# TRIAGE method stabilizes 4‑bit RL training for Qwen models

> Researchers introduce TRIAGE, a direction‑aware fix that keeps 4‑bit execution while matching full‑precision performance on math benchmarks.

Oossa · 2026-10-07 · https://oossa.com/en/triage-method-stabilizes-4-bit-rl-training-for-qwen-models

Zhen Li and team released a new technique called TRIAGE on Oct 7 2026. It works inside native NVFP4, a 4‑bit weight‑and‑activation format, to balance policy‑gradient updates during reinforcement learning. The method diagnoses mismatches at the segment level and applies bounded repairs, keeping the low‑precision execution unchanged. Tests on Qwen3‑4B and Qwen3‑30B‑A3B showed stable training and full‑precision quality on five math reasoning benchmarks, while delivering up to 2.3 times faster rollout speed than BF16.

## The facts

- Event date: 2026-10-07 ("Wed Oct 07 2026")
- TRIAGE gave up to 2.3× higher rollout throughput than BF16

## Why it matters

It lets developers train large language models faster with 4‑bit hardware without losing accuracy, saving compute costs.

## Sources & references

1. [TRIAGE: Direction-Aware Mismatch Stabilization of Native NVFP4 Reinforcement Learning](https://arxiv.org/abs/2610.07043) – arXiv cs.LG, 2026-10-07

Last updated: 2026-10-07