Researchers from Princeton University and Meta’s FAIR division developed Self-Direction, a method that converts an AI agent’s past mistakes into reusable skill modules. Each correction is stored in a rank-16 LoRA adapter that modifies roughly 0.02% of the model’s parameters without requiring retraining. On Qwen3.6-27B models, success rates increased by 17.0 percentage points on GAIA, 13.8 points on Terminal-Bench 2.1 and 8.9 points on Claw-Eval. The rate at which agents resolved previously encountered mistakes nearly doubled, rising from 9.7% to 18.9%, while the correction rate for first-time mistakes remained at 67.3%. The modular design could reduce computing costs and allow teams to inspect, update or remove individual skills without altering the rest of the system.
Source: Read the original article
