Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning

Researchers introduced Molt, an open-source PyTorch-native framework for agentic reinforcement learning.

It supports asynchronous training for multimodal and mixture-of-experts policies while staying compact and end-to-end readable.

The paper reports comparable performance to a Megatron-based system under matched conditions.

Molt aims to reduce engineering overhead and make rapid RL algorithm iteration easier.