Jet-RL: Enabling On-Policy FP8 Reinforcement Learning with Unified Training and Rollout Precision Flow
Haocheng Xi 1,3 Charlie Ruan 3 Peiyuan Liao 4 Yujun Lin 1 Han Cai 1 Yilong Zhao 3 Shuo Yang 3 Kurt Keutzer 3 Song Han 1,2 Ligeng Zhu 1, † {\dagger} 1 NVIDIA 2 MIT 3 UC Berkeley 4 Independent Researcher · 2026
Read Jet-RL: Enabling On-Policy FP8 Reinforcement Learning with Unified Training and Rollout Precision Flow by Haocheng Xi 1,3 Charlie Ruan 3 Peiyuan Liao 4 Yujun Lin 1 Han Cai 1 Yilong Zhao 3 Shuo Yang 3 Kurt Keutzer 3 Song Han 1,2 Ligeng Zhu 1, † {\dagger} 1 NVIDIA 2 MIT 3 UC Berkeley 4 Independent Researcher (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-14243.txt.
Description
Haocheng Xi and a team from NVIDIA, MIT, UC Berkeley, and other labs present Jet-RL in this 2026 paper. Their topic is training large language models with reinforcement learning, the trial-and-error method that builds reasoning skills. Such training is slow and costly. The rollout phase, where the model drafts many trial replies, takes over 70 percent of the total time.
Jet-RL tries to cut that cost with FP8, a compact number format that uses eight bits per value. The team built one shared flow for training and rollout, so both phases use the same low precision. The goal is on-policy learning that runs faster without a new split between the two phases.
Catalog record
| CREATOR | Haocheng Xi 1,3 Charlie Ruan 3 Peiyuan Liao 4 Yujun Lin 1 Han Cai 1 Yilong Zhao 3 Shuo Yang 3 Kurt Keutzer 3 Song Han 1,2 Ligeng Zhu 1, † {\dagger} 1 NVIDIA 2 MIT 3 UC Berkeley 4 Independent Researcher |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-14243