AREX-2 introduces self-improving agent trained on long-horizon reflective tasks

A new paper introduces AREX-2, a framework designed to advance the self-improving capability of LLM-based agents. The authors define self-improvement as the ability to iteratively refine a solution at test time, relying on two core capabilities: reflection, which produces a better solution than the current one, and long-horizon execution, which sustains effective iteration over many rounds.

The authors hypothesize that these capabilities are domain-agnostic and can be learned from supervised data. To test this, they synthesize long-horizon improvement trajectories from machine learning and algorithmic programming tasks, domains that provide verifiable feedback and reward sustained iteration.

AREX-2 is built on the Qwen3.8-27B model and is evaluated on multiple benchmarks. On MLE-bench Lite, it achieves a score of 81.8, and on Frontier-CS, it scores 70.7. The agent also transfers to deep research tasks, with scores of 84.0 on BrowseComp, 52.6 on HLE, 92.2 on GAIA, and 93.8 on DeepSearchQA. Performance improves as the agent’s iteration budget increases.

The paper concludes that long-horizon reflective data is an effective route toward building self-improving agents, suggesting a path to more autonomous and continuously improving AI systems.