arXiv subjects

Xinyi Wu

Publications and source records attributed to Xinyi Wu.

At least 19 recordsLinked to original sources

arXiv2026

Generalizable and Computational Efficient Channel Extrapolation for 6G: A Configurable AI-Driven Framework Built from a Modular Perspective

Acquiring channel state information (CSI) with manageable overhead has been essential to provide high-performance communication services, which is extremely challenging in the emerging sixth generation (6G) mobile network. Channel extrapolation has been proposed to infer complete CSI using a small portion of known CSI, its performance can be dramatically enhanced by artificial intelligence (AI). However, AI-driven channel extrapolation suffers from poor generalization across scenarios and high computational complexity, which is common in the broad research of AI and large language models. Inspired by the modular function of human brain, we propose a configurable AI-driven framework to achieve generalizable and computational efficient channel extrapolation from a modular perspective. We propose a three-stage framework, consisting of experts emergent, experts construction and experts selection. This framework assumes that CSI correlations can be captured by a small number of specialized functional modules (experts) that are activated differently across scenarios. Such modularity emerges in the experts emergent stage via pre-training using CSI data covering comprehensive scenarios. The neurons with similar weight-space patterns are grouped as experts in the experts construction stage. A lightweight gating function is added to control the routing of experts and is fine-tuned for each scenario in the experts selection stage. Simulation results demonstrate that the proposed three-stage framework reduce the channel extrapolation error and computational complexities dramatically by $1.1-19.1$ db and $38$ \%, respectively. In addition, attributed to the proposed experts emergent and section modules, the proposed framework outperforms its counterpart mix-of-expert model dramatically in terms of channel extrapolation performance.

eess.SP↗

arXiv2026

How Data Shapes RoPE Frequency Usage: From Positional Scale Matching to Length Generalization

Rotary Position Embeddings (RoPE) provide transformers with a fixed grid of positional frequencies, yet trained models use these frequencies highly non-uniformly. We study what determines this frequency usage and propose a data-centered explanation: RoPE frequencies are selected to match the relative-distance structure of the training data. Viewing each frequency as a positional lens, we formalize a field-resolution tradeoff and show that, for a data-induced dependency profile of width $W$, the optimal frequency scales as $1/W$. This frequency-matching principle explains controlled observations on synthetic and text-based data, and suggests that the mid-low frequency bands observed in language models arise from the multi-scale dependency structure of natural language. We further connect frequency selection to position-interpolation-based length generalization: scaling frequencies down expands the effective field while reducing resolution. This helps when longer-context dependencies are approximate dilations of those seen during training, but can fail when relevant dependencies do not scale with context length. Empirically, we show that natural language exhibits approximate self-similarity across positional scales, explaining why test-time frequency scaling can support long-context generalization. Overall, our results identify a data-driven mechanism behind emergent RoPE frequency usage and show that long-context generalization depends on two forms of scale matching: between learned frequencies and training-time dependencies, and between frequency scaling and how those dependencies extend to longer contexts.

cs.LG↗

arXiv2026

ChronoPhyBench: Do MLLMs Truly Understand the World or Merely Exploit Language Priors?

Recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated remarkable proficiency in open-world reasoning and understanding. However, a critical ambiguity persists: it remains unclear whether these models genuinely synthesize cross-modal information to construct physically grounded reasoning chains, or if they merely exploit strong language priors to mask single-modality reliance, thereby hallucinating advanced multimodal capabilities. Motivated by this, and to rigorously mitigate language modality bias and shortcuts, we propose a novel multimodal Chrono}logical Physical Dynamics Reasoning Benchmark ChronoPhyBench, which unifies next state prediction with Visual Question Answering (VQA) paradigms by conditioning on historical video context and textual captions to enforce models to deduce subsequent physical states through both single image selection and the inherently more complex task of multiple frame chronological sorting. Concurrently, we construct a large-scale multimodal reasoning dataset curated using the ChronoPhyBench criteria, comprising over 10,000 long-form videos paired with meticulously annotated captions, totaling 5M tokens. Our experimental evaluations reveal a stark contrast to conclusions drawn by previous benchmarks. The capacity of current open-source models to perform physically grounded multimodal reasoning remains in its infancy. Ultimately, this work seeks to systematically stress-test the reasoning capabilities of multimodal models, quantify hallucination rates, and advance the development of Physical AI, thereby providing the community with a robust and transparent evaluation framework toward Artificial General Intelligence (AGI).

cs.CV↗

arXiv2026

EduStory: A Unified Framework for Pedagogically-Consistent Multi-Shot STEM Instructional Video Generation

Long-horizon video generation has advanced in visual quality, yet existing methods still struggle to maintain knowledge consistency and coherent pedagogical narratives across multi-shot instructional videos, especially in STEM domains. To address these challenges, we propose EduStory, a unified framework for reliable instructional video generation. EduStory integrates pedagogical state modeling to track persistent knowledge states, script-guided structured control to organize multi-shot narratives, and learning-oriented evaluation metrics to assess knowledge fidelity and constraint satisfaction. To support rigorous evaluation, we further introduce EduVideoBench, a diagnostic benchmark with multi-granularity annotations, including pedagogical storyboards, shot-level semantics, and knowledge state transitions, together with baseline tasks for controllable instructional video generation. Extensive experiments demonstrate that domain-aware state modeling and structured control substantially reduce narrative breakdown and improve alignment with instructional intent. These results highlight the significance of domain-specific structural constraints and tailored benchmarks for advancing reliable, controllable, and also trustworthy long-horizon video generation.

cs.CV↗

arXiv2026

Unveiling the Resilience of LLM-Enhanced Search Engines against Black-Hat SEO Manipulation

The emergence of Large Language Model-enhanced Search Engines (LLMSEs) has revolutionized information retrieval by integrating web-scale search capabilities with AI-powered summarization. While these systems demonstrate improved efficiency over traditional search engines, their security implications against well-established black-hat Search Engine Optimization (SEO) attacks remain unexplored. In this paper, we present the first systematic study of SEO attacks targeting LLMSEs. Specifically, we examine ten representative LLMSE products (e.g., ChatGPT, Gemini) and construct SEO-Bench, a benchmark comprising 1,000 real-world black-hat SEO websites, to evaluate both open- and closed-source LLMSEs. Our measurements show that LLMSEs mitigate over 99.78% of traditional SEO attacks, with the phase of retrieval serving as the primary filter, intercepting the vast majority of malicious queries. We further propose and evaluate seven LLMSEO attack strategies, demonstrating that off-the-shelf LLMSEs are vulnerable to LLMSEO attacks, i.e., rewritten-query stuffing and segmented texts double the manipulation rate compared to the baseline. This work offers the first in-depth security analysis of the LLMSE ecosystem, providing practical insights for building more resilient AI-driven search systems. We have responsibly reported the identified issues to major vendors.

cs.CR↗

arXiv2026

AI/ML for mobile networks: Current status in Rel. 19 and challenges ahead

The transformative power of artificial intelligence (AI) and machine learning (ML) is recognized as a key enabler for sixth generation (6G) mobile networks by both academia and industry. Research on AI/ML in mobile networks has been ongoing for years, and the 3rd generation partnership project (3GPP) launched standardization efforts to integrate AI into mobile networks. However, a comprehensive review of the current status and challenges of the standardization of AI/ML for mobile networks is still missing. To this end, we provided a comprehensive review of the standardization efforts by 3GPP on AI/ML for mobile networks. This includes an overview of the general AI/ML framework, representative use cases (i.e., CSI feedback, beam management and positioning), and corresponding evaluation matrices. We emphasized the key research challenges on dataset preparation, generalization evaluation and baseline AI/ML models selection. Using CSI feedback as a case study, given the test dataset 2, we demonstrated that the pre-training-fine-tuning paradigm (i.e., pre-training using dataset 1 and fine-tuning using dataset 2) outperforms training on dataset 2. Moreover, we observed the highest performance enhancements in Transformer-based models through fine-tuning, showing its great generalization potential at large floating-point operations (FLOPs). Finally, we outlined future research directions for the application of AI/ML in mobile networks.

eess.SP↗

arXiv2026

Channel Extrapolation for MIMO Systems with the Assistance of Multi-path Information Induced from Channel State Information

Acquiring channel state information (CSI) through traditional methods, such as channel estimation, is increasingly challenging for the emerging sixth generation (6G) mobile networks due to high overhead. To address this issue, channel extrapolation techniques have been proposed to acquire complete CSI from a limited number of known CSIs. To improve extrapolation accuracy, environmental information, such as visual images or radar data, has been utilized, which poses challenges including additional hardware, privacy and multi-modal alignment concerns. To this end, this paper proposes a novel channel extrapolation framework by leveraging environment-related multi-path characteristics induced directly from CSI without integrating additional modalities. Specifically, we propose utilizing the multi-path characteristics in the form of power-delay profile (PDP), which is acquired using a CSI-to-PDP module. CSI-to-PDP module is trained in an AE-based framework by reconstructing the PDPs and constraining the latent low-dimensional features to represent the CSI. We further extract the total power & power-weighted delay of all the identified paths in PDP as the multi-path information. Building on this, we proposed a MAE architecture trained in a self-supervised manner to perform channel extrapolation. Unlike standard MAE approaches, our method employs separate encoders to extract features from the masked CSI and the multi-path information, which are then fused by a cross-attention module. Extensive simulations demonstrate that this framework improves extrapolation performance dramatically, with a minor increase in inference time (around 0.1 ms). Furthermore, our model shows strong generalization capabilities, particularly when only a small portion of the CSI is known, outperforming existing benchmarks.

eess.SP↗

arXiv2026

WebTrap Park: An Automated Platform for Systematic Security Evaluation of Web Agents

Web Agents are increasingly deployed to perform complex tasks in real web environments, yet their security evaluation remains fragmented and difficult to standardize. We present WebTrap Park, an automated platform for systematic security evaluation of Web Agents through direct observation of their concrete interactions with live web pages. WebTrap Park instantiates three major sources of security risk into 1,226 executable evaluation tasks and enables action based assessment without requiring agent modification. Our results reveal clear security differences across agent frameworks, highlighting the importance of agent architecture beyond the underlying model. WebTrap Park is publicly accessible at https://security.fudan.edu.cn/webagent and provides a scalable foundation for reproducible Web Agent security evaluation.

cs.AI↗

arXiv2026

When Bots Take the Bait: Exposing and Mitigating the Emerging Social Engineering Attack in Web Automation Agent

Web agents, powered by large language models (LLMs), are increasingly deployed to automate complex web interactions. The rise of open-source frameworks (e.g., Browser Use, Skyvern-AI) has accelerated adoption, but also broadened the attack surface. While prior research has focused on model threats such as prompt injection and backdoors, the risks of social engineering remain largely unexplored. We present the first systematic study of social engineering attacks against web automation agents and design a pluggable runtime mitigation solution. On the attack side, we introduce the AgentBait paradigm, which exploits intrinsic weaknesses in agent execution: inducement contexts can distort the agent's reasoning and steer it toward malicious objectives misaligned with the intended task. On the defense side, we propose SUPERVISOR, a lightweight runtime module that enforces environment and intention consistency alignment between webpage context and intended goals to mitigate unsafe operations before execution. Empirical results show that mainstream frameworks are highly vulnerable to AgentBait, with an average attack success rate of 67.5% and peaks above 80% under specific strategies (e.g., trusted identity forgery). Compared with existing lightweight defenses, our module can be seamlessly integrated across different web automation frameworks and reduces attack success rates by up to 78.1% on average while incurring only a 7.7% runtime overhead and preserving usability. This work reveals AgentBait as a critical new threat surface for web agents and establishes a practical, generalizable defense, advancing the security of this rapidly emerging ecosystem. We reported the details of this attack to the framework developers and received acknowledgment before submission.

cs.CR↗

arXiv2026

AI-Driven Channel State Information (CSI) Extrapolation for 6G: Current Situations, Challenges and Future Research

CSI extrapolation is an effective method for acquiring channel state information (CSI), essential for optimizing performance of sixth-generation (6G) communication systems. Traditional channel estimation methods face scalability challenges due to the surging overhead in emerging high-mobility, extremely large-scale multiple-input multiple-output (EL-MIMO), and multi-band systems. CSI extrapolation techniques mitigate these challenges by using partial CSI to infer complete CSI, significantly reducing overhead. Despite growing interest, a comprehensive review of state-of-the-art (SOTA) CSI extrapolation techniques is lacking. This paper addresses this gap by comprehensively reviewing the current status, challenges, and future directions of CSI extrapolation for the first time. Firstly, we analyze the performance metrics specific to CSI extrapolation in 6G, including extrapolation accuracy, adaption to dynamic scenarios and algorithm costs. We then review both model-driven and artificial intelligence (AI)-driven approaches for time, frequency, antenna, and multi-domain CSI extrapolation. Key insights and takeaways from these methods are summarized. Given the promise of AI-driven methods in meeting performance requirements, we also examine the open-source channel datasets and simulators that could be used to train high-performance AI-driven CSI extrapolation models. Finally, we discuss the critical challenges of the existing research and propose perspective research opportunities.

eess.SP↗

arXiv2025

Rethinking Reasoning: A Survey on Reasoning-based Backdoors in LLMs

With the rise of advanced reasoning capabilities, large language models (LLMs) are receiving increasing attention. However, although reasoning improves LLMs' performance on downstream tasks, it also introduces new security risks, as adversaries can exploit these capabilities to conduct backdoor attacks. Existing surveys on backdoor attacks and reasoning security offer comprehensive overviews but lack in-depth analysis of backdoor attacks and defenses targeting LLMs' reasoning abilities. In this paper, we take the first step toward providing a comprehensive review of reasoning-based backdoor attacks in LLMs by analyzing their underlying mechanisms, methodological frameworks, and unresolved challenges. Specifically, we introduce a new taxonomy that offers a unified perspective for summarizing existing approaches, categorizing reasoning-based backdoor attacks into associative, passive, and active. We also present defense strategies against such attacks and discuss current challenges alongside potential directions for future research. This work offers a novel perspective, paving the way for further exploration of secure and trustworthy LLM communities.

cs.CR↗

arXiv2025

P2P: A Poison-to-Poison Remedy for Reliable Backdoor Defense in LLMs

During fine-tuning, large language models (LLMs) are increasingly vulnerable to data-poisoning backdoor attacks, which compromise their reliability and trustworthiness. However, existing defense strategies suffer from limited generalization: they only work on specific attack types or task settings. In this study, we propose Poison-to-Poison (P2P), a general and effective backdoor defense algorithm. P2P injects benign triggers with safe alternative labels into a subset of training samples and fine-tunes the model on this re-poisoned dataset by leveraging prompt-based learning. This enforces the model to associate trigger-induced representations with safe outputs, thereby overriding the effects of original malicious triggers. Thanks to this robust and generalizable trigger-based fine-tuning, P2P is effective across task settings and attack types. Theoretically and empirically, we show that P2P can neutralize malicious backdoors while preserving task performance. We conduct extensive experiments on classification, mathematical reasoning, and summary generation tasks, involving multiple state-of-the-art LLMs. The results demonstrate that our P2P algorithm significantly reduces the attack success rate compared with baseline models. We hope that the P2P can serve as a guideline for defending against backdoor attacks and foster the development of a secure and trustworthy LLM community.

cs.CR↗

arXiv2025

DevilSight: Augmenting Monocular Human Avatar Reconstruction through a Virtual Perspective

We present a novel framework to reconstruct human avatars from monocular videos. Recent approaches have struggled either to capture the fine-grained dynamic details from the input or to generate plausible details at novel viewpoints, which mainly stem from the limited representational capacity of the avatar model and insufficient observational data. To overcome these challenges, we propose to leverage the advanced video generative model, Human4DiT, to generate the human motions from alternative perspective as an additional supervision signal. This approach not only enriches the details in previously unseen regions but also effectively regularizes the avatar representation to mitigate artifacts. Furthermore, we introduce two complementary strategies to enhance video generation: To ensure consistent reproduction of human motion, we inject the physical identity into the model through video fine-tuning. For higher-resolution outputs with finer details, a patch-based denoising algorithm is employed. Experimental results demonstrate that our method outperforms recent state-of-the-art approaches and validate the effectiveness of our proposed strategies.

cs.CV↗

arXiv2025

An Evaluation of DUSt3R/MASt3R/VGGT 3D Reconstruction on Photogrammetric Aerial Blocks

State-of-the-art 3D computer vision algorithms continue to advance in handling sparse, unordered image sets. Recently developed foundational models for 3D reconstruction, such as Dense and Unconstrained Stereo 3D Reconstruction (DUSt3R), Matching and Stereo 3D Reconstruction (MASt3R), and Visual Geometry Grounded Transformer (VGGT), have attracted attention due to their ability to handle very sparse image overlaps. Evaluating DUSt3R/MASt3R/VGGT on typical aerial images matters, as these models may handle extremely low image overlaps, stereo occlusions, and textureless regions. For redundant collections, they can accelerate 3D reconstruction by using extremely sparsified image sets. Despite tests on various computer vision benchmarks, their potential on photogrammetric aerial blocks remains unexplored. This paper conducts a comprehensive evaluation of the pre-trained DUSt3R/MASt3R/VGGT models on the aerial blocks of the UseGeo dataset for pose estimation and dense 3D reconstruction. Results show these methods can accurately reconstruct dense point clouds from very sparse image sets (fewer than 10 images, up to 518 pixels resolution), with completeness gains up to +50% over COLMAP. VGGT also demonstrates higher computational efficiency, scalability, and more reliable camera pose estimation. However, all exhibit limitations with high-resolution images and large sets, as pose reliability declines with more images and geometric complexity. These findings suggest transformer-based methods cannot fully replace traditional SfM and MVS, but offer promise as complementary approaches, especially in challenging, low-resolution, and sparse scenarios.

cs.CV↗

arXiv2025

Towards channel foundation models (CFMs): Motivations, methodologies and opportunities

Artificial intelligence (AI) has emerged as a pivotal enabler for next-generation wireless communication systems. However, conventional AI-based models encounter several limitations, such as heavy reliance on labeled data, limited generalization capability, and task-specific design. To address these challenges, this paper introduces, for the first time, the concept of channel foundation models (CFMs)-a novel and unified framework designed to tackle a wide range of channel-related tasks through a pretrained, universal channel feature extractor. By leveraging advanced AI architectures and self-supervised learning techniques, CFMs are capable of effectively exploiting large-scale unlabeled data without the need for extensive manual annotation. We further analyze the evolution of AI methodologies, from supervised learning and multi-task learning to self-supervised learning, emphasizing the distinct advantages of the latter in facilitating the development of CFMs. Additionally, we provide a comprehensive review of existing studies on self-supervised learning in this domain, categorizing them into generative, discriminative and the combined paradigms. Given that the research on CFMs is still at an early stage, we identify several promising future research directions, focusing on model architecture innovation and the construction of high-quality, diverse channel datasets.

eess.SP↗

arXiv2025

Affective-ROPTester: Capability and Bias Analysis of LLMs in Predicting Retinopathy of Prematurity

Despite the remarkable progress of large language models (LLMs) across various domains, their capacity to predict retinopathy of prematurity (ROP) risk remains largely unexplored. To address this gap, we introduce a novel Chinese benchmark dataset, termed CROP, comprising 993 admission records annotated with low, medium, and high-risk labels. To systematically examine the predictive capabilities and affective biases of LLMs in ROP risk stratification, we propose Affective-ROPTester, an automated evaluation framework incorporating three prompting strategies: Instruction-based, Chain-of-Thought (CoT), and In-Context Learning (ICL). The Instruction scheme assesses LLMs' intrinsic knowledge and associated biases, whereas the CoT and ICL schemes leverage external medical knowledge to enhance predictive accuracy. Crucially, we integrate emotional elements at the prompt level to investigate how different affective framings influence the model's ability to predict ROP and its bias patterns. Empirical results derived from the CROP dataset yield two principal observations. First, LLMs demonstrate limited efficacy in ROP risk prediction when operating solely on intrinsic knowledge, yet exhibit marked performance gains when augmented with structured external inputs. Second, affective biases are evident in the model outputs, with a consistent inclination toward overestimating medium- and high-risk cases. Third, compared to negative emotions, positive emotional framing contributes to mitigating predictive bias in model outputs. These findings highlight the critical role of affect-sensitive prompt engineering in enhancing diagnostic reliability and emphasize the utility of Affective-ROPTester as a framework for evaluating and mitigating affective bias in clinical language modeling systems.

cs.AI↗

arXiv2025

From Query to Explanation: Uni-RAG for Multi-Modal Retrieval-Augmented Learning in STEM

In AI-facilitated teaching, leveraging various query styles to interpret abstract educational content is crucial for delivering effective and accessible learning experiences. However, existing retrieval systems predominantly focus on natural text-image matching and lack the capacity to address the diversity and ambiguity inherent in real-world educational scenarios. To address this limitation, we develop a lightweight and efficient multi-modal retrieval module, named Uni-Retrieval, which extracts query-style prototypes and dynamically matches them with tokens from a continually updated Prompt Bank. This Prompt Bank encodes and stores domain-specific knowledge by leveraging a Mixture-of-Expert Low-Rank Adaptation (MoE-LoRA) module and can be adapted to enhance Uni-Retrieval's capability to accommodate unseen query types at test time. To enable natural language educational content generation, we integrate the original Uni-Retrieval with a compact instruction-tuned language model, forming a complete retrieval-augmented generation pipeline named Uni-RAG. Given a style-conditioned query, Uni-RAG first retrieves relevant educational materials and then generates human-readable explanations, feedback, or instructional content aligned with the learning objective. Experimental results on SER and other multi-modal benchmarks show that Uni-RAG outperforms baseline retrieval and RAG systems in both retrieval accuracy and generation quality, while maintaining low computational cost. Our framework provides a scalable, pedagogically grounded solution for intelligent educational systems, bridging retrieval and generation to support personalized, explainable, and efficient learning assistance across diverse STEM scenarios.

cs.AI↗

arXiv2025

The effects of the orbital configurations of mergers on reshaping galaxy structures

We performed a systematic analysis of how the orbital configuration of a merger can influence the structural formation of remnant galaxies using 531 merger pairs selected from IllustrisTNG-100. We comprehensively described the merger orbital configuration, considering the relative orbital motion of the merger pair and their disk orientations. We quantified the galaxy structures by dynamically defining four components: bulge, disk, warm component, and hot inner stellar halo. For mergers on spiral-in orbits, the disk planes of the two merging galaxies tend to be aligned with the orbital plane, leading to higher fractions for the disk and warm components, as well as lower fractions for the bulge and hot inner stellar halo components in the remnant galaxy. For mergers on direct collision orbits, the disk planes of the two galaxies tend to be perpendicular to the orbital plane, leading to lower fractions for disk and warm components, as well as higher fractions of the bulge and hot inner stellar halo in the remnant. Mergers can lead to either an increase or decrease in the disk and bulge mass fraction in the remnant compared to the progenitor galaxy, depending on the merger orbital configurations; however, in 93% of cases, mergers cause an increase in the hot inner stellar halo. As a result, the luminosity fraction of the hot inner stellar halo (but not the bulge) in galaxies at $z=0$ is highly correlated with its total ex situ stellar mass. In addition, we find that merger on spiral-in orbits is one of the possible reasons for the formation of recently discovered red but HI-rich (RR) galaxies.

astro-ph.GA↗