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

2026-07-22

AI Summary

This work addresses the high framework complexity and substantial iteration costs in reinforcement learning agent research, often caused by frequent algorithmic modifications. To this end, the authors propose a lightweight, PyTorch-native training framework that treats agents as ordinary programs, enabling an end-to-end, human-readable, and modifiable training loop with consistent semantics and aligned versions. The framework supports asynchronous training of multimodal and mixture-of-experts policies using only self-generated data. Under a compatible asynchronous protocol, it achieves statistical performance equivalence to state-of-the-art Megatron-based systems while significantly reducing development complexity, thereby offering both simplicity and high performance.

Abstract

Agentic reinforcement learning research is constant algorithm modification, new estimators, new pipeline stages, new rollout schemes, and in mainstream frameworks each change threads through layers of trainer, distributed backend, and rollout glue: the cost lands on the researcher at every iteration. Molt is a PyTorch-native training framework built to keep that cost small: a codebase compact and clean enough for a researcher to hold in their head, and for an AI coding assistant to read and reason about in its entirety, so the algorithm flow can be traced and changed end to end. The agent is an ordinary program, and one asynchronous loop trains multimodal and mixture-of-experts policies while never training on a token it did not generate, consistent in tokens, policy versions, and model semantics. Leanness does not cost performance: under a matched, fully asynchronous protocol, Molt is statistically comparable to a state-of-the-art Megatron-based stack. Molt is open source and provides recipes and containers at https://github.com/NVIDIA-NeMo/labs-molt.

Problem

• agentic reinforcement learning
• training framework
• algorithm iteration
• research overhead
• PyTorch-native

Innovation

• agentic reinforcement learning
• PyTorch-native framework
• asynchronous training
• mixture-of-experts
• token-consistent policy