The FAST framework improves parallel reinforcement learning for autonomous driving by mitigating the straggler effect during environment sampling.
RealityBridge addresses the simulation-to-reality gap in autonomous driving by improving the fidelity of editable 3D Gaussian Splatting scenes.
A longitudinal study of over three million learning interactions reveals that while generative AI reduces time spent on math problems, it may negatively impact long-term knowledge retention.
WrAFT is a modular automated writing evaluation system designed to provide scoring and feedback for argumentative essays using various large language models.
The proposed Turbo Connection architecture aims to enhance Transformer reasoning capabilities by allowing information flow across different layers to overcome fixed-depth computation limits.
The INFORM interpretability method allows for the analysis and decoupling of expert interaction policies within multi-LLM orchestration systems.
BiasLab is a new multilingual evaluation framework designed to measure and mitigate systematic biases in large language models within human resources and workplace applications.
A new study identifies that multi-hop reasoning failures in transformers are caused by a lack of supervision on bridge entities and proposes a method to mitigate this issue.
PiCSAR is proposed as a training-free method to improve the accuracy of reasoning models by selecting the most reliable chain of thought without requiring ground-truth labels.
Researchers present CRINN, a reinforcement learning-based approach to optimize approximate nearest-neighbor search algorithms for improved performance in AI retrieval systems.
Researchers developed a text-centered multimodal system designed to recognize human ambivalence and hesitancy in video data.
MoP-JEPA is a proposed world model architecture that uses a mixture of predictors to better handle stochastic transitions in latent state prediction.
The proposed PsPLUG plugin aims to resolve conflicts between user-specific personalization and explicit style instructions in large language models.
A new theoretical model called AIQI has been developed as the first model-free reinforcement learning agent proven to be asymptotically optimal.
StoryTeller is a training-free method designed to improve long-form audio description by maintaining narrative context across video scenes.
A new agentic system utilizes summary-based memory to detect long-term conversational scams that occur over extended periods.
Xiaomi-Robotics-U0 introduces a world foundation model designed to improve visual consistency and geometric coherence for embodied robotic applications.
Researchers have developed a new simulation testbed designed to evaluate how AI models handle complex, long-horizon business decision-making.
Researchers have developed a relational database bridge to connect bibliographic metadata with formalized mathematical proof libraries.
This paper proposes a modular post-training framework for large language models that decouples policy exploration from reward optimization to improve efficiency and transferability.
The PHILIA agent architecture is presented to support long-term human-robot coexistence through improved memory, coordination, and intent translation.
An empirical study examines how AI-powered coding assistants influence the security of software by analyzing developer interactions with security APIs.
A self-supervised foundation model has been created to improve the efficiency and scalability of cardiac CT segmentation and clinical phenotyping.
The PRISM Edit method proposes a new approach to model editing that allows LLMs to distinguish between current facts and historical temporal contexts.