arXiv subjects

Qi Wang

Publications and source records attributed to Qi Wang.

At least 253 records · Page 14Linked to original sources

arXiv2025

DynTok: Dynamic Compression of Visual Tokens for Efficient and Effective Video Understanding

Typical video modeling methods, such as LLava, represent videos as sequences of visual tokens, which are then processed by the LLM backbone for effective video understanding. However, this approach leads to a massive number of visual tokens, especially for long videos. A practical solution is to first extract relevant visual information from the large visual context before feeding it into the LLM backbone, thereby reducing computational overhead. In this work, we introduce DynTok, a novel \textbf{Dyn}amic video \textbf{Tok}en compression strategy. DynTok adaptively splits visual tokens into groups and merges them within each group, achieving high compression in regions with low information density while preserving essential content. Our method reduces the number of tokens to 44.4% of the original size while maintaining comparable performance. It further benefits from increasing the number of video frames and achieves 65.3% on Video-MME and 72.5% on MLVU. By applying this simple yet effective compression method, we expose the redundancy in video token representations and offer insights for designing more efficient video modeling techniques.

cs.CL↗

arXiv2025

Strategy-Augmented Planning for Large Language Models via Opponent Exploitation

Efficiently modeling and exploiting opponents is a long-standing challenge in adversarial domains. Large Language Models (LLMs) trained on extensive textual data have recently demonstrated outstanding performance in general tasks, introducing new research directions for opponent modeling. Some studies primarily focus on directly using LLMs to generate decisions based on the elaborate prompt context that incorporates opponent descriptions, while these approaches are limited to scenarios where LLMs possess adequate domain expertise. To address that, we introduce a two-stage Strategy-Augmented Planning (SAP) framework that significantly enhances the opponent exploitation capabilities of LLM-based agents by utilizing a critical component, the Strategy Evaluation Network (SEN). Specifically, in the offline stage, we construct an explicit strategy space and subsequently collect strategy-outcome pair data for training the SEN network. During the online phase, SAP dynamically recognizes the opponent's strategies and greedily exploits them by searching best response strategy on the well-trained SEN, finally translating strategy to a course of actions by carefully designed prompts. Experimental results show that SAP exhibits robust generalization capabilities, allowing it to perform effectively not only against previously encountered opponent strategies but also against novel, unseen strategies. In the MicroRTS environment, SAP achieves a $85.35\%$ performance improvement over baseline methods and matches the competitiveness of reinforcement learning approaches against state-of-the-art (SOTA) rule-based AI. Our code is available at https://github.com/hsushuai/SAP.

cs.AI↗

arXiv2025

What Makes a Good Reasoning Chain? Uncovering Structural Patterns in Long Chain-of-Thought Reasoning

Recent advances in reasoning with large language models (LLMs) have popularized Long Chain-of-Thought (LCoT), a strategy that encourages deliberate and step-by-step reasoning before producing a final answer. While LCoTs have enabled expert-level performance in complex tasks, how the internal structures of their reasoning chains drive, or even predict, the correctness of final answers remains a critical yet underexplored question. In this work, we present LCoT2Tree, an automated framework that converts sequential LCoTs into hierarchical tree structures and thus enables deeper structural analysis of LLM reasoning. Using graph neural networks (GNNs), we reveal that structural patterns extracted by LCoT2Tree, including exploration, backtracking, and verification, serve as stronger predictors of final performance across a wide range of tasks and models. Leveraging an explainability technique, we further identify critical thought patterns such as over-branching that account for failures. Beyond diagnostic insights, the structural patterns by LCoT2Tree support practical applications, including improving Best-of-N decoding effectiveness. Overall, our results underscore the critical role of internal structures of reasoning chains, positioning LCoT2Tree as a powerful tool for diagnosing, interpreting, and improving reasoning in LLMs.

cs.AI↗

arXiv2025

ScreenExplorer: Training a Vision-Language Model for Diverse Exploration in Open GUI World

The rapid progress of large language models (LLMs) has sparked growing interest in building Artificial General Intelligence (AGI) within Graphical User Interface (GUI) environments. However, existing GUI agents based on LLMs or vision-language models (VLMs) often fail to generalize to novel environments and rely heavily on manually curated, diverse datasets. To overcome these limitations, we introduce ScreenExplorer, a VLM trained via Group Relative Policy Optimization(GRPO) in real, dynamic, and open-ended GUI environments. Innovatively, we introduced a world-model-based curiosity reward function to help the agent overcome the cold-start phase of exploration. Additionally, distilling experience streams further enhances the model's exploration capabilities. Our training framework enhances model exploration in open GUI environments, with trained models showing better environmental adaptation and sustained exploration compared to static deployment models. Our findings offer a scalable pathway toward AGI systems with self-improving capabilities in complex interactive settings.

cs.AI↗

arXiv2025

Guiding the Experts: Semantic Priors for Efficient and Focused MoE Routing

Mixture-of-Experts (MoE) models have emerged as a promising direction for scaling vision architectures efficiently. Among them, Soft MoE improves training stability by assigning each token to all experts via continuous dispatch weights. However, current designs overlook the semantic structure which is implicitly encoded in these weights, resulting in suboptimal expert routing. In this paper, we discover that dispatch weights in Soft MoE inherently exhibit segmentation-like patterns but are not explicitly aligned with semantic regions. Motivated by this observation, we propose a foreground-guided enhancement strategy. Specifically, we introduce a spatially aware auxiliary loss that encourages expert activation to align with semantic foreground regions. To further reinforce this supervision, we integrate a lightweight LayerScale mechanism that improves information flow and stabilizes optimization in skip connections. Our method necessitates only minor architectural adjustments and can be seamlessly integrated into prevailing Soft MoE frameworks. Comprehensive experiments on ImageNet-1K and multiple smaller-scale classification benchmarks not only showcase consistent performance enhancements but also reveal more interpretable expert routing mechanisms.

cs.CV↗

arXiv2025

DualComp: End-to-End Learning of a Unified Dual-Modality Lossless Compressor

Most learning-based lossless compressors are designed for a single modality, requiring separate models for multi-modal data and lacking flexibility. However, different modalities vary significantly in format and statistical properties, making it ineffective to use compressors that lack modality-specific adaptations. While multi-modal large language models (MLLMs) offer a potential solution for modality-unified compression, their excessive complexity hinders practical deployment. To address these challenges, we focus on the two most common modalities, image and text, and propose DualComp, the first unified and lightweight learning-based dual-modality lossless compressor. Built on a lightweight backbone, DualComp incorporates three key structural enhancements to handle modality heterogeneity: modality-unified tokenization, modality-switching contextual learning, and modality-routing mixture-of-experts. A reparameterization training strategy is also used to boost compression performance. DualComp integrates both modality-specific and shared parameters for efficient parameter utilization, enabling near real-time inference (200KB/s) on desktop CPUs. With much fewer parameters, DualComp achieves compression performance on par with the SOTA LLM-based methods for both text and image datasets. Its simplified single-modality variant surpasses the previous best image compressor on the Kodak dataset by about 9% using just 1.2% of the model size.

cs.CV↗

arXiv2025

Clapper: Compact Learning and Video Representation in VLMs

Current vision-language models (VLMs) have demonstrated remarkable capabilities across diverse video understanding applications. Designing VLMs for video inputs requires effectively modeling the temporal dimension (i.e. capturing dependencies across frames) and balancing the processing of short and long videos. Specifically, short videos demand preservation of fine-grained details, whereas long videos require strategic compression of visual information to handle extensive temporal contexts efficiently. However, our empirical analysis reveals a critical limitation: most existing VLMs suffer severe performance degradation in long video understanding tasks when compressing visual tokens below a quarter of their original visual tokens. To enable more effective modeling of both short and long video inputs, we propose Clapper, a method that utilizes a slow-fast strategy for video representation and introduces a novel module named TimePerceiver for efficient temporal-spatial encoding within existing VLM backbones. By using our method, we achieves 13x compression of visual tokens per frame (averaging 61 tokens/frame) without compromising QA accuracy. In our experiments, Clapper achieves 62.0% on VideoMME, 69.8% on MLVU, and 67.4% on TempCompass, all with fewer than 6,000 visual tokens per video. The code will be publicly available on the homepage.

cs.CV↗

arXiv2025

FLARE: Robot Learning with Implicit World Modeling

We introduce $\textbf{F}$uture $\textbf{LA}$tent $\textbf{RE}$presentation Alignment ($\textbf{FLARE}$), a novel framework that integrates predictive latent world modeling into robot policy learning. By aligning features from a diffusion transformer with latent embeddings of future observations, $\textbf{FLARE}$ enables a diffusion transformer policy to anticipate latent representations of future observations, allowing it to reason about long-term consequences while generating actions. Remarkably lightweight, $\textbf{FLARE}$ requires only minimal architectural modifications -- adding a few tokens to standard vision-language-action (VLA) models -- yet delivers substantial performance gains. Across two challenging multitask simulation imitation learning benchmarks spanning single-arm and humanoid tabletop manipulation, $\textbf{FLARE}$ achieves state-of-the-art performance, outperforming prior policy learning baselines by up to 26%. Moreover, $\textbf{FLARE}$ unlocks the ability to co-train with human egocentric video demonstrations without action labels, significantly boosting policy generalization to a novel object with unseen geometry with as few as a single robot demonstration. Our results establish $\textbf{FLARE}$ as a general and scalable approach for combining implicit world modeling with high-frequency robotic control.

cs.RO↗

arXiv2025

Universal expansion of human mobility across urban scales

Human mobility is a fundamental process underpinning socioeconomic life and urban structure. Classic theories, such as egocentric activity spaces and central place theory, provide crucial insights into specific facets of movement, like home-centricity and hierarchical spatial organization. However, identifying universal characteristics or an underlying principle that quantitatively links these disparate perspectives has remained a challenge. Here, we reveal such a connection by analyzing the spatial structure of individual daily mobility trajectories using network-based modules. We discover a universal scaling law: the spatial extent (radius) of these mobility modules expands sublinearly with increasing distance from home, a pattern consistent across three orders of magnitude. Furthermore, we demonstrate that these modules precisely map onto the nested hierarchy of urban systems, corresponding to local, city-level, and regional scales as distance from home increases. These findings deepen our understanding of human mobility dynamics and demonstrate the profound connection between classical urban theory, human geography, and mobility studies.

physics.soc-ph↗

arXiv2025

Switching exploration modes in human mobility

Recent advances in human mobility research have revealed consistent pairwise characteristics in movement behavior, yet existing mobility models often overlook the spatial and topological structure of mobility networks. By analyzing millions of devices' anonymized cell phone trajectories, we uncover a distinct modular organization within these networks, demonstrating that movements within spatial modules differ significantly from those between modules. This finding challenges the conventional assumption of uniform mobility dynamics and underscores the influence of heterogeneous environments on human movement. Inspired by switching behaviors in animal movement patterns, we introduce a novel "switch mechanism" to differentiate movement modes, allowing our model to accurately reproduce both the modular structures of trajectory networks and spatial mobility patterns. Our results provide new insights into the dynamics of human mobility and its impact on network formation, with broad applications in traffic prediction, disease transmission modeling, and urban planning. Beyond advancing the theoretical and practical understanding of mobility networks, this work opens new avenues for understanding societal dynamics at large.

physics.soc-ph↗

arXiv2025

Deeply Nonlinear Magnonic Directional Coupler

Dipolar coupling between closely spaced magnetic waveguides enables the design of magnonic directional couplers - universal devices capable of functioning as signal combiners, power splitters, demultiplexers, and more. The wavelength-dependent coupling, combined with the weak nonlinear variation of a spin wave's wavelength at constant-frequency, introduces power-dependent characteristics of directional couplers. This property has been leveraged in the development of magnonic logic elements and other applications. Here, we explore another nonlinear phenomenon in a directional coupler arising purely from the nonlinear frequency shift of spin waves. We show that a strong nonlinear frequency shift causes the coupler to behave as if composed of nonidentical waveguides, suppressing the energy transfer between the waveguides. The transition from complete to negligible energy transfer exhibits a sharp threshold behavior, where the critical power is determined by the coupling strength and nonlinear frequency shift parameter. Based on these findings, a switchable directional coupler as a critical component for future integrated magnonic circuits is designed and validated by micromagnetic simulations.

cond-mat.mes-hall↗

arXiv2025

Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets

Large language models (LLMs) have shown great potential as general-purpose AI assistants across various domains. To fully leverage this potential in specific applications, many companies provide fine-tuning API services, enabling users to upload their own data for LLM customization. However, fine-tuning services introduce a new safety threat: user-uploaded data, whether harmful or benign, can break the model's alignment, leading to unsafe outputs. Moreover, existing defense methods struggle to address the diversity of fine-tuning datasets (e.g., varying sizes, tasks), often sacrificing utility for safety or vice versa. To address this issue, we propose Safe Delta, a safety-aware post-training defense method that adjusts the delta parameters (i.e., the parameter change before and after fine-tuning). Specifically, Safe Delta estimates the safety degradation, selects delta parameters to maximize utility while limiting overall safety loss, and applies a safety compensation vector to mitigate residual safety loss. Through extensive experiments on four diverse datasets with varying settings, our approach consistently preserves safety while ensuring that the utility gain from benign datasets remains unaffected.

cs.LG↗

arXiv2025

Conversations With The Stressed Body: Facilitating Stress Self-Disclosure Among Adolescent Girls Through An Embodied Approach

Adolescent girls face significant mental health challenges during their transition to adulthood, often experiencing heightened stress from various sources. While various interactive technologies for self-disclosure had been explored to support stress relief, little is known about how to encourage stress-related self-disclosure through an embodied approach. This study presents a co-design workshop centred on Embodied Probes, a series of artefacts and activities incorporating embodied methods and technologies. During the workshop, nine participants aged 15 to 18 engaged with their bodies, expressed bodily sensations through tangible means, and designed embodied prototypes tailored to their personal needs for stress perception and relief. The workshop revealed insights into somatic symptoms, sources, and coping strategies for stress among adolescent girls, as well as how embodied methods can support their stress self-disclosure. This paper contributes to the HCI community by offering design implications on leveraging embodied technologies to support self-disclosure for young women's mental well-being.

cs.HC↗

arXiv2025

Learning Hierarchical Interaction for Accurate Molecular Property Prediction

Discovering molecules with desirable molecular properties, including ADMET profiles, is of great importance in drug discovery. Existing approaches typically employ deep learning models, such as Graph Neural Networks (GNNs) and Transformers, to predict these molecular properties by learning from diverse chemical information. However, these models often fail to efficiently capture and utilize the hierarchical nature of molecular structures, and often lack mechanisms for effective interaction among multi-level features. To address these limitations, we propose a Hierarchical Interaction Message Passing Mechanism, which serves as the foundation of our novel model, the Hierarchical Interaction Message Net (HimNet). Our method enables interaction-aware representation learning across atomic, motif, and molecular levels via hierarchical attention-guided message passing. This design allows HimNet to effectively balance global and local information, ensuring rich and task-relevant feature extraction for downstream property prediction tasks, such as Blood-Brain Barrier Permeability (BBBP). We systematically evaluate HimNet on eleven datasets, including eight widely-used MoleculeNet benchmarks and three challenging, high-value datasets for metabolic stability, malaria activity, and liver microsomal clearance, covering a broad range of pharmacologically relevant properties. Extensive experiments demonstrate that HimNet achieves the best or near-best performance in most molecular property prediction tasks. Furthermore, our method exhibits promising hierarchical interpretability, aligning well with chemical intuition on representative molecules. We believe that HimNet offers an accurate and efficient solution for molecular activity and ADMET property prediction, contributing to advanced decision-making in the early stages of drug discovery.

q-bio.BM↗

arXiv2025

Motion Design Principles for Accessible Video-based Learning: Addressing Cognitive Challenges for Deaf and Hard of Hearing Learners

Deaf and Hard-of-Hearing (DHH) learners face unique challenges in video-based learning due to the complex interplay between visual and auditory information in videos. Traditional approaches to making video content accessible primarily focus on captioning, but these solutions often neglect the cognitive demands of processing both visual and textual information simultaneously. This paper introduces a set of \textit{Motion} design guidelines, aimed at mitigating these cognitive challenges and improving video learning experiences for DHH learners. Through a two-phase research, we identified five key challenges, including misaligned content and visual overload. We proposed five design principles accordingly. User study with 16 DHH participants showed that improving visual-audio relevance and guiding visual attention significantly enhances the learning experience by reducing physical demand, alleviating temporal pressure, and improving learning satisfaction. Our findings highlight the potential of Motion design to transform educational content for DHH learners, and we discuss implications for inclusive video learning tools.

cs.HC↗

arXiv2025

PeSANet: Physics-encoded Spectral Attention Network for Simulating PDE-Governed Complex Systems

Accurately modeling and forecasting complex systems governed by partial differential equations (PDEs) is crucial in various scientific and engineering domains. However, traditional numerical methods struggle in real-world scenarios due to incomplete or unknown physical laws. Meanwhile, machine learning approaches often fail to generalize effectively when faced with scarce observational data and the challenge of capturing local and global features. To this end, we propose the Physics-encoded Spectral Attention Network (PeSANet), which integrates local and global information to forecast complex systems with limited data and incomplete physical priors. The model consists of two key components: a physics-encoded block that uses hard constraints to approximate local differential operators from limited data, and a spectral-enhanced block that captures long-range global dependencies in the frequency domain. Specifically, we introduce a novel spectral attention mechanism to model inter-spectrum relationships and learn long-range spatial features. Experimental results demonstrate that PeSANet outperforms existing methods across all metrics, particularly in long-term forecasting accuracy, providing a promising solution for simulating complex systems with limited data and incomplete physics.

cs.LG↗

arXiv2025

Leveraging Pretrained Diffusion Models for Zero-Shot Part Assembly

3D part assembly aims to understand part relationships and predict their 6-DoF poses to construct realistic 3D shapes, addressing the growing demand for autonomous assembly, which is crucial for robots. Existing methods mainly estimate the transformation of each part by training neural networks under supervision, which requires a substantial quantity of manually labeled data. However, the high cost of data collection and the immense variability of real-world shapes and parts make traditional methods impractical for large-scale applications. In this paper, we propose first a zero-shot part assembly method that utilizes pre-trained point cloud diffusion models as discriminators in the assembly process, guiding the manipulation of parts to form realistic shapes. Specifically, we theoretically demonstrate that utilizing a diffusion model for zero-shot part assembly can be transformed into an Iterative Closest Point (ICP) process. Then, we propose a novel pushing-away strategy to address the overlap parts, thereby further enhancing the robustness of the method. To verify our work, we conduct extensive experiments and quantitative comparisons to several strong baseline methods, demonstrating the effectiveness of the proposed approach, which even surpasses the supervised learning method. The code has been released on https://github.com/Ruiyuan-Zhang/Zero-Shot-Assembly.

cs.CV↗

arXiv2025

Impact of Initial-State Nuclear and Sub-Nucleon Structures on Ultra-Central Puzzle in Heavy Ion Collisions

Hydrodynamic models fail to describe the near-equal $v_2/v_3$ ratio observed in ultra-central heavy-ion collisions, despite their success in other centrality classes. This discrepancy stems from shear viscosity suppressing higher-order geometric eccentricities, resulting in underestimated $v_3$ when using the conventional QGP viscosity coefficient. We explore two initial-state modifications to resolve this puzzle: (1) enforcing a minimum nucleon separation distance to homogenize distributions, and (2) amplifying sub-nucleon structures to reduce initial eccentricity. Using TRENTo initial conditions and 3+1D viscous hydrodynamic model CLVisc, both approaches significantly lower geometric eccentricity, reduce required viscosity, and narrow the $v_2$-$v_3$ gap in ultra-central collisions. Our results implicate initial-state nuclear and sub-nucleon structures as critical factors in addressing this puzzle. Resolving it would advance nuclear structure studies and improve precision in extracting QGP transport coefficients (e.g., shear viscosity), bridging microscopic nuclear features to macroscopic quark-gluon plasma properties.

nucl-th↗