arXiv subjects

Wenqing Wang

Publications and source records attributed to Wenqing Wang.

At least 19 recordsLinked to original sources

arXiv2026

SCOUT: Active Information Foraging for Long-Text Understanding with Decoupled Epistemic States

Long-Text Understanding (LTU) at million-token scale requires balancing reasoning fidelity with computational efficiency. Frontier long-context LLMs can process millions of token contexts end-to-end, but they suffer from high token consumption and attention dilution. In parallel, specialized LTU agents often sacrifice fidelity through task-agnostic abstractions like graph construction or indexing. We identify a key insight for LTU: query-relevant information is typically sparse relative to the full document, so effective reasoning should rely on a query-sufficient subset rather than the entire context. To address this, we propose SCOUT, a new paradigm for LTU that shifts from passive processing to active information foraging. It treats the document as an explorable environment and answers from a compact, provenance-grounded epistemic state. Guided by state-level gap diagnosis, SCOUT adaptively alternates between coarse-to-fine exploration and anchored state updates that progressively contract its epistemic state toward query sufficiency. Experiments show that SCOUT matches state-of-the-art proprietary models while reducing token consumption by up to 8x. Moreover, SCOUT remains stable as context length scales, substantially alleviating the practical cost-performance trade-off.

cs.CL↗

arXiv2026

Enhancing Requirements Traceability Link Recovery: A Novel Approach with T-SimCSE

Requirements traceability plays an important role in ensuring software quality and responding to changes in requirements. Requirements trace links (such as the links between requirements and other software artifacts) underpin the modeling and implementation of requirements traceability. With the rapid development of artificial intelligence, more and more pre-trained language models (PLMs) techniques are applied to the automatic recovery of requirements trace links. However, the requirements traceability links recovered by these approaches are not accurate enough, and many approaches require a large labeled dataset for training. Currently, there are very few labeled datasets available. To address these limitations, this paper proposes a novel requirements traceability link recovery approach called T-SimCSE, which is based on a PLM -- SimCSE. SimCSE has the advantages of not requiring labeled data, having broad applicability, and performing well. T-SimCSE firstly uses the SimCSE model to calculate the similarity between requirements and target artifacts, and employs a new metric (i.e. specificity) to reorder those target artifacts. Finally, the trace links are created between the requirement and the top-K target artifacts. We have evaluated T-SimCSE on ten public datasets by comparing them with other approaches. The results show that T-SimCSE achieves superior performance in terms of recall and Mean Average Precision (MAP).

cs.SE↗

arXiv2026

MERGETUNE: Continued Fine-Tuning of Vision-Language Models

Fine-tuning vision-language models (VLMs) such as CLIP often leads to catastrophic forgetting of pretrained knowledge. Prior work primarily aims to mitigate forgetting during adaptation; however, forgetting often remains inevitable during this process. We introduce a novel paradigm, continued fine-tuning (CFT), which seeks to recover pretrained knowledge after a zero-shot model has already been adapted. We propose a simple, model-agnostic CFT strategy (named MERGETUNE) guided by linear mode connectivity (LMC), which can be applied post hoc to existing fine-tuned models without requiring architectural changes. Given a fine-tuned model, we continue fine-tuning its trainable parameters (e.g., soft prompts or linear heads) to search for a continued model which has two low-loss paths to the zero-shot (e.g., CLIP) and the fine-tuned (e.g., CoOp) solutions. By exploiting the geometry of the loss landscape, the continued model implicitly merges the two solutions, restoring pretrained knowledge lost in the fine-tuned counterpart. A challenge is that the vanilla LMC constraint requires data replay from the pretraining task. We approximate this constraint for the zero-shot model via a second-order surrogate, eliminating the need for large-scale data replay. Experiments show that MERGETUNE improves the harmonic mean of CoOp by +5.6% on base-novel generalisation without adding parameters. On robust fine-tuning evaluations, the LMC-merged model from MERGETUNE surpasses ensemble baselines with lower inference cost, achieving further gains and state-of-the-art results when ensembled with the zero-shot model. Our code is available at https://github.com/Surrey-UP-Lab/MERGETUNE.

cs.CV↗

arXiv2025

SagaScale: A Realistic, Scalable, and High-Quality Long-Context Benchmark Built from Full-Length Novels

Large Language Models (LLMs) have shown significant progress, but understanding long and complex documents remains challenging. Many long-context benchmarks have been proposed, but they face several limitations, including task realism, data scalability, and data quality. To this end, we introduce SagaScale, a realistic, scalable, and high-quality long-context benchmark built from full-length novels. The entire benchmark is constructed using an automated data collection pipeline that utilizes external resources (e.g., Wikipedia pages) to curate question-answer pairs. Critically, these external resources are provided only for benchmark construction and not during evaluation, which allows LLMs to curate complex questions that go beyond what they can answer during evaluation. SagaScale is also bilingual and offers the largest context length to date, with average token counts exceeding 250K for English novels and 320K for Chinese novels. Our evaluation across 12 frontier LLMs and three long-context methods -- Na\"ive RAG, Agentic RAG, and Long Context -- yields key insights, including: (1) Directly supplying the full context to the LLM can outperform other methods by a large margin; (2) Most LLMs still struggle with lengthy contexts, but Gemini-2.5-Pro stands out as an exception; and (3) Agentic RAG effectively addresses the retrieval bottleneck in Na\"ive RAG. Finally, we publicly release the SagaScale benchmark and our data collection codebase to facilitate future research.

cs.CL↗

arXiv2025

BeautyGuard: Designing a Multi-Agent Roundtable System for Proactive Beauty Tech Compliance through Stakeholder Collaboration

As generative AI enters enterprise workflows, ensuring compliance with legal, ethical, and reputational standards becomes a pressing challenge. In beauty tech, where biometric and personal data are central, traditional reviews are often manual, fragmented, and reactive. To examine these challenges, we conducted a formative study with six experts (four IT managers, two legal managers) at a multinational beauty company. The study revealed pain points in rule checking, precedent use, and the lack of proactive guidance. Motivated by these findings, we designed a multi-agent "roundtable" system powered by a large language model. The system assigns role-specialized agents for legal interpretation, checklist review, precedent search, and risk mitigation, synthesizing their perspectives into structured compliance advice. We evaluated the prototype with the same experts using System Usability Scale(SUS), The Official NASA Task Load Index(NASA-TLX), and interviews. Results show exceptional usability (SUS: 77.5/100) and minimal cognitive workload, with three key findings: (1) multi-agent systems can preserve tacit knowledge into standardized workflows, (2) information augmentation achieves higher acceptance than decision automation, and (3) successful enterprise AI should mirror organizational structures. This work contributes design principles for human-AI collaboration in compliance review, with broader implications for regulated industries beyond beauty tech.

cs.HC↗

arXiv2025

Dynamic Avatar-Scene Rendering from Human-centric Context

Reconstructing dynamic humans interacting with real-world environments from monocular videos is an important and challenging task. Despite considerable progress in 4D neural rendering, existing approaches either model dynamic scenes holistically or model scenes and backgrounds separately aim to introduce parametric human priors. However, these approaches either neglect distinct motion characteristics of various components in scene especially human, leading to incomplete reconstructions, or ignore the information exchange between the separately modeled components, resulting in spatial inconsistencies and visual artifacts at human-scene boundaries. To address this, we propose {\bf Separate-then-Map} (StM) strategy that introduces a dedicated information mapping mechanism to bridge separately defined and optimized models. Our method employs a shared transformation function for each Gaussian attribute to unify separately modeled components, enhancing computational efficiency by avoiding exhaustive pairwise interactions while ensuring spatial and visual coherence between humans and their surroundings. Extensive experiments on monocular video datasets demonstrate that StM significantly outperforms existing state-of-the-art methods in both visual quality and rendering accuracy, particularly at challenging human-scene interaction boundaries.

cs.CV↗

arXiv2025

A-IPO: Adaptive Intent-driven Preference Optimization

Human preferences are diverse and dynamic, shaped by regional, cultural, and social factors. Existing alignment methods like Direct Preference Optimization (DPO) and its variants often default to majority views, overlooking minority opinions and failing to capture latent user intentions in prompts. To address these limitations, we introduce \underline{\textbf{A}}daptive \textbf{\underline{I}}ntent-driven \textbf{\underline{P}}reference \textbf{\underline{O}}ptimization (\textbf{A-IPO}). Specifically,A-IPO introduces an intention module that infers the latent intent behind each user prompt and explicitly incorporates this inferred intent into the reward function, encouraging stronger alignment between the preferred model's responses and the user's underlying intentions. We demonstrate, both theoretically and empirically, that incorporating an intention--response similarity term increases the preference margin (by a positive shift of $\lambda\,\Delta\mathrm{sim}$ in the log-odds), resulting in clearer separation between preferred and dispreferred responses compared to DPO. For evaluation, we introduce two new benchmarks, Real-pref, Attack-pref along with an extended version of an existing dataset, GlobalOpinionQA-Ext, to assess real-world and adversarial preference alignment. Through explicit modeling of diverse user intents,A-IPO facilitates pluralistic preference optimization while simultaneously enhancing adversarial robustness in preference alignment. Comprehensive empirical evaluation demonstrates that A-IPO consistently surpasses existing baselines, yielding substantial improvements across key metrics: up to +24.8 win-rate and +45.6 Response-Intention Consistency on Real-pref; up to +38.6 Response Similarity and +52.2 Defense Success Rate on Attack-pref; and up to +54.6 Intention Consistency Score on GlobalOpinionQA-Ext.

cs.CL↗

arXiv2025

Dissecting Embodied Abilities in Multimodal Language Models through Skill-level Evaluation and Diagnosis

Understanding the capability bottlenecks of embodied multimodal large language models (MLLMs) is crucial for improving embodied agents. However, existing embodied benchmarks mainly focus on task-level evaluation and fail to provide actionable insights into the underlying causes of model failures. To address this limitation, we introduce BEAR, a benchmark that decomposes embodied tasks into 14 atomic skills for fine-grained skill-level evaluation. BEAR comprises 4,469 interleaved image-video-text samples spanning 14 skills across 6 categories, ranging from low-level perception to high-level planning. We evaluate 20 MLLMs on BEAR under a hierarchical skill-level diagnosis framework and uncover two key findings: (1) perceptual capabilities are major bottlenecks behind reasoning failures, and (2) current models suffer from unstable spatiotemporal modeling that remains largely unexposed in prior benchmarks. Motivated by these findings, we further propose BEAR-Agent, a multimodal conversational agent that augments MLLMs with visual and spatial reasoning tools. BEAR-Agent substantially improves performance across embodied skills, achieving a relative improvement of 17.5% on GPT-5 over the base model on BEAR, while also outperforming strong baselines in both simulation and real-world robotic experiments. Project page: https://bear-official66.github.io/

cs.CV↗

arXiv2025

MARS: Co-evolving Dual-System Deep Research via Multi-Agent Reinforcement Learning

Large Reasoning Models (LRMs) face two fundamental limitations: excessive token consumption when overanalyzing simple information processing tasks, and inability to access up-to-date knowledge beyond their training data. We introduce MARS (Multi-Agent System for Deep ReSearch), a novel co-evolution framework that jointly optimizes dual cognitive systems through multi-agent reinforcement learning. Unlike prior approaches that employ fixed or independently-trained summarizers, MARS enables System 1 (fast, intuitive processing) and System 2 (deliberate reasoning) to co-adapt through shared trajectory rewards, developing complementary strategies where System 1 learns to distill information specifically useful for System 2's reasoning. We extend Group Relative Policy Optimization (GRPO) for multi-agent settings with three key innovations: (1) decoupled gradient computation ensuring proper credit assignment despite shared rewards, (2) bin-packing optimization for efficient parallel information processing, and (3) advantage-weighted balanced sampling preventing training imbalance. Extensive experiments demonstrate that MARS (8B), trained under a challenging Zero RL setting without any supervised fine-tuning, achieves 8.17% on HLE -- outperforming WebThinker (32B with SFT, 6.87%) and narrowing the gap with proprietary models like Claude 3.7 Sonnet (7.89%) -- while achieving an average gain of 8.9% across 7 knowledge-intensive tasks.

cs.AI↗

arXiv2025

Dual-Stage Safe Herding Framework for Adversarial Attacker in Dynamic Environment

Recent advances in robotics have enabled the widespread deployment of autonomous robotic systems in complex operational environments, presenting both unprecedented opportunities and significant security problems. Traditional shepherding approaches based on fixed formations are often ineffective or risky in urban and obstacle-rich scenarios, especially when facing adversarial agents with unknown and adaptive behaviors. This paper addresses this challenge as an extended herding problem, where defensive robotic systems must safely guide adversarial agents with unknown strategies away from protected areas and into predetermined safe regions, while maintaining collision-free navigation in dynamic environments. We propose a hierarchical hybrid framework based on reach-avoid game theory and local motion planning, incorporating a virtual containment boundary and event-triggered pursuit mechanisms to enable scalable and robust multi-agent coordination. Simulation results demonstrate that the proposed approach achieves safe and efficient guidance of adversarial agents to designated regions.

cs.RO↗

arXiv2025

Encouraging Good Processes Without the Need for Good Answers: Reinforcement Learning for LLM Agent Planning

The functionality of Large Language Model (LLM) agents is primarily determined by two capabilities: action planning and answer summarization. The former, action planning, is the core capability that dictates an agent's performance. However, prevailing training paradigms employ end-to-end, multi-objective optimization that jointly trains both capabilities. This paradigm faces two critical challenges: imbalanced optimization objective allocation and scarcity of verifiable data, making it difficult to enhance the agent's planning capability. To address these challenges, we propose Reinforcement Learning with Tool-use Rewards (RLTR), a novel framework that decouples the training process to enable a focused, single-objective optimization of the planning module. Crucially, RLTR introduces a reward signal based on tool-use completeness to directly evaluate the quality of tool invocation sequences. This method offers a more direct and reliable training signal than assessing the final response content, thereby obviating the need for verifiable data. Our experiments demonstrate that RLTR achieves an 8%-12% improvement in planning performance compared to end-to-end baselines. Moreover, this enhanced planning capability, in turn, translates to a 5%-6% increase in the final response quality of the overall agent system.

cs.LG↗

arXiv2025

RealTalk: Realistic Emotion-Aware Lifelike Talking-Head Synthesis

Emotion is a critical component of artificial social intelligence. However, while current methods excel in lip synchronization and image quality, they often fail to generate accurate and controllable emotional expressions while preserving the subject's identity. To address this challenge, we introduce RealTalk, a novel framework for synthesizing emotional talking heads with high emotion accuracy, enhanced emotion controllability, and robust identity preservation. RealTalk employs a variational autoencoder (VAE) to generate 3D facial landmarks from driving audio, which are concatenated with emotion-label embeddings using a ResNet-based landmark deformation model (LDM) to produce emotional landmarks. These landmarks and facial blendshape coefficients jointly condition a novel tri-plane attention Neural Radiance Field (NeRF) to synthesize highly realistic emotional talking heads. Extensive experiments demonstrate that RealTalk outperforms existing methods in emotion accuracy, controllability, and identity preservation, advancing the development of socially intelligent AI systems.

cs.CV↗

arXiv2025

Set Invariance with Probability One for Controlled Diffusion: Score-based Approach

Given a controlled diffusion and a connected, bounded, Lipschitz set, when is it possible to guarantee controlled set invariance with probability one? In this work, we answer this question by deriving the necessary and sufficient conditions for the same in terms of gradients of certain log-likelihoods -- a.k.a. score vector fields -- for two cases: given finite time horizon and infinite time horizon. The deduced conditions comprise a score-based test that provably certifies or falsifies the existence of Markovian controllers for given controlled set invariance problem data. Our results are constructive in the sense when the problem data passes the proposed test, we characterize all controllers guaranteeing the desired set invariance. When the problem data fails the proposed test, there does not exist a controller that can accomplish the desired set invariance with probability one. The computation in the proposed tests involve solving certain Dirichlet boundary value problems, and in the finite horizon case, can also account for additional constraint of hitting a target subset at the terminal time. We illustrate the results using several semi-analytical and numerical examples.

math.OC↗

arXiv2025

Markov Kernels, Distances and Optimal Control: A Parable of Linear Quadratic Non-Gaussian Distribution Steering

For a controllable linear time-varying (LTV) pair $(\boldsymbol{A}_t,\boldsymbol{B}_t)$ and $\boldsymbol{Q}_{t}$ positive semidefinite, we derive the Markov kernel for the It\^{o} diffusion ${\mathrm{d}}\boldsymbol{x}_{t}=\boldsymbol{A}_{t}\boldsymbol{x}_t {\mathrm{d}} t + \sqrt{2}\boldsymbol{B}_{t}{\mathrm{d}}\boldsymbol{w}_{t}$ with an accompanying killing of probability mass at rate $\frac{1}{2}\boldsymbol{x}^{\top}\boldsymbol{Q}_{t}\boldsymbol{x}$. This Markov kernel is the Green's function for an associated linear reaction-advection-diffusion partial differential equation. Our result generalizes the recently derived kernel for the special case $\left(\boldsymbol{A}_t,\boldsymbol{B}_t\right)=\left(\boldsymbol{0},\boldsymbol{I}\right)$, and depends on the solution of an associated Riccati matrix ODE. A consequence of this result is that the linear quadratic non-Gaussian Schr\"{o}dinger bridge is exactly solvable. This means that the problem of steering a controlled LTV diffusion from a given non-Gaussian distribution to another over a fixed deadline while minimizing an expected quadratic cost can be solved using dynamic Sinkhorn recursions performed with the derived kernel. Our derivation for the $\left(\boldsymbol{A}_t,\boldsymbol{B}_t,\boldsymbol{Q}_t\right)$-parametrized kernel pursues a new idea that relies on finding a state-time dependent distance-like functional given by the solution of a deterministic optimal control problem. This technique breaks away from existing methods, such as generalizing Hermite polynomials or Weyl calculus, which have seen limited success in the reaction-diffusion context. Our technique uncovers a new connection between Markov kernels, distances, and optimal control. This connection is of interest beyond its immediate application in solving the linear quadratic Schr\"{o}dinger bridge problem.

math.OC↗

arXiv2025

HandSCS: Structural Coordinate Space for Animatable Hand Gaussian Splatting

Photorealistic and animatable hand avatars are essential for applications such as AR/VR, gaming, and telepresence. Recent 3D Gaussian Splatting (3DGS) based avatar methods enable real-time rendering of articulated humans, but modeling hands remains challenging due to their compact structure, frequent self-occlusions, and complex finger interactions. Existing approaches primarily rely on pose-driven transformations while representing Gaussians in Euclidean space, lacking an explicit structural association with the underlying skeleton, which makes preserving fine-grained hand structures under complex articulations difficult. In this work, we introduce the Structural Coordinate Space (SCS), a skeleton-relative representation that assigns each Gaussian primitive an explicit structural coordinate with respect to the articulated hand skeleton. SCS is constructed using a hybrid static-virtual bone basis together with a distance-rotation structural descriptor that encodes the geometric relationship between Gaussians and bones. Based on SCS, we enforce both intra-pose and cross-pose structural consistency by combining per-Gaussian residual embeddings for local appearance modeling with structural correspondence across poses. Experiments demonstrate that our approach significantly improves structural consistency and preserves fine geometric details under challenging hand articulations compared with existing 3DGS-based avatar methods.

cs.CV↗

arXiv2025

Interpersonal Memory Matters: A New Task for Proactive Dialogue Utilizing Conversational History

Proactive dialogue systems aim to empower chatbots with the capability of leading conversations towards specific targets, thereby enhancing user engagement and service autonomy. Existing systems typically target pre-defined keywords or entities, neglecting user attributes and preferences implicit in dialogue history, hindering the development of long-term user intimacy. To address these challenges, we take a radical step towards building a more human-like conversational agent by integrating proactive dialogue systems with long-term memory into a unified framework. Specifically, we define a novel task named Memory-aware Proactive Dialogue (MapDia). By decomposing the task, we then propose an automatic data construction method and create the first Chinese Memory-aware Proactive Dataset (ChMapData). Furthermore, we introduce a joint framework based on Retrieval Augmented Generation (RAG), featuring three modules: Topic Summarization, Topic Retrieval, and Proactive Topic-shifting Detection and Generation, designed to steer dialogues towards relevant historical topics at the right time. The effectiveness of our dataset and models is validated through both automatic and human evaluations. We release the open-source framework and dataset at https://github.com/FrontierLabs/MapDia.

cs.CL↗

arXiv2025

Ductility mechanisms in complex concentrated refractory alloys from atomistic fracture simulations

The striking variation in damage tolerance among refractory complex concentrated alloys is examined through the analysis of atomistic fracture simulations, contrasting behavior in elemental Nb with that in brittle NbMoTaW and ductile Nb45Ta25Ti15Hf15. We employ machine-learning interatomic potentials (MLIPs), including a new MLIP developed for NbTaTiHf, in atomistic simulations of crack tip extension mechanisms based on analyses of atomistic fracture resistance curves. While the initial behavior of sharp cracks shows good correspondence with the Rice theory, fracture resistance curves reveal marked changes in fracture modes for the complex alloys as crack extension proceeds. In NbMoTaW, compositional complexity appears to promote dislocation nucleation relative to pure Nb, despite theoretical predictions that the alloy should be relatively more brittle. In Nb45Ta25Ti15Hf15, alloying not only changes the fracture mode relative to elemental Nb, but promotes dislocation accumulation at the crack tip, leading to higher resistance to crack propagation.

cond-mat.mtrl-sci↗

arXiv2025

EDA-Q: Electronic Design Automation for Superconducting Quantum Chip

Electronic Design Automation (EDA) plays a crucial role in classical chip design and significantly influences the development of quantum chip design. However, traditional EDA tools cannot be directly applied to quantum chip design due to vast differences compared to the classical realm. Several EDA products tailored for quantum chip design currently exist, yet they only cover partial stages of the quantum chip design process instead of offering a fully comprehensive solution. Additionally, they often encounter issues such as limited automation, steep learning curves, challenges in integrating with actual fabrication processes, and difficulties in expanding functionality. To address these issues, we developed a full-stack EDA tool specifically for quantum chip design, called EDA-Q. The design workflow incorporates functionalities present in existing quantum EDA tools while supplementing critical design stages such as device mapping and fabrication process mapping, which users expect. EDA-Q utilizes a unique architecture to achieve exceptional scalability and flexibility. The integrated design mode guarantees algorithm compatibility with different chip components, while employing a specialized interactive processing mode to offer users a straightforward and adaptable command interface. Application examples demonstrate that EDA-Q significantly reduces chip design cycles, enhances automation levels, and decreases the time required for manual intervention. Multiple rounds of testing on the designed chip have validated the effectiveness of EDA-Q in practical applications.

cs.ET↗