REGEN paper proposes replay-recycling for expert-to-generalist LLM distillation via offline RL

A revised arXiv paper introduces REGEN, a method that recycles replay data to distill specialized expert policies into a more general model using offline reinforcement learning. The approach targets the cost of scaling online RL, which is widely used to develop long-horizon reasoning and tool-use abilities in large language models. The v3 revision appears in both the cs.AI and cs.LG listings.

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.