Researchers proposed a dual-graph verification framework to help organizations transition legacy IT security documents into machine-readable compliance formats.
Researchers are analyzing the effectiveness of output perturbation defenses against distillation attacks on black-box language models.
A new training-free method for diffusion and flow matching models accelerates generation by utilizing x-prediction to reduce the number of required neural function evaluations.
The paper proposes a spatiotemporal tokenization method to enable cross-subject and multi-session modeling of widefield calcium imaging data in neuroscience.
The Filtered Reasoning Score is a new evaluation metric designed to assess the quality of reasoning in large language models by focusing on their most confident outputs.
EHR-MPC is a new framework that uses generative electronic health record models as patient digital twins to optimize sepsis treatment strategies dynamically during inference.
A new tool-adaptive reranker has been proposed to improve the accuracy and efficiency of information retrieval by selectively invoking external tools.
A new benchmark called BackendForge is introduced to evaluate the ability of agentic LLMs to generate and deploy functional code within backend service environments.
Researchers investigated how message formatting affects information fidelity and generation quality when LLM agents pass information across multiple hops.
Tool-MCoT introduces a small language model optimized for content safety moderation by leveraging external tools to reduce computational costs and latency.
Researchers have developed AMID, an autonomous multi-agent framework designed to streamline the development and validation of medical imaging models.
Researchers propose a mathematical framework to replace similarity heuristics in semantic caching with governed conversational demand classes.
Anthropic's latest advertisement has sparked strong emotional reactions and discomfort among viewers.
Researchers have proposed an autonomous framework designed to extract structured scientific knowledge from unstructured literature to accelerate AI-driven materials research.
This paper introduces TS-Mask VLA, a vision-language-action model that utilizes two-dimensional temporal-spatial masking to improve action generation for embodied agents.
AMT-X is a new multi-turn red-teaming framework designed to improve LLM safety evaluations by addressing the limitations of single-turn attack datasets and scoring methods.
The study demonstrates that structured, object-centric slot representations can improve robotic manipulation policies without requiring increased model capacity.
The paper explores how the reliability of value estimation affects policy optimization in offline-to-online reinforcement learning for robotic manipulation.
Researchers compared Graph Attention Networks and BERT models to evaluate their effectiveness in mimicking human player behavior and strategies in puzzle games.
Researchers introduced Engram, a module that adds conditional memory to Transformers to enable efficient knowledge lookup as an alternative to standard Mixture-of-Experts scaling.
Researchers have proposed a security decision support system that uses a multi-agent framework to recommend security controls based on minimal user requirements.
The paper presents OpenProver, an open-source, agentic system that integrates a Planner-Worker-Verifier architecture for automated theorem proving using Lean 4.
Researchers are utilizing imitation learning to develop autonomous navigation systems for soft robotic guidewires used in endovascular surgery.
Researchers have developed PREF-Gate, a decision framework designed to prevent invalid neighborhood risk data from compromising graph-based fraud detection.