AI
What’s new As AI "teams" of LLM agents grow, the cloud bill can explode. BAMAS is a framework that designs these teams with a dollar cap in mind. How it works * Picks the right mix of models: Uses integer linear programming to balance accuracy vs. price. * Plans
AI
Meet MAESTRO: an AI coach for teamwork Training teams of AI agents is hard: you must design the right "score" (rewards) and the right practice plan (curriculum). Get either wrong and learning stalls. MAESTRO uses a large language model (LLM) as an offline training architect—so it guides
AIResearch
Long, well-sourced answers are hard for AI to learn: most training rewards short, easily graded Q&A. This work introduces Reinforcement Learning with Evolving Rubrics (RLER) - grading guides that co-evolve with the model, so feedback stays aligned with what the model actually explores during multi-step research. Built with
AI
Teach Robots to Learn Like People Building versatile embodied AI (agents that see, read, and act) is hard because real-world data is scarce and training is expensive. We introduce Deliberate Practice Policy Optimization (DPPO), a metacognitive "metaloop" that alternates between: * Supervised fine-tuning to expand general competence * Reinforcement learning
AI
Building capable embodied AI is hard: real-world data is scarce and training is expensive. This paper introduces Deliberate Practice Policy Optimization (DPPO), a coach-like training loop that helps vision-language agents learn more from less. How it works: the system alternates between learning from examples (to expand skills) and trial-and-error reinforcement
ReinforcementLearning
Reinforcement learning (RL) wins in games and simulators—but deploying it on real products is a different story. Gabriel Dulac‑Arnold, Daniel Mankowitz, and Todd Hester outline nine must-solve challenges before RL can safely power real-world systems. * Safety & constraints: avoid harmful actions while learning. * Sample efficiency: learn from limited,
Robotics
What's new PhysWorld lets robots learn skills from AI-generated videos—without collecting any real robot data. How it works * Given a task and one photo of the scene, a video model "imagines" a demonstration. * PhysWorld reconstructs a physics-based world from that video. * An object-centric learner turns
AI
AI finally masters Stratego—on a budget Stratego has long been a worst-case test for AI: you must plan deeply while most pieces are hidden. Big-budget attempts still fell short of top humans. This paper reports a step change: an AI that reaches vastly superhuman Stratego performance—trained for only
AI
Big win for AI in hidden-information games. Stratego has long resisted top-tier AI—even with training runs costing millions. This new study cracks it. Researchers achieved vastly superhuman Stratego play using self-play reinforcement learning plus a smart test-time search designed for imperfect information. The surprise: it worked with only a
Robotics
What if robots could learn new tasks just by watching AI-generated videos - and still obey physics? PhysWorld is a new framework that turns text-and-image prompts into physically executable robot skills. * It first generates a task-conditioned video from a single scene image and a command. * Then it reconstructs the scene&
AI
AI that sees, thinks—and uses tools Meet DeepEyesV2, a multimodal “agentic” model that doesn’t just read text and look at images—it can call external tools like code runners and web search, then weave the results into its reasoning. Key ideas: * Two-stage training: a cold-start phase teaches basic
AI
Struggling to compare or improve your automations? Opus offers a simple, quantitative way to score and optimize any Workflow—so you can pick what works, fix what doesn’t, and waste fewer resources. What’s inside * Opus Workflow Reward: a probabilistic score that blends likelihood of success, resource usage, and