Amazon Science
·
1 month ago
Amazon's Project Eluna and related research propose four approaches to ground AI agents in physical environments: physics-guided deep learning, uncertainty-aware reasoning, bridging text-to-numerical gaps, and verifier-augmented grounding. The uncertainty-aware reasoning framework achieved over 25% reduction in expected calibration error, while the adapting-while-learning framework achieved 29% higher answer accuracy on physical-science datasets. These techniques enable AI agents to reason reliably in high-stakes physical settings by respecting physical laws and constraints rather than producing dangerous hallucinations.
Amazon Science
·
1 month ago
Researchers published a paper formalizing the intent-execution gap in AI agents, where mismatches occur between what language models intend and what the harness software actually executes, demonstrating that closing this gap without task-specific tuning achieves state-of-the-art results on benchmarks including SWE-Pro and Terminal-Bench2. The study introduces Simple Strands Agent, a lightweight harness implementation, and identifies concrete design principles such as requiring stronger text anchors for code edits, providing diff feedback after execution, and balancing reasoning with tool interactions through model-specific nudging. These findings suggest that optimal agent performance requires tight model-harness codesign rather than optimizing either component independently, with effective strategies varying across model families.
Hugging Face Blog
·
1 month ago
OpenEnv, a tool for creating execution environments where AI agents can interact with terminals and browsers, is now governed by a committee including Meta, Microsoft, Nvidia, and Hugging Face to standardize how these environments are built and deployed. The project is moving from a reward-definition framework to a protocol layer using familiar APIs like Gymnasium's reset() and step() functions, with environments served over HTTP and WebSocket with Docker packaging. This shift enables any open-source model to work with any environment without custom code, allowing the community to train specialized agents efficiently across different infrastructure platforms.
