MEDA: Medical-Oriented Activation Editing for Hallucination Mitigation in Medical Large Vision-Language Model

Tianbo Wang, Yuqing Ma, Lingyan Meng, Zhange Zhang, Kewei Liao, Jian Yang, Simin Li, Jinyang Guo, Xianglong Liu

Github

• This paper proposes MEDA, the first efficient activation editing paradigm tailored to medical scenarios, which steers LVLMs toward medically grounded expertise for mitigating hallucination.

AFTER: Mitigating the Object Hallucination of LVLM via Adaptive Factual-Guided Activation Editing

Tianbo Wang, Yuqing Ma, Kewei Liao, Zhange Zhang, Simin Li, Jinyang Guo, Xianglong Liu

Github

• This paper proposes AFTER, a novel adaptive factual-guided activation editing approach to mitigate object hallucinations and language bias in Large Vision-Language Models.

Query-Routed Activation Editing with Truth-hierarchical Preference Optimization

Kewei Liao*, Tianbo Wang*(Equal Contribution), Yuqing Ma, Zhange Zhang, Zhicheng Geng, Xiaowei Zhao, Jiakai Wang, Xianglong Liu

Github

• This paper proposes QRAE, a novel query-routed activation editing framework to leverage query-specific semantics for adaptive hallucination mitigation in LLMs.

CASE: Conflict-assessed Knowledge-sensitive Neuron Tuning for Lifelong Model Editing

Zhange Zhang, Yuqing Ma, Yulong Wang, Tianbo Wang, Jiakai Wang, Simin Li, Xianglong Liu

Github

• This paper proposes CASE, a novel conflict-assessed knowledge-sensitive neuron tuning approach to address subspace allocation issues and editing conflicts in lifelong model editing.

Token-Aware Editing of Internal Activations for Large Language Model Alignment

Tianbo Wang, Yuqing Ma, Kewei Liao, Chengzhao Yang, Zhange Zhang, Jiakai Wang, Xianglong Liu

Github

• This paper proposes TAE, a novel token-aware activation editing approach to utilize token-level alignment information for superior inference-time behavior mitigation in LLMs.

Towards Universal X-ray Security Inspection: A Benchmark and Stereoscopic-Aware Oriented Prohibited Item Detection Framework

Tianbo Wang, Kewei Liao, Zhange Zhang, Yuqing Ma, Hongping Zhi, Aishan Liu, Ruihao Gong, Xianglong Liu

Github

• This paper proposes SWEAR, a novel stereoscopic-aware oriented detection framework to suppress background interference and improve prohibited item detection in oriented X-ray security inspection.

Conflict-Aware Knowledge Editing in the Wild: Semantic-Augmented Graph Representation for Unstructured Text

Zhange Zhang, Zhicheng Geng, Yuqing Ma, Tianbo Wang, Kai Lv, Xianglong Liu

[Github]

• This paper proposes CAKE, a novel conflict-aware knowledge editing framework to enable precise knowledge extraction and refinement from wild unstructured text in LLMs.

Lexical Diversity-aware Relevance Assessment for Retrieval-Augmented Generation

Zhange Zhang, Yuqing Ma, Yulong Wang, Shan He, Tianbo Wang, Siqi He, Jiakai Wang, Xianglong Liu

Github

• This paper proposes DRAG, a novel lexical diversity-aware relevance assessment method to achieve granular relevance analysis and improve document retrieval precision in RAG systems.

Few-shot X-ray Prohibited Item Detection: A Benchmark and Weak-feature Enhancement Network

Renshuai Tao, Tianbo Wang, Ziyang Wu, Cong Liu, Aishan Liu, Xianglong Liu

Github

• This paper proposes WEN, a novel weak-feature enhancement network to overcome performance degradation caused by occlusion and color fading in few-shot X-ray prohibited item detection.

Exploring Endogenous Shift for Cross-domain Detection: A Large-scale Benchmark and Perturbation Suppression Network

Renshuai Tao, Hainan Li, Tianbo Wang, Yanlu Wei, Yifu Ding, Bowei Jin, Hongping Zhi, Xianglong Liu, Aishan Liu

Github

• This paper proposes PSN, a novel perturbation suppression network to handle endogenous domain shifts caused by intrinsic imaging mechanisms in cross-domain object detection.