Decision-Oriented Uncertainty Quantification for Risk Control in Earth System Spatiotemporal Foundation Models
本文提出了一种面向决策的不确定性量化框架，用于地球系统时空基础模型的风险控制，通过预测分布和决策风险适配器来提高高影响应用中的决策可靠性。
0 citationsRead paper
Selected work
Representative Papers
MedTRACE: Tool-Augmented Multimodal Clinical Reasoning Agents for Evidence-Grounded Decision-Making
本文提出MedTRACE，通过多模态编码、工具辅助决策和证据验证来提高临床决策的准确性、可解释性和可靠性。
0 citationsRead paper
AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition
This work addresses the challenges of on-device speech emotion recognition, where large self-supervised models incur prohibitive computational costs, and existing knowledge distillation approaches suffer from unreliable teacher predictions and neglect of inter-sample relational structures. To overcome these limitations, the authors propose an adaptive multi-teacher relational distillation framework. This framework incorporates a one-class SVM–based mechanism to assess teacher reliability and dynamically weight their predictions, while also introducing a relational distillation loss that preserves cross-sample semantic structure by aligning similarity matrices between teachers and students. Experiments on IEMOCAP and CREMA-D demonstrate that four lightweight student models consistently outperform single-teacher baselines, and ablation studies confirm the complementary benefits of the two core components.
0 citationsRead paper
Comment and Control: Hijacking Agentic Workflows via Context-Grounded Evolution
This work demonstrates that large language model (LLM) agents integrated into automation platforms are vulnerable to manipulation via malicious inputs—such as GitHub comments—leading to risks including credential leakage and arbitrary command execution. To systematically uncover these novel attack surfaces in agent workflows, the authors propose “Context-Anchored Evolution,” a method that combines static path feasibility analysis, dynamic prompt provenance tracing, and runtime capability assessment within a three-stage contextual framework, enhanced by evolutionary input generation to achieve targeted hijacking of LLM agents. The resulting JAW framework successfully exploits 4,714 GitHub workflows and 8 n8n templates, compromising 15 widely used GitHub Actions and 2 official n8n nodes. Vulnerabilities identified through this approach have been acknowledged and patched by vendors including GitHub, Google, and Anthropic.
0 citationsRead paper
IIR-VLM: In-Context Instance-level Recognition for Large Vision-Language Models
This work addresses the challenge that large vision-language models struggle to distinguish individual instances—such as specific people or objects—thereby limiting their applicability in personalized scenarios. To overcome this limitation, the authors propose an auxiliary visual encoding mechanism that leverages a pre-trained instance-level recognition expert model to provide specialized features to the large vision-language model. This enables the model to perform few-shot, context-aware learning and achieve fine-grained understanding of novel instances from a single example, without requiring extensive instance-specific data or additional training. The approach is the first to enable in-context one-shot instance-level recognition in large vision-language models and supports cross-category instance perception. Evaluated on both existing and newly curated multi-category benchmarks—including faces, persons, pets, and general objects—the method significantly outperforms current state-of-the-art approaches.
0 citationsRead paper
Recent publications
Latest Papers
Decision-Oriented Uncertainty Quantification for Risk Control in Earth System Spatiotemporal Foundation Models
本文提出了一种面向决策的不确定性量化框架，用于地球系统时空基础模型的风险控制，通过预测分布和决策风险适配器来提高高影响应用中的决策可靠性。
0 citationsRead paper
MedTRACE: Tool-Augmented Multimodal Clinical Reasoning Agents for Evidence-Grounded Decision-Making
本文提出MedTRACE，通过多模态编码、工具辅助决策和证据验证来提高临床决策的准确性、可解释性和可靠性。
0 citationsRead paper
AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition
This work addresses the challenges of on-device speech emotion recognition, where large self-supervised models incur prohibitive computational costs, and existing knowledge distillation approaches suffer from unreliable teacher predictions and neglect of inter-sample relational structures. To overcome these limitations, the authors propose an adaptive multi-teacher relational distillation framework. This framework incorporates a one-class SVM–based mechanism to assess teacher reliability and dynamically weight their predictions, while also introducing a relational distillation loss that preserves cross-sample semantic structure by aligning similarity matrices between teachers and students. Experiments on IEMOCAP and CREMA-D demonstrate that four lightweight student models consistently outperform single-teacher baselines, and ablation studies confirm the complementary benefits of the two core components.
0 citationsRead paper
Comment and Control: Hijacking Agentic Workflows via Context-Grounded Evolution
This work demonstrates that large language model (LLM) agents integrated into automation platforms are vulnerable to manipulation via malicious inputs—such as GitHub comments—leading to risks including credential leakage and arbitrary command execution. To systematically uncover these novel attack surfaces in agent workflows, the authors propose “Context-Anchored Evolution,” a method that combines static path feasibility analysis, dynamic prompt provenance tracing, and runtime capability assessment within a three-stage contextual framework, enhanced by evolutionary input generation to achieve targeted hijacking of LLM agents. The resulting JAW framework successfully exploits 4,714 GitHub workflows and 8 n8n templates, compromising 15 widely used GitHub Actions and 2 official n8n nodes. Vulnerabilities identified through this approach have been acknowledged and patched by vendors including GitHub, Google, and Anthropic.
0 citationsRead paper
IIR-VLM: In-Context Instance-level Recognition for Large Vision-Language Models
This work addresses the challenge that large vision-language models struggle to distinguish individual instances—such as specific people or objects—thereby limiting their applicability in personalized scenarios. To overcome this limitation, the authors propose an auxiliary visual encoding mechanism that leverages a pre-trained instance-level recognition expert model to provide specialized features to the large vision-language model. This enables the model to perform few-shot, context-aware learning and achieve fine-grained understanding of novel instances from a single example, without requiring extensive instance-specific data or additional training. The approach is the first to enable in-context one-shot instance-level recognition in large vision-language models and supports cross-category instance perception. Evaluated on both existing and newly curated multi-category benchmarks—including faces, persons, pets, and general objects—the method significantly outperforms current state-of-the-art approaches.
0 citationsRead paper