-based mechanism to evaluate the marginal contribution of each step. By accurately filtering deceptive reasoning samples, EntroCoT constructs a high-quality dataset where every intermediate step in each reasoning trace facilitates the final answer. Extensive experiments on mathematical benchmarks demonstrate that fine-tuning on the subset constructed by EntroCoT consistently outperforms the baseslines of full-dataset supervision.

Zhewen Tan, Wenhan Yu, Jianfeng Si et al.

In recent years, safety risks associated with large language models have become increasingly prominent, highlighting the urgent need to mitigate the generation of toxic and harmful content. The mainstream paradigm for LLM safety alignment typically adopts a collaborative framework involving three roles: an attacker for adversarial prompt generation, a defender for safety defense, and an evaluator for response assessment. In this paper, we propose a closed-loop reinforcement learning framework called TriPlay-RL that enables iterative and co-improving collaboration among three roles with near-zero manual annotation. Experimental results show that the attacker preserves high output diversity while achieving a 20%-50% improvement in adversarial effectiveness; the defender attains 10%-30% gains in safety performance without degrading general reasoning capability; and the evaluator continuously refines its fine-grained judgment ability through iterations, accurately distinguishing unsafe responses, simple refusals, and useful guidance. Overall, our framework establishes an efficient and scalable paradigm for LLM safety alignment, enabling continuous co-evolution within a unified learning loop.

Wang Chen, Wenhan Yu, Guanqiang Qi et al. · baidu, tsinghua

Retrieval-Augmented Generation (RAG) has become a core paradigm in document question answering tasks. However, existing methods have limitations when dealing with multimodal documents: one category of methods relies on layout analysis and text extraction, which can only utilize explicit text information and struggle to capture images or unstructured content; the other category treats document segmentation as visual input and directly passes it to visual language models (VLMs) for processing, yet it ignores the semantic advantages of text, leading to suboptimal retrieval and generation results. To address these research gaps, we propose the Co-Modality-based RAG (CMRAG) framework, which can simultaneously leverage texts and images for more accurate retrieval and generation. Our framework includes two key components: (1) a Unified Encoding Model (UEM) that projects queries, parsed text, and images into a shared embedding space via triplet-based training, and (2) a Unified Co-Modality-informed Retrieval (UCMR) method that statistically normalizes similarity scores to effectively fuse cross-modal signals. To support research in this direction, we further construct and release a large-scale triplet dataset of (query, text, image) examples. Experiments demonstrate that our proposed framework consistently outperforms single-modality--based RAG in multiple visual document question-answering (VDQA) benchmarks. The findings of this paper show that integrating co-modality information into the RAG framework in a unified manner is an effective approach to improving the performance of complex VDQA systems.

Peizhuang Cong, Wenpu Liu, Wenhan Yu et al.

Large language models (LLMs) have demonstrated remarkable success across various tasks, accompanied by a continuous increase in their parameter size. Parameter-efficient fine-tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA), address the challenges of fine-tuning LLMs by significantly reducing the number of trainable parameters. Recent studies have integrated LoRA with Mixture of Experts (MoE) architectures, leveraging multiple adapter experts and gating mechanisms to further improve fine-tuning performance. However, existing approaches primarily focus on adjusting the allocations of adapter experts per layer to optimize the introduced trainable parameter size, while neglecting a critical factor of adapters' rank. To this end, we propose a hierarchical scheme for expert allocation and rank configuration, HILO, which dynamically adjusts the number and rank of adapter experts across layers, matching the varying representational complexity of model layers in adapter-granularity. Extensive experiments on multiple benchmark tasks demonstrate that HILO outperforms existing methods in accuracy while introducing fewer trainable parameters, providing an efficient and practical solution for fine-tuning LLMs.

Guanhua Chen, Wenhan Yu, Xiao Lu et al.

While Retrieval-Augmented Generation (RAG) plays a crucial role in the application of Large Language Models (LLMs), existing retrieval methods in knowledge-dense domains like law and medicine still suffer from a lack of multi-perspective views, which are essential for improving interpretability and reliability. Previous research on multi-view retrieval often focused solely on different semantic forms of queries, neglecting the expression of specific domain knowledge perspectives. This paper introduces a novel multi-view RAG framework, MVRAG, tailored for knowledge-dense domains that utilizes intention-aware query rewriting from multiple domain viewpoints to enhance retrieval precision, thereby improving the effectiveness of the final inference. Experiments conducted on legal and medical case retrieval demonstrate significant improvements in recall and precision rates with our framework. Our multi-perspective retrieval approach unleashes the potential of multi-view information enhancing RAG tasks, accelerating the further application of LLMs in knowledge-intensive fields.

Wenhan Yu, Wang Chen, Guanqiang Qi et al.

Document Visual Question Answering (DocVQA) is a fundamental task for multimodal document understanding and a key testbed for vision language reasoning. However, most existing DocVQA datasets are limited to the page level and lack fine grained spatial grounding, constraining the interpretability and reasoning capability of Vision Language Models (VLMs). To address this gap, we introduce BBox DocVQA a large scale, bounding box grounded dataset designed to enhance spatial reasoning and evidence localization in visual documents. We further present an automated construction pipeline, Segment Judge and Generate, which integrates a segment model for region segmentation, a VLM for semantic judgment, and another advanced VLM for question answer generation, followed by human verification for quality assurance. The resulting dataset contains 3.6 K diverse documents and 32 K QA pairs, encompassing single and multi region as well as single and multi page scenarios. Each QA instance is grounded on explicit bounding boxes, enabling fine grained evaluation of spatial semantic alignment. Benchmarking multiple state of the art VLMs (e.g., GPT 5, Qwen2.5 VL, and InternVL) on BBox DocVQA reveals persistent challenges in spatial grounding and reasoning accuracy. Furthermore, fine tuning on BBox DocVQA substantially improves both bounding box localization and answer generation, validating its effectiveness for enhancing the reasoning ability of VLMs. Our dataset and code will be publicly released to advance research on interpretable and spatially grounded vision language reasoning.

Wenhan Yu, Terence Jie Chua, Jun Zhao

The Metaverse is gaining attention among academics as maturing technologies empower the promises and envisagements of a multi-purpose, integrated virtual environment. An interactive and immersive socialization experience between people is one of the promises of the Metaverse. In spite of the rapid advancements in current technologies, the computation required for a smooth, seamless and immersive socialization experience in the Metaverse is overbearing, and the accumulated user experience is essential to be considered. The computation burden calls for computation offloading, where the integration of virtual and physical world scenes is offloaded to an edge server. This paper introduces a novel Quality-of-Service (QoS) model for the accumulated experience in multi-user socialization on a multichannel wireless network. This QoS model utilizes deep reinforcement learning approaches to find the near-optimal channel resource allocation. Comprehensive experiments demonstrate that the adoption of the QoS model enhances the overall socialization experience.