Geometric Factual Recall in Transformers

Shauli Ravfogel, Gilad Yehudai, Joan Bruna, Alberto Bietti · 2026

Read Geometric Factual Recall in Transformers by Shauli Ravfogel, Gilad Yehudai, Joan Bruna, Alberto Bietti (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-2605-12426.txt.

Description

Ravfogel and three colleagues argue that transformers can memorise facts geometrically, not just as key-value pairs. The standard picture needs parameters that grow with the fact count, since weight matrices store pairs of embeddings. In a controlled test where a one-layer transformer must learn random subject-to-attribute maps, they prove a logarithmic embedding size suffices. Subject embeddings store sums of attribute vectors, and a small network picks the right attribute by gating. The account extends to multi-step chains of relational questions.

Catalog record

CREATORShauli Ravfogel, Gilad Yehudai, Joan Bruna, Alberto Bietti
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE91929
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.12426

Browse the full library · About the Tower of Babel