Summary
Most agent builders have written the same loop: try something, check the result, decide whether to retry, keep the best version, and know when to stop. AREX-2 from the Beijing Academy of Artificial Intelligence (BAAI) asks whether the model itself could handle that work instead of the surrounding harness.
This video from Sam Witteveen examines the model and its accompanying paper. AREX-2 is a 27B open-weights model under the Apache 2 license, fine-tuned from Qwen 3.8 27B for long-running agent tasks. It is designed around two ideas, reflection and long-horizon execution, which together let it use a multi-hour budget to keep improving its answers. The training recipe is supervised fine-tuning rather than reinforcement learning, selecting trajectories by how they end and keeping failed rounds in context while training only on steps that make progress.
On the reported benchmarks, AREX-2 scores 81.8 on MLE-Bench Lite, ahead of GPT 5.6 Soul at 72.7, and 70.7 on Frontier CS, the best among open-weights models though about six points behind GPT 5.6 Soul. Its score climbs from 54 after one hour to 70.7 after five. The video also covers the skills ablation, and what the results mean for developers building their own agent harnesses and fine-tuning on agent traces.
📺 Source: Sam Witteveen · Published November 06, 2026
🏷️ Format: Deep Dive
