Agentic reinforcement learning (RL) enhances LLM training by enabling agents to learn decision-making processes through interactive, multi-step trajectories in simulated or real environments. The article details practical challenges and solutions for training GPT-OSS with agentic RL, including fixing on-policy integrity issues, memory inefficiencies, and training-inference mismatches.

OpenAI and Amazon have formed a strategic partnership to bring OpenAI’s Frontier platform to AWS, enhancing AI infrastructure, enabling custom models, and supporting enterprise AI agents.

Google DeepMind has launched Nano Banana 2, an AI image generation model that combines the advanced features of Nano Banana Pro with the speed of Gemini Flash, enabling fast, high-quality image generation and consistent subject handling across Google products.

CourtGuard is a model-agnostic framework that enables zero-shot policy adaptation for LLM safety by using an adversarial debate mechanism grounded in external policies. It outperforms existing methods across safety benchmarks, achieves high accuracy on out-of-domain tasks, and automates dataset curation and auditing without retraining.

Decoder-based Sense Knowledge Distillation (DSKD) introduces a framework to integrate structured lexical knowledge—like word senses—into decoder-style large language models during training, improving semantic accuracy without requiring dictionary lookups at inference. The method enhances knowledge distillation for generative models by enabling them to inherit structured semantics while maintaining efficient training and performance on diverse benchmarks.

GitHub Copilot has introduced new features to enhance AI-powered code generation, improving developer efficiency through better integration with the GitHub ecosystem and expanded capabilities in generative AI and machine learning.

OpenAI and the Pacific Northwest National Laboratory have launched DraftNEPABench, a benchmark that assesses AI coding agents’ ability to speed up federal environmental permitting processes, potentially cutting drafting time by up to 15%. The tool aims to modernize infrastructure reviews through automated NEPA documentation.

Google is partnering with the Massachusetts AI Hub to offer free AI and career training to residents through its Grow with Google program, including access to AI Professional and Career Certificates.

MobilityBench is a comprehensive benchmark that evaluates LLM-powered route-planning agents using real-world user queries from Amap, offering reproducible, end-to-end testing across diverse mobility scenarios. It assesses performance in areas like instruction understanding, planning, tool use, and efficiency, revealing that models excel at basic tasks but struggle with personalized, preference-constrained routing.

dLLM is an open-source framework that unifies and standardizes core components of diffusion language modeling, enabling reproducibility, customization, and easy deployment of both large and small DLMs. It provides transparent, accessible recipes for building small DLMs from existing models like BERT or autoregressive LMs and includes released checkpoints to accelerate research.