A provocative new algorithm of AI learning: not inside the neural weights, but in the recursive coupling between a frozen LLM, persistent harness state and environmental feedback.
WikiSkill is an offline, outer-loop skill-development system for advanced tasks and human queries.
Can a frozen LLM accumulate genuine procedural knowledge without updating a single parameter? WikiSkill introduces a file-based learning architecture that converts immutable agent trajectories into persistent Wiki patterns, lets an LLM-based ReAct proposer compile this cross-iteration knowledge into executable SKILL.md files, and admits each modification only through validation-gated rollback, while preserving even rejected experiments as future evidence.
Across five models and five benchmarks, this external skill compiler delivers gains of up to 23.9 points, enables a skilled 9B model to outperform an unskilled 27B model, and reveals a provocative new algorithm of AI learning: not inside the neural weights, but in the recursive coupling between a frozen LLM, persistent harness state and environmental feedback.
All rights w/ authors:
WikiSkill: Compiling Agent Experience into
Persistent Knowledge for Skill Evolution
Liyan Tang1, Cyrus Rashtchian1, Chun-Sung Ferng1, Andrew Tomkins1, Da-Cheng Juan1 and Tu Vu1,2
from
1 Google Research,
2 Virginia Tech
#airesearch
#aiexplained
#aiskills
source
Leave a Reply