ByteDance-backed team open-sources Axion1.0: two-stage SFT+RL agent training beats GPT-6 on HLE

机器之心 · wechat · 2026-11-17

A team from Duke, ByteDance, UMD and University at Buffalo released Axion1.0, an agent foundation model trained with a minimal pipeline of just two SFT stages plus RL — no mid-training, DPO, or online policy distillation — and open-sourced the full data construction, training code, data and models.

Pipeline

Results

The method also works on small models (4B/9B reach 53.1%/63.1% on HLE-130 with SFT only) and closes a 9.2-point base-model gap to 0.8 points, showing the data recipe transfers across Qwen bases.

Original post →

More from coding & agent

coding & agent channel →