Princeton and Meta FAIR's Self-Direction turns agent errors into micro-LoRA adapters

Researchers from Princeton and Meta FAIR published Self-Direction (arXiv:2610.31846), which distills an agent's execution mistakes into micro-LoRA adapters covering about 0.02% of base model parameters while keeping base weights frozen. On GAIA, Claw-Eval and Terminal-Bench 2.1 the method raised task completion by 8.9–17.0 percentage points and nearly doubled correction reliability.

• Adapters update ~0.02% of model parameters with base weights frozen
• Task completion rose 8.9–17.0 points on GAIA, Claw-Eval, Terminal-Bench 2.1
• Reliability of repeat corrections nearly doubled
• Adapters train locally in seconds on consumer hardware