The application of vision-language models in medicine is critically hampered by the scarcity of high-quality, expert-annotated data. Supervised fine-tuning on existing datasets often leads to poor generalization on unseen modalities and tasks, while reinforcement learning, a promising alternative, is stymied by the lack of reliable reward signals in this data-scarce domain. To address this challenge, we propose a Generative Reward Learning framework that establishes a self-improving training cycle. The framework jointly develops a data generator and a reward model, enabling the automated and continuous creation of high-quality multimodal medical data that serves as an effective training source for post-training. Our experiments demonstrate that supervised fine-tuning using the generated data already surpasses models trained on large-scale human-curated datasets. More importantly, when the generated data is further leveraged for reinforcement learning via Group Relative Policy Optimization, the resulting model achieves state-of-the-art cross-modality and cross-task generalization, significantly outperforming specialized reinforcement-learning-based methods. Notably, a compact model trained under this framework attains performance competitive with foundation models containing more than an order of magnitude more parameters. These results suggest a new paradigm for data-efficient learning in high-stakes medical domains, shifting the bottleneck from data scarcity to data generation and unlocking the potential of reinforcement learning for building robust and generalizable medical AI systems.

Xianda Zheng, Huan Gao, Meng-Fen Chiang, Michael J. Witbrock, Kaiqi Zhao, Shangyang Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Pre-trained large-scale models have brought about significant advancements in Natural Language Processing (NLP), inspiring their application to other domains, including physiological signals. However, the use of pre-trained models in the analysis of physiological signals, particularly electroencephalograms (EEG), remains limited. Specifically, the application of self-supervised pre-training methods to EEG signals faces three key challenges: (1) Pre-training self-supervised models demands substantial computational resources; (2) The low signal-to-noise ratio (SNR) of EEG signals can hinder the representation learning capabilities of these models; (3) There is a lack of high-quality training datasets. High-quality EEG data is not as readily available as natural language data, as EEG signals are typically collected from limited datasets and predominantly contain normal signals. To address these challenges, we propose PLM2EEG, a method that leverages pre-trained language models for EEG analysis tasks. The two core components of PLM2EEG are as follows: (1) EEG Tokenization, where fixed-length EEG segments from each channel are treated as tokens, encapsulating local feature information and aligning with the input dimensions of pre-trained language models. Channel and positional embeddings are added to each token to preserve the spatiotemporal integrity of the signals. (2) Frozen Pretrained Model, which uses a pre-trained large language model with its self-attention and feedforward layers retained. The model is then fine-tuned specifically for EEG-related tasks, showcasing its adaptability to this domain. Experimental results demonstrate that PLM2EEG significantly outperforms existing self-supervised pre-trained models on EEG tasks across two large datasets. It enhances cross-dataset learning and sets new benchmarks for EEG analysis.

Federated Learning (FL) is a groundbreaking approach that enables multiple clients to jointly train deep learning models by pooling their data, while addressing privacy and bandwidth issues that prevent direct data sharing. This approach is particularly suitable for building strong and widely applicable graph models, given the increasing amounts of graph data stored across different locations. However, FL for subgraph models faces significant challenges, such as the diversity of data and the risk of attacks, which can affect the strength and reliability of these models. In response to these challenges, our research delves into the complexities of FL for subgraphs from an information theory perspective. We identify a major issue that affects the performance of graph models: the bias in the optimization goal of the commonly used FedAVG training method. To address this, we propose InfoFedGNN, an innovative FL framework for subgraphs that is based on the Information Bottleneck principle. InfoFedGNN is designed to overcome the problem of Non-Independent and Identically Distributed (non-i.i.d.) data in FL and to significantly improve its defense against security threats. Our thorough evaluation of InfoFedGNN on five public datasets, with both uniform and diverse data distributions, highlights its improved defense capabilities and better training outcomes. These results confirm the effectiveness of InfoFedGNN in enhancing the security and efficiency of FL, demonstrating its potential to push forward the development of federated graph models.

Brain simulation builds dynamical models to mimic the structure and functions of the brain, while brain-inspired computing (BIC) develops intelligent systems by learning from the structure and functions of the brain. The two fields are intertwined and should share a common programming framework to facilitate each other's development. However, none of the existing software in the fields can achieve this goal, because traditional brain simulators lack differentiability for training, while existing deep learning (DL) frameworks fail to capture the biophysical realism and complexity of brain dynamics. In this paper, we introduce BrainPy, a differentiable brain simulator developed using JAX and XLA, with the aim of bridging the gap between brain simulation and BIC. BrainPy expands upon the functionalities of JAX, a powerful AI framework, by introducing complete capabilities for flexible, efficient, and scalable brain simulation. It offers a range of sparse and event-driven operators for efficient and scalable brain simulation, an abstraction for managing the intricacies of synaptic computations, a modular and flexible interface for constructing multi-scale brain models, and an object-oriented just-in-time compilation approach to handle the memory-intensive nature of brain dynamics. We showcase the efficiency and scalability of BrainPy on benchmark tasks, and highlight its differentiable simulation for biologically plausible spiking models.

Access control vulnerabilities that lead to elevated privileges are among the most dangerous vulnerabilities in Web applications. Most of the existing detection methods use dynamic or static analysis techniques alone, which suffer from high manual involvement, low automation, high leakage rate, low page coverage, and other deficiencies. To this end, this paper proposes a novel access control vulnerability detection method (DetAC) based on a sitemap model with global information representation. This method first constructs a static site-wide sitemap model based on the page link addresses in the Web application source code through static analysis techniques. After that, the application is logged in and executed dynamically with different role users. During this process, execution traces and request parameters are collected and converted into annotations to fill the corresponding edges of the static site-wide sitemap model. Then, the sitemap model with global information representation is obtained. This model can represent both the global control flow and data flow of the application. Then DetAC analyzes the role-based and user-based access control policies of the Web application based on the node reachability and annotated data features of the model. And according to the information such as role, user, and access resources, it generates attack vectors to achieve different roles and the same role of different users to access each other’s resources. Finally, access control vulnerabilities are detected based on the equivalence of the results obtained using attack vector access and normal access to the Web application server. DetAC was validated on five real open-source Web applications, and the results showed that DetAC successfully detected up to 12 access control vulnerabilities, which are more than those of the traditional seven tools. The dynamic analysis page coverage rate was significantly improved during the detection process, reaching an average of 91.37%.