Princeton and Meta propose Self-Direction, saving agent fixes as LoRA adapters

Princeton and Meta researchers propose Self-Direction, a method that trains an AI agent's failure lessons into LoRA adapters, which a router reloads on similar tasks.

Each adapter is about 0.02% of the model's size. Rohan Paul, who shared the paper on X, reports that with Qwen3.6-27B, GAIA task success rose from 38.8% to 55.8%.

The paper says 67.3% of repairs still worked by the third encounter with a task, compared with 42.9% for plain retrying. The paper's claims come from its own experiments on the GAIA benchmark.