Tingchao Fu, Wenkai Wang, Fanxiao Li, Huadong Zhang, Jinhong Zhang, Dayang Li, Yunyun Dong, Renyang Liu, Wei Zhou. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Even when factually correct, social-media news previews (image-headline pairs) can induce interpretation drift: by selectively omitting crucial context, they lead readers to form judgments that diverge from what the full article supports.This covert harm is subtler than explicit misinformation, yet remains underexplored.To address this gap, we develop a multistage pipeline that simulates preview-based and context-based understanding, enabling construction of the MM-MISLEADING benchmark.Using MM-MISLEADING, we systematically evaluate open-source LVLMs and uncover pronounced blind spots in omission-based misleadingness detection.We further propose OM-GUARD, which combines (1) Interpretation-Aware Fine-Tuning for misleadingness detection and (2) Rationale-Guided Misleading Content Correction, where explicit rationales guide headline rewriting to reduce misleading impressions.Experiments show that OM-GUARD lifts an 8B model's detection accuracy to the level of a 235B LVLM while delivering markedly stronger end-to-end correction.Further analysis shows that misleadingness usually arises from local narrative shifts, such as missing background, instead of global frame changes, and identifies image-driven cases where text-only correction fails, underscoring the need for visual interventions.

As industrial equipment becomes increasingly complex and intelligent, fault diagnosis (FD) technology has emerged as a critical means to ensure system reliability and safety, spurring the development of various real-time online monitoring techniques. Traditional fault diagnosis methods primarily rely on single data sources and specific algorithms, which makes it challenging to effectively integrate the multimodal data captured by diverse sensors and often overlooks the vital role of human expertise in the diagnostic process. By leveraging a universal fault diagnosis framework that combines Large Language Models (LLMs) with multimodal data, existing methods can be seamlessly integrated. LLMs possess powerful capabilities in natural language understanding, knowledge integration, and reasoning, enabling them to analyze text, images, signals, and other types of multimodal information to facilitate zero-shot and few-shot knowledge reasoning and fault diagnosis. This paper systematically reviews the development of LLM- and multimodal data-based fault diagnosis technologies, outlines key techniques such as data-driven processing, feature extraction, and feature fusion within LLM frameworks, and analyzes the technological advancements fostered by these methods. It also summarizes the advantages and limitations of this approach in fault diagnosis and health state assessment, and offers an outlook on the future trends and challenges of applying LLMs in multimodal fault diagnosis. The aim is to provide technical guidance for researchers and engineers, thereby accelerating the innovation and application of intelligent fault diagnosis technologies.

Deep Reinforcement Learning (DRL) has achieved remarkable success in various fields by leveraging neural networks. However, applying DRL to control complex robot systems faces challenges, such as hard to converge and robust control to accommodate different environments. In this paper, we propose IDM-TD3, a new DRL framework which introduces an Inverse Dynamic Model into TD3 as the output map of the actor network to decouple the kinematics and dynamics of the control system. This decoupled configuration not only permits online fine-tuning of the IDM within target robot environment for a better control performence, but also facilitates the seamless transference of experiential knowledge across agents with akin kinematic features. Experimental results show that without pretraining, the convergence performance of the proposed method is comparable to that of our baseline algorithm TD3. If the kinematic network is pretrained using expert policies (even from environments with different dynamic parameters), we achieve much better convergence than TD3 and its combination with behavioral cloning. Moreover, by fine-tuning the IDM, our method exhibits robust control even in environments with distinct dynamic differences. This shows the promising application of our IDM-TD3 in many fields, particularly in addressing the generalization problem or harnessing pre-existing experiences for the training of nascent agents.

The growing frequency of medical emergencies, such as cardiac arrest and stroke, highlights the need for timely first aid interventions. Traditional emergency medical services (EMS) face challenges like delayed response times, resource disparities, and high-risk environments, often leading to inadequate care. Recent advancements in robotics, artificial intelligence, and cloud computing have enabled the development of first aid robots, which can provide automated, efficient medical assistance in critical situations. This paper explores the motivation for first aid robots, reviews the state of the art, and identifies key challenges. Finally, we propose a technical framework for potential implementation. Despite having promising potential to enhance emergency medical services and preserve lives, further research and multidisciplinary collaboration are needed for their effective deployment.

Dayang Li, Fanxiao Li, Bingbing Song, Li Tang, Wei Zhou. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

When using current head-mounted displays (HMDs), users with optical aberrations need to wear the equipment on the top of their own glasses. As both the HMDs and the glasses require to be tightly attached to faces, wearing them together is very inconvenient and uncomfortable, and thus degrades user experiences heavily. In this paper, we propose a real-time image pre-correction technique to correct the aberrations purely by software. Users can take off their own glasses and enjoy the virtual reality (VR) experience through an ordinary HMD freely and comfortably. Furthermore, as our technique is not related to hardware, it is compatible with all the current commercial HMDs. Our technique is based on the observation that the refractive errors majorly cause the ideal retinal image to be convolved by certain kernels. So we pre-correct the image on the display according to the specific aberrations of a user, aiming to maximize the similarity between the convolved retinal image and the ideal image. To achieve real-time performance, we modify the energy function to have linear solutions and implement the optimization fully on GPU. The experiments and the user study indicate that without any changes on hardware, we generate better viewing experience of HMDs for users with optical aberrations.