Researchers have introduced AREX-2, a system aimed at advancing the self-improving capability of LLM agents, which the authors define as the ability to iteratively refine a solution at test time. The work is presented as an AI research paper and is accompanied by a public code repository and model collection.
The authors argue that this self-improvement rests on two complementary abilities: reflection, which produces a solution better than the current one, and long-horizon execution, which keeps the iteration effective across many rounds. Their central hypothesis is that both capabilities are domain-agnostic and can therefore be learned in settings that are well suited for supervision.
To test this, the team synthesized long-horizon improvement trajectories from machine learning and algorithmic programming tasks — two domains chosen because they offer verifiable feedback and reward sustained iteration. The resulting agent, built on Qwen3.8-27B, was trained on this data.
According to the paper, the trained agent achieves strong results on MLE-bench Lite (81.8) and Frontier-CS (70.7), and transfers to deep research tasks with 84.0 on BrowseComp, 52.6 on HLE, 92.2 on GAIA, and 93.8 on DeepSearchQA. The authors also report that performance keeps improving as the budget of rounds grows.
The paper concludes that long-horizon reflective data is an effective route toward self-improving agents. A homepage and model collection are linked alongside the publication.