Researchers propose Recursive Environment Scaling (RES), a framework for making agent benchmarks more challenging without altering the underlying tasks. The method transforms existing benchmark environments so that agents face greater difficulty while the core objectives remain the same.

According to the paper's abstract, the goal is to use environment scaling as a lever to increase benchmark difficulty. Because the task objectives are fixed, performance changes can be attributed to the environment's complexity rather than to a different problem.

RES is positioned as relevant to general long-horizon agents, suggesting the authors intend the scaling method to apply across a range of sequential decision-making benchmarks. The abstract excerpt does not specify the exact transformation rules, so implementation details are not yet public.