This paper presents a multi-agent AI architecture that achieves 92.1% reliability by organizing specialized AI agents into teams with opposing roles and hierarchical oversight, similar to corporate organizational …
GEPA introduces a novel prompt optimization approach that uses natural language reflection and Pareto-based evolutionary search to optimize compound AI systems, achieving superior performance compared to reinforcement …
A randomized controlled study examining how AI assistance affects skill formation in software development, finding that while AI can improve productivity, it significantly impairs conceptual understanding, code reading, …
This paper presents a comprehensive framework for authenticated delegation to AI agents, extending OAuth 2.0 and OpenID Connect protocols to enable secure, accountable, and auditable delegation of authority from users to …
FuzzingBrain V2 is a multi-agent LLM system that combines semantic code analysis with fuzzing-based verification to automatically discover and reproduce software vulnerabilities. The system achieved 90% detection rate on …
An in-depth exploration of world models in AI, covering their definition, implementation approaches (generative vs predictive), and practical applications from autonomous vehicles to interactive environments and agent …
A comprehensive research symposium featuring leading AI researchers discussing frontiers in artificial intelligence, from scientific applications and safety concerns to the future of AI development and its global impact.
Andrej Karpathy discusses the transition from traditional programming to AI-assisted development, exploring concepts like ‘vibe coding’ and ‘agentic engineering’ while examining how LLMs represent …
LeWorldModel introduces the first stable end-to-end Joint Embedding Predictive Architecture (JEPA) that learns world models from raw pixels using only two loss terms, achieving 48× faster planning than …
V-JEPA 2 is a self-supervised video model trained on over 1 million hours of internet video that achieves state-of-the-art performance on motion understanding and video question-answering tasks. The model can be adapted …
This paper presents Janus-Q, a novel framework that uses hierarchical-gated reward modeling to train large language models for event-driven financial trading, achieving superior performance by directly mapping financial …
This comprehensive guide examines the complete lifecycle of code large language models, from pre-training and supervised fine-tuning to reinforcement learning and deployment as autonomous agents. The paper provides …
AutoTTS introduces an environment-driven framework for automatically discovering test-time scaling strategies for LLMs, shifting from manual heuristic design to automated controller synthesis through offline replay …
