A new loss function, Centroid-Guided Contrastive Loss (CGCL), is proposed for structured fraudulent job posting detection. CGCL unifies classification and clustering to achieve high accuracy and meaningful structure in latent-space representations, capturing subtleties among fake posts. Experiments demonstrate state-of-the-art performance on the EMSCAD benchmark dataset.

Researchers proposed DSRec, a novel dual-interest cross-SSM model for sequential product recommendation. DSRec improves upon existing methods by capturing item polysemy and dynamic behavior across different temporal granularities. The model uses two SSM encoders: a full-sequence Mamba for long-term modeling and a time-modulated SSM for short-term intent. Experiments show DSRec outperforms other state-of-the-art methods on public benchmarks.

Researchers introduced the mean-expansion layer to accelerate action-value learning in reinforcement learning, improving performance in deep Q-networks and implicit quantile networks by sharing values across actions within a state.

This research explores how the choice of p in lp norms affects the sparsity and smoothness of adversarial attacks against deep neural networks. The study finds that the optimal p value depends on the task and that lp norms with p in [1.3, 1.5] yield the best trade-off between sparse and smooth attacks.

Researchers investigated the role of momentum parameters in Adam, a widely used optimizer for large-scale training. They found that tying the two momentum parameters (β_1 = β_2) can preserve performance and identified a mechanism behind this phenomenon. This study contributes to the understanding of Adam's behavior and may inform the development of more efficient optimizers.

Researchers propose Riemannian--Lorentz Parameter Fusion (RLPF), a method for merging pre-trained Vision Transformers and state-space models by aligning parameter groups by semantic role, enabling orders-of-magnitude savings versus retraining. Initial results show improved accuracy on CIFAR-10, Oxford-IIIT Pet, and ImageNet-1K.

A new deep architecture is proposed for route planning that jointly optimizes cost functions and route-ranking models. It outperforms state-of-the-art methods in route quality and customizability. While this work is not directly related to AI agents, it may be of interest to those working on AI-related route-planning tasks or navigation services.

This tutorial bridges the gap between operations research and management science (OR/MS) and deep learning for sequential decision making under uncertainty. It presents a framework for integrating OR/MS approaches with neural architectures, highlighting the complementary strengths of both fields.

AuthorMix proposes a lightweight, modular, and interpretable style transfer framework that allows for rapid training of specialized adaptation models for each new target using layer-wise adapter mixing via reinforcement learning. This approach improves meaning preservation and ranks first on the combined style-meaning score among all baselines, including GPT-5.1.

This research explores the application of reinforcement learning to adaptively couple full and reduced order models in hybrid simulations, particularly in transient problems where localized features propagate through the domain. The approach uses deep Q-networks to select between subdomain-local full order models and pre-trained operator inference reduced order models to balance accuracy, cost, and model-switching frequency.

Researchers proposed R-U-Net, an ECG delineation model that pairs ResNet-18 with a U-Net decoder, achieving improved performance over previous models. This advancement has implications for AI models interpreting ECGs and may inform the design of AI architectures for similar tasks.