Researchers propose AREX-2, a self-improving LLM agent that iteratively refines solutions through reflection and long-horizon execution. They demonstrate the agent's effectiveness on various benchmarks, showing potential for sustained improvement.