Large language models (LLMs) have demonstrated immense utility across various industries.However, as LLMs advance, the risk of harmful outputs increases due to incorrect or malicious prompts.While current methods effectively address jailbreak risks, they share common limitations: 1) Judging harmful outputs from the prefill-level lacks utilization of the model's decoding outputs, leading to relatively lower effectiveness and robustness.2) Rejecting potentially harmful outputs based on a single evaluation can significantly impair the model's helpfulness.To address the above issues, we examine LLMs' capability to recognize harmful outputs, revealing and quantifying their proficiency in assessing the danger of previous tokens.Motivated by pilot experiment results, we design a robust defense mechanism at the decoding level.Our novel decoder-oriented, step-by-step defense architecture corrects the outputs of harmful queries directly rather than rejecting them outright.We introduce speculative decoding to enhance usability and facilitate deployment to boost safe decoding speed.Extensive experiments demonstrate that our method improves model security without compromising reasoning speed.Notably, our method leverages the model's ability to discern hazardous information, maintaining its helpfulness compared to existing methods 1 .

Existing multimodal entity and relation extraction tasks primarily focus on text-to-text or text-to-visual entity relations, overlooking real-world complexities involving visual-to-text and visual-to-visual cases, thus failing to capture the richer semantic structures in complex cross-modal interactions. To address the limitations, we propose a new task, Unconstrained Multimodal Entity and Relation Extraction (U-MERE), which jointly extracts arbitrary visual and textual entities, and their relations from image-text pairs. To accomplish U-MERE, we construct UMERE-Bench, a benchmark with over 9,000 samples that comprehensively covers four cross-modal entity relation directions and three task settings. Given the difficulty of jointly modeling diverse directions of cross-modal entity relations, we introduce Collaborative Modeling and Order-Sensitive (CMOS), which collaboratively guides large vision-language models (LVLMs) to decompose task complexity and mitigates generation order bias from fixed target relation sequences. CMOS employs small models to generate candidate entities, guiding LVLMs to capture key information and jointly optimizes multiple feasible relation orderings to reduce order dependency. Additionally, we design a Multimodal Order-aware Matching (MOM) evaluation method to align predictions with ground truth for precise assessment. Experimental results reveal that current LVLMs show limited performance on U-MERE, underscoring its inherent challenges, while CMOS consistently achieves superior performance across multiple advanced LVLMs, demonstrating its effectiveness and generalization capability. The dataset and code will be available in https://github.com/jiaweidoris/U-MERE.

Hallucination in large vision-language models (LVLMs) is a significant challenge, i.e., generating objects that are not present in the visual input, which significantly compromises the reliability of models. Recent studies often attribute hallucinations to a lack of visual understanding, yet ignore a more fundamental issue: the model's inability to effectively extract or decouple visual features. In this paper, we revisit the hallucinations in LVLMs from an architectural perspective, investigating whether the primary cause lies in the visual encoder (feature extraction) or the modal alignment module (feature decoupling). Motivated by our preliminary findings, we propose a parameter-efficient fine-tuning strategy, PATCH, to mitigate hallucinations in LVLMs. This plug-and-play method can be integrated into various LVLMs, leveraging adaptive virtual tokens to extract object features from bounding boxes, thereby addressing hallucinations stemming from inadequate feature decoupling. PATCH achieves state-of-the-art performance across multiple multi-modal hallucination datasets and demonstrates significant improvements in general capabilities. We hope this work provides deeper insights into the underlying causes of hallucinations in LVLMs, fostering further advancements and innovation in this field. The code will be available at https://github.com/YuyingShang/PATCH.