Edit Knowledge, Not Just Facts via Multi-Step Reasoning over Background Stories

Ya Gao, Kalle Kujanpää, Pekka Marttinen, Harri Valpola, Alexander Ilin · 2026

Read Edit Knowledge, Not Just Facts via Multi-Step Reasoning over Background Stories by Ya Gao, Kalle Kujanpää, Pekka Marttinen, Harri Valpola, Alexander Ilin (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-02028.txt.

Description

Ya Gao and colleagues argue knowledge update is a reasoning problem, not memorization, so models should train where new information solves a task. Existing knowledge editing emphasizes atomic facts, improving recall but failing to weave updated information into a coherent framework. Their strategy rests on three principles. New knowledge arrives as a background story, and models train on self-generated multi-hop questions. Distillation forces a student model to absorb the teacher's reasoning without access. Experiments show models trained this way leverage new knowledge on questions combining multiple new facts.

Catalog record

CREATORYa Gao, Kalle Kujanpää, Pekka Marttinen, Harri Valpola, Alexander Ilin
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE156079
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602.02028

Browse the full library · About the Tower of Babel