Do LLMs dream of elephants (when told not to)? Latent concept association and associative memory in transformers

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Jiang, Yibo, Rajendran, Goutham, Ravikumar, Pradeep, Aragam, Bryon
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916497747607552
author Jiang, Yibo
Rajendran, Goutham
Ravikumar, Pradeep
Aragam, Bryon
author_facet Jiang, Yibo
Rajendran, Goutham
Ravikumar, Pradeep
Aragam, Bryon
contents Large Language Models (LLMs) have the capacity to store and recall facts. Through experimentation with open-source models, we observe that this ability to retrieve facts can be easily manipulated by changing contexts, even without altering their factual meanings. These findings highlight that LLMs might behave like an associative memory model where certain tokens in the contexts serve as clues to retrieving facts. We mathematically explore this property by studying how transformers, the building blocks of LLMs, can complete such memory tasks. We study a simple latent concept association problem with a one-layer transformer and we show theoretically and empirically that the transformer gathers information using self-attention and uses the value matrix for associative memory.
format Preprint
id arxiv_https___arxiv_org_abs_2406_18400
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Do LLMs dream of elephants (when told not to)? Latent concept association and associative memory in transformers
Jiang, Yibo
Rajendran, Goutham
Ravikumar, Pradeep
Aragam, Bryon
Computation and Language
Machine Learning
Large Language Models (LLMs) have the capacity to store and recall facts. Through experimentation with open-source models, we observe that this ability to retrieve facts can be easily manipulated by changing contexts, even without altering their factual meanings. These findings highlight that LLMs might behave like an associative memory model where certain tokens in the contexts serve as clues to retrieving facts. We mathematically explore this property by studying how transformers, the building blocks of LLMs, can complete such memory tasks. We study a simple latent concept association problem with a one-layer transformer and we show theoretically and empirically that the transformer gathers information using self-attention and uses the value matrix for associative memory.
title Do LLMs dream of elephants (when told not to)? Latent concept association and associative memory in transformers
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2406.18400