The authors evaluated agents on GAIA, covering research and assistant tasks; Claw-Eval, covering tool-use workflows; and Terminal-Bench 2.1, covering terminal tasks.
Small learned adapters help AI agents avoid repeating their mistakes
Across three benchmarks, Kansal and colleagues report higher task success when agents convert reflections on failures into reusable parameter updates.
Big Tech
Yuval Kansal · Karen Ullrich · Polina Kirichenko · Ilia Kulikov · Niraj K. Jha · Mohammad Pezeshki · +1 more
Princeton University · Meta FAIR
Research Digest··2 min read
Kansal and colleagues introduce Self-Direction, a method that turns an agent’s reflection on a failed task into a small, separately stored model adapter.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.