Curvature-Independent Regret Bounds for Distributed Online Optimization on Hadamard Manifolds

A new arXiv paper studies decentralized online optimization where the decision variables live on Hadamard manifolds, a setting with negative curvature. The authors derive regret guarantees that do not depend on the manifold's curvature, avoiding the finite lower bound on sectional curvature required by earlier analyses built on geodesic convexity. The results apply to distributed multi-agent settings where agents coordinate over a network while optimizing online.

Online Reinforcement Learning Applied Inside Met Office Unified Model

A study from the Met Office pairs its Unified Model with a distributed reinforcement-learning setup so that agents can be trained while the numerical weather model runs. The work targets machine-learned corrections that stay stable as the underlying forecast model evolves, rather than being trained offline on frozen data. The authors describe the coupling architecture that lets model and agent processes communicate across distributed infrastructure.

Hugging Face details async GRPO with LoRA on Jobs using no NCCL

A Hugging Face blog post describes a method for running asynchronous GRPO reinforcement-learning training with LoRA adapters across the company's Jobs infrastructure. Instead of relying on NCCL collectives for inter-worker communication, the setup uses an object-storage bucket plus a proxy to pass data between the policy and training components. The approach is presented as a practical way to scale online RL fine-tuning without tightly coupled GPU networking.

WHY IT MATTERS ↘By replacing NCCL with object storage and a proxy, Hugging Face’s setup lowers the networking bar for online RL fine-tuning, letting teams use cheaper, loosely coupled or preemptible GPUs instead of high-bandwidth clusters. That could reduce costs and widen who can train reasoning models, while shifting operational trade-offs toward storage latency, checkpoint security, and reproducibility controls.