Preprint reports agent success gains from saving corrections in adapters

Researchers report in a preprint that Self-Direction, a method for retaining lessons from failed tasks, improved Qwen-based agents' success rates by 8.9 to 17.0 percentage points across research, tool-use and terminal benchmarks. After a failure, the agent reviews its actions, extracts a lesson and trains a small adapter while keeping the original model weights frozen. Each adapter adds parameters equivalent to ~0.02% of the base model's parameter count. The agent stores adapters in a library and selects them for related tasks later. The reported experiments cover GAIA, Claw-Eval and Terminal-Bench 2.1 using open-weight Qwen-family models.