Self-Direction turns agent mistakes into LoRA lessons, lifting GAIA success from 38.8% to 55.8%

rohanpaul_ai · x · 2026-10-30

A Princeton × Meta FAIR paper introduces Self-Direction: when an agent fails, it reflects on its trajectory and distills the lesson into a LoRA adapter of only 0.02% of the base model's size, stored in a persistent lesson library with a router that loads the closest adapter for similar future tasks.

Key points:

Motivation: in-context reflections are forgotten across sessions; full fine-tuning is costly and can disrupt existing skills.
Tested on Qwen-family models across GAIA, Claw-Eval, and Terminal-Bench 2.1, improving success rates by 8.9–17.0 points; GAIA rose from 38.8% to 55.8%, beating in-context reflection and a Skills.md file.
Of 185 earlier mistakes, 67.3% of repairs still worked on the third encounter vs 42.9% for plain retrying.
A handful of adapters generalize to dozens of new tasks without extra training.

Practical takeaway: if your agents repeat mistakes, save each fix as a small reloadable adapter, not just prompt notes.

Original post →

Self-Direction turns agent mistakes into LoRA lessons, lifting GAIA success from 38.8% to 55.8%

More from coding & agent