Recursive Environment Scaling for General Long-Horizon Agents

Recursive Environment Scaling for General Long-Horizon Agents. By Lingxiao Jing and twelve others, from the University of Electronic Science and Technology of China, the Hong Kong University of Science and Technology, and four other institutions. Published on the twelfth of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Agent benchmarks need to get harder, but making a task harder usually means changing what the task is. The authors want difficulty to rise while the core objective of each task stays exactly where it was, so that the harder version can still be compared with the original.

Their answer is to scale the environment rather than the goal. Recursive Environment Scaling, or RES, takes existing benchmark environments and transforms them, while preserving their objectives and their completion requirements. The task asks for the same thing. The world the agent works in is what changes.

A proposer makes the changes. It alters information access, resource relations, and execution dependencies. Each resulting candidate is then validated by execution-grounded checks, and the changes that are accepted are inherited across rounds, which is where the recursion comes in: every round builds on the environment the last round left behind. The output is a family of related environments, usable both for paired evaluation against the original and for collecting training trajectories.

The experiments start from two hundred and ninety-four initial tasks drawn from three benchmarks. Task success drops for all six evaluation models, and on Claw-Eval-Live the decreases run from fifty-seven point one four up to seventy-seven point one four percentage points. On the training side, with equal trajectory counts and identical optimization settings, fine-tuning on RES trajectories beats fine-tuning on the original trajectories in all nine model and benchmark comparisons. The margins run from seven points up to fourteen point six one points. Ablations single out recursive inheritance and dependency composition as important contributors to the added difficulty.

One transformation serves two jobs. The same scaled environments that expose weaker agents also supply the trajectories that make agents stronger, and the authors read that as support for environment scaling in evaluation and training alike.