ai-trend-notifierKO
← wiki

$ cat wiki/papers/2026/2607.21653-molt-agentic-rl.md

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

paperupdated 2026-07-28created 2026-07-28

TL;DR

NVIDIA NeMo's Molt is a compact, researcher-friendly PyTorch-native RL training framework for agentic scenarios that matches Megatron-scale performance without its complexity. HF Daily #1 July 27 with 605 upvotes.

Authors & Org

NVIDIA NeMo team — GitHub: NVIDIA-NeMo/labs-molt

Method

Molt composes three components:

  1. vLLM rollout engines — efficient token generation with policy-correct sampling
  2. Single FSDP2 policy actor on NeMo AutoModel — eliminates multi-process training complexity
  3. Ray async queue — decoupled coordination between rollout and training

Key design principle: training only on tokens the policy itself generated (not reference or offline data), which matters for agentic RL stability and avoids distribution shift artifacts that affect frameworks trained on mixed data.

Results

Achieves performance comparable to Megatron-based systems on standard agentic RL benchmarks while requiring significantly less infrastructure expertise to deploy and modify.

Significance

Open Questions

  • How does it handle very long trajectories (cf. SEED's 45K-token trajectories)?
  • Integration path to Cosmos (physical AI RL training)?
  • Benchmark: tested against Ring-Zero at 1T scale?

Cite

arXiv: 2607.21653 (2026-07-27)

Referenced by

Sources