arXiv subjects

Yanchen Yin

Publications and source records attributed to Yanchen Yin.

3 recordsLinked to original sources

arXiv2026

HoloGeo: Mitigating Landmark Bias in Geo-localization via Evidence-Driven Reasoning

Recent advances in Vision-Language Models (VLMs) have significantly improved image geo-localization, yet existing models remain susceptible to landmark bias, causing them to overlook geographical cues or form spurious correlations, ultimately resulting in inaccurate localization. To systematically investigate this issue, we first design two quantitative metrics, Bias Intensity (BI) and Bias Harmfulness (BH), to characterize the impact of landmarks exerted on model reasoning, and establish a comprehensive benchmark, LandmarkBias-3K. To mitigate landmark bias, we further propose an evidence-driven reasoning framework, HoloGeo, to improve the reliability of geo-localization. HoloGeo is supported by a high-quality dataset, BF-30k, annotated with structured multi-evidence bias-free reasoning chains. By incorporating multi-dimensional rewards, HoloGeo explicitly encourages balanced attention over diverse visual cues and achieves evidence-driven joint reasoning. Extensive experiments demonstrate that HoloGeo not only maintains excellent performance on IM2GPS3K and YFCC4k but also significantly outperforms existing open-source VLMs on LandmarkBias-3K, validating its effectiveness for robust geospatial reasoning.

cs.CV↗

arXiv2026

Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models

Jailbreak attacks bypass LLM safety alignment, yet their mechanisms remain poorly understood. We provide evidence that attacks do not comprehensively eliminate safety features, but instead selectively suppress specific attention heads. We identify two functionally differentiated types: Adversarially Compromised Heads (ACHs) concentrated in early layers, which are suppressed under attacks, and Safety-Aligned Heads (SAHs) in mid-layers, which maintain robust activations even when attacks succeed. Ablation studies support the causal role of ACHs and the contribution of SAHs to robust activations: suppressing a small number of ACHs is sufficient to induce jailbreak-like behavior on normally refused inputs, while removing SAHs substantially weakens mid-layer safety activations. Token-level attribution further shows that ACH suppression is driven specifically by attack-template tokens, providing a mechanistic account of why attacks can bypass refusal decisions through ACH suppression while leaving internal safety signals sustained by SAHs -- a phenomenon we term Robust Harmful Features. To validate the practical significance of this robustness, we show that simply reading these persistent activations -- without any training -- yields competitive aggregate detection performance with strong adversarial robustness.

cs.CR↗

arXiv2025

Asynchronous Stochastic Block Projection Algorithm for Solving Linear Systems under Predefined Communication Patterns

This paper proposes an event-triggered asynchronous distributed randomized block Kaczmarz projection (ER-AD-RBKP) algorithm for efficiently solving large-scale linear systems in resource-constrained and communication-unstable environments. The algorithm enables each agent to update its local state estimate independently and engage in communication only when specific triggering conditions are satisfied, thereby significantly reducing communication overhead. At each iteration, agents perform projections using randomly selected partial local data blocks to lower per-iteration computational costs and enhance scalability. By defining events that ensure strong connectivity in the communication graph, we derive the sufficient conditions for global convergence under a probabilistic framework, proving that the algorithm converges exponentially in expectation as long as no extreme events (e.g., permanent agent disconnection) occur. Besides, for inconsistent systems, auxiliary variables are incorporated to transform the problem into an equivalent consistent formulation, and theoretical error bounds are derived. Moreover, we implement the ER-AD-RBKP algorithm in an asynchronous communication environment built on ROS2, a distributed middleware framework for real-time robotic systems. We evaluate the algorithm under various settings, including varying numbers of agents, neighborhood sizes, communication intervals, and failure scenarios such as communication disruptions and processing faults. Experimental results demonstrate the robust performance of the proposed algorithm in terms of computational efficiency, communication cost, and system resilience, highlighting its strong potential for practical applicability in real-world distributed systems.

math.NA↗