We're Cooked! - Probing LLM Political Alignment Via Conflict-Framed Recipe Translation
研究探讨了大型语言模型在翻译含政治倾向术语的食谱时如何展现隐含的政治立场，通过对比不同模型对冲突框架词汇的处理方式揭示其潜在政治倾向。
0 citationsRead paper
Selected work
Representative Papers
Federated Attack Campaign Detection via Contrastive Encoding of Threat Indicators in Gradient Updates
本文提出FedIoC框架，通过联邦学习和对比编码方法解决跨组织网络攻击检测问题，无需直接共享敏感威胁情报。
0 citationsRead paper
LeakageBench: Document-Level Leakage Risk for Redacting Personally Identifiable Information in Document Images
本文针对文档图像中个人身份信息（PII）泄露风险问题，通过构建LeakageBench基准数据集并评估多种检测方法来解决文档级PII去除不彻底的问题。
0 citationsRead paper
Beyond Clear Skies: Synthetic Seasonal and Weather Variations for Real-World Drone Detection
This study addresses the critical challenge of data scarcity in UAV detection under adverse weather and seasonal variations by introducing SynDroneVision-Weather, the first systematic synthetic dataset for urban environments. Leveraging a game engine for high-fidelity rendering and automatic annotation, this dataset enables controllable environmental perturbations across diverse meteorological and seasonal conditions, facilitating clean-to-adverse comparative analysis. Experimental results demonstrate that SynDroneVision-Weather serves as an effective complement to general synthetic data, significantly enhancing the robustness of YOLO-series models against complex appearance changes. Specifically, it effectively reduces both missed detections and false alarm rates in real-world scenarios. These findings validate the pivotal role of domain-specific synthetic data in bridging the sim-to-real gap and improving detection performance under challenging environmental conditions.
0 citationsRead paper
Early Yield Prediction for Sugar Beet Fields using Satellite Data -- Learnings from Specialized Vision Transformers
This study addresses the challenge of accurately predicting final sugar beet yield and effectively identifying low-yielding fields during the early growth stage using remote sensing satellite data. To this end, we propose a novel approach that integrates agronomic domain knowledge with machine learning by developing a customized Vision Transformer model based on Sentinel-2 multispectral imagery. Our design employs an exceptionally small patch size and incorporates all available spectral bands, thereby overcoming limitations of conventional architectures. We further introduce an innovative rank-based mechanism for early detection of underperforming fields, which identifies anomalous plots without requiring absolute yield labels. Experimental results demonstrate that the proposed method consistently detects a substantial proportion of low-yield areas during the initial growth phases across multiple years, exhibiting strong generalization capability and practical applicability.
0 citationsRead paper
Recent publications
Latest Papers
We're Cooked! - Probing LLM Political Alignment Via Conflict-Framed Recipe Translation
研究探讨了大型语言模型在翻译含政治倾向术语的食谱时如何展现隐含的政治立场，通过对比不同模型对冲突框架词汇的处理方式揭示其潜在政治倾向。
0 citationsRead paper
Federated Attack Campaign Detection via Contrastive Encoding of Threat Indicators in Gradient Updates
本文提出FedIoC框架，通过联邦学习和对比编码方法解决跨组织网络攻击检测问题，无需直接共享敏感威胁情报。
0 citationsRead paper
LeakageBench: Document-Level Leakage Risk for Redacting Personally Identifiable Information in Document Images
本文针对文档图像中个人身份信息（PII）泄露风险问题，通过构建LeakageBench基准数据集并评估多种检测方法来解决文档级PII去除不彻底的问题。
0 citationsRead paper
Beyond Clear Skies: Synthetic Seasonal and Weather Variations for Real-World Drone Detection
This study addresses the critical challenge of data scarcity in UAV detection under adverse weather and seasonal variations by introducing SynDroneVision-Weather, the first systematic synthetic dataset for urban environments. Leveraging a game engine for high-fidelity rendering and automatic annotation, this dataset enables controllable environmental perturbations across diverse meteorological and seasonal conditions, facilitating clean-to-adverse comparative analysis. Experimental results demonstrate that SynDroneVision-Weather serves as an effective complement to general synthetic data, significantly enhancing the robustness of YOLO-series models against complex appearance changes. Specifically, it effectively reduces both missed detections and false alarm rates in real-world scenarios. These findings validate the pivotal role of domain-specific synthetic data in bridging the sim-to-real gap and improving detection performance under challenging environmental conditions.
0 citationsRead paper
Early Yield Prediction for Sugar Beet Fields using Satellite Data -- Learnings from Specialized Vision Transformers
This study addresses the challenge of accurately predicting final sugar beet yield and effectively identifying low-yielding fields during the early growth stage using remote sensing satellite data. To this end, we propose a novel approach that integrates agronomic domain knowledge with machine learning by developing a customized Vision Transformer model based on Sentinel-2 multispectral imagery. Our design employs an exceptionally small patch size and incorporates all available spectral bands, thereby overcoming limitations of conventional architectures. We further introduce an innovative rank-based mechanism for early detection of underperforming fields, which identifies anomalous plots without requiring absolute yield labels. Experimental results demonstrate that the proposed method consistently detects a substantial proportion of low-yield areas during the initial growth phases across multiple years, exhibiting strong generalization capability and practical applicability.
0 citationsRead paper