Researchers introduced the Legal Multi-Agent Debate framework to evaluate how different multi-agent structures and aggregation methods perform in legal textual entailment tasks.
Researchers have developed a framework called BREW that enables AI agents to store and retrieve past interaction experiences to improve performance on multi-step tasks.
This paper proposes a neuro-agentic control framework that combines deep learning with LLM-powered agents to manage security controls in industrial IoT environments safely.
The authors introduce Latent Reward Steering, an inference-time framework designed to improve reasoning in large language models by dynamically adjusting cognitive behaviors.
Researchers have introduced M4V, a multimodal Mamba-based architecture designed to improve the computational efficiency of text-to-video generation compared to traditional Transformer models.
The DaDaDa dataset is introduced to help researchers study and develop effective pricing mechanisms for data products in digital marketplaces.
A study investigates whether large language models exhibit geopolitical bias by analyzing how they evaluate policies based on the country associated with them.
Transition Matching Distillation is introduced as a method to compress complex video diffusion models into efficient, few-step generators suitable for real-time applications.
A new multi-agent framework addresses context explosion and consistency issues in automated fictional worldbuilding by using hierarchical context compression and iterative review.
The paper proposes a provenance-based defense mechanism using a dual-LLM architecture to prevent prompt injection attacks in agentic systems.
Researchers developed a method for multi-attribute steering of language models that allows for simultaneous, non-conflicting adjustments to model behavior during inference.
The study explores the integration of evolutionary computation into autonomous systems to enhance scientific discovery and exploration.
The authors introduce WILDTRACE, a new benchmark designed to evaluate how effectively AI models can integrate dispersed evidence across long-context documents.
A new test-time training framework allows robot foundation models to adapt to new tasks by observing human demonstration videos.
ALICE is a new unified foundation model for pathology that integrates knowledge from multiple vision and vision-language experts through a multi-stage distillation process.
LieBN introduces a batch normalization technique designed specifically for manifold-valued data operating over Lie groups to improve deep neural network performance.
Microsoft Research has released Aurora 1.5, an updated open foundation model for weather and climate forecasting that introduces hourly resolution and probabilistic ensemble forecasting.
Researchers introduce a mechanism for large language models to self-modify and consolidate memories, addressing current limitations in continuous learning.
Researchers have proposed a contrastive approach to weak-to-strong generalization to mitigate noise and bias when training powerful models using outputs from smaller, weaker models.
VEXAIoT introduces an AI agent-based approach for automating the identification and exploitation of security vulnerabilities in IoT devices.
The paper proposes a generative communications paradigm for 6G networks that integrates large AI models to improve semantic understanding and content generation.
The authors introduce LP2Graph, a method designed to mine and structure linear programming formulations from academic literature to assist in railway rescheduling.
OpenAI is sunsetting its task-oriented browser, ChatGPT Atlas, to streamline its productivity features and better compete with Anthropic.
Ant Group's Lingbo has launched LingBot-VA 2.0, a native world action model designed to improve robot execution speed and generalization through physical world interaction.