[KongchangAI]
ConceptProgram-Aided Patchable Skill Evolution

PASE

An agent skill evolution framework that decomposes a skill into an executable program module and a natural language description, and achieves targeted updating and iterative evolution of the skill through program patches.

Timeline (last 90 days)

Nov 11

The PASE framework represents agent skills as a separation between executable programs and natural language descriptions: the executable program carries the procedural logic, while the natural language retains descriptive knowledge.

Unverified50%
Nov 11

Experiments on multiple models and multiple benchmarks show that PASE consistently outperforms all baselines, with gains exceeding 2.8%.

Unverified50%
Nov 11

PASE performs better in analyses of evolutionary stability and generalization ability.

Unverified50%
Nov 11

PASE's data and code will be released via GitHub.

Unverified50%
Nov 11

The PASE paper emphasizes that programmatic knowledge plays a key role in skill evolution.

Unverified50%

All Facts (5)

Source Articles