This paper proposes a learnable Dirichlet-process cache that stores only novel inputs to bridge the gap between state-space models and attention mechanisms.
ActiveFly-Bench is a new benchmark designed to evaluate aerial embodied perception by linking high-level task understanding, behavior planning, and low-level control for unmanned aerial vehicles.
This study analyzes the internal attention dynamics of vision-language models to explain why visual grounding degrades and proposes scheduling visual relay windows to stabilize reasoning.
The authors introduce SWE-MERA, a dynamic benchmark designed to mitigate data contamination and improve the evaluation of LLMs on software engineering tasks.
Researchers released FindMyText, an open-source Python package designed to detect near-verbatim text containment within large web-crawled datasets using document fingerprinting.
Researchers have introduced a new benchmark dataset of mathematical constant formulas to evaluate the advanced mathematical reasoning capabilities of AI systems.
This paper examines the phenomenon of model collapse from both engineering and creative perspectives, exploring how recursive training on AI-generated data affects model output.
Embodied-R1.5 is a new foundation model designed to unify embodied reasoning and physical intelligence using a large-scale dataset of 15 billion tokens.
This paper provides a comprehensive survey and a new two-level taxonomy of Graph Neural Network methodologies applied across knowledge graph technologies.
Researchers propose a new speculative decoding method that improves large language model inference speed by utilizing progressive tree drafting to better exploit parallel processing.
GrandCode is a new multi-agent reinforcement learning system that aims to achieve competitive programming performance at the level of human grandmasters.
A new metric called the 50%-task-completion time horizon has been proposed to better compare AI performance against human capabilities in software development tasks.
Agentic-DPO is a new training framework designed to optimize AI agent policies on expert trajectories by teaching them to avoid plausible mistakes rather than just imitating sequences.
IdeaTrail is a new dataset designed to capture the complete, multi-stage workflow of AI agents performing scientific ideation and research tasks.
The Mako project introduces a self-evolving agentic operating system capable of autonomously synthesizing and testing new security exploits.
A new scheduling method for heterogeneous edge GPUs aims to improve the efficiency and throughput of vision transformer models in autonomous vehicle applications.
Researchers developed ReflectWorld-MM, a multimedia memory system that organizes long-term video stream data around persistent entities rather than individual frames.
The study identifies a critical inefficiency in Chain-of-Thought prompting where models generate redundant but logically valid reasoning steps that current evaluators fail to penalize.
Researchers propose a formal theory of least autonomy as a security principle to constrain the permissions and workflow capabilities of agentic AI systems.
This paper outlines a deployment-focused pathway for transitioning medical AI agents from simple assistants to autonomous clinical systems.
MemDecay is a new memory management technique that optimizes LLM agent inference by using region-aware KV cache eviction based on semantic structure.
The proposed Gefen optimizer reduces the memory footprint of large-scale model training by sharing second-moment estimates and quantizing first-moment states.
The authors introduce FARS, a fully automated system that enables AI agents to conduct research, generate hypotheses, and write manuscripts at scale.
Researchers have introduced SDABench, a new benchmark designed to evaluate the scientific data analysis capabilities of large language models across six distinct areas.