pipette
ENEnglish

Procedural Memory for Language Agents: A Neurosymbolic Perspective on Explicit Goal Structure

Jonathan Leung, Zhiqi Shen, Deheng Ye

Preprint

En palabras de los autores

Large language models remain unreliable on long-horizon tasks requiring procedural reasoning. We argue that a key cause is the absence of explicit procedural memory: accessible, modifiable knowledge of how tasks are accomplished, distinct from semantic and episodic memory. Goal-Oriented Graphs offer such a representation, encoding goals and prerequisite dependencies as inspectable symbolic structure. We examine their construction, the challenges they pose, and principles for integrating them with neural reasoning.

Resultado principalEl resumen no menciona limitaciones.

Apareció: martes, 20 de octubre. arXiv. Preprint, todavía sin revisión por pares.