Recursive Environment Scaling for General Long-Horizon Agents

Recursive Environment Scaling hardens agent benchmarks and improves fine-tuning by up to 14.61 points.

Recursive Environment Scaling (RES) increases agent-benchmark difficulty without changing core objectives by altering information access, resource relations, and execution dependencies, then validating candidates and inheriting accepted changes. On 294 initial tasks from three benchmarks, all six evaluated models succeed less often, with drops of 57.14 to 77.14 percentage points on Claw-Eval-Live. With equal trajectory counts and identical optimization, fine-tuning on RES trajectories beats original-trajectory supervised fine-tuning in all nine model-benchmark comparisons by 7.00 to 14.61 points. Ablations point to recursive inheritance and dependency composition as key sources of difficulty.