Definition
Plain language
An open-source training framework for reinforcement learning on language models.
As stated in the literature
A distributed RL training framework for LLMs commonly used in agent post-training pipelines, notable for FSDP-based sharding rather than DeepSpeed.
Why it matters: It is one of the few open frameworks built for the specific demands of large-scale LLM RL, which differs significantly from standard supervised training.
For example, an open-source team training an agent with GRPO across multiple nodes might use verl to coordinate rollouts, scoring, and weight updates.
Heard on the show
“The mixed-policy methods — LUFFY, ReLIFT, all of them — were implemented in a different framework called verl.”Episode 009 — How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers