Saved in:
Bibliographic Details
Main Authors: Dugan, Owen, Garcia, Roberto, Junkins, Ronny, Liu, Jerry, Zinsley, Dylan, Eyuboglu, Sabri, Rudra, Atri, Ré, Chris
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2512.00207
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911294083301376
author Dugan, Owen
Garcia, Roberto
Junkins, Ronny
Liu, Jerry
Zinsley, Dylan
Eyuboglu, Sabri
Rudra, Atri
Ré, Chris
author_facet Dugan, Owen
Garcia, Roberto
Junkins, Ronny
Liu, Jerry
Zinsley, Dylan
Eyuboglu, Sabri
Rudra, Atri
Ré, Chris
contents The success of large language models (LLMs) can be attributed in part to their ability to efficiently store factual knowledge as key-value mappings within their MLP parameters. Recent work has proposed explicit weight constructions to build such fact-storing MLPs, providing an improved understanding of LLM fact storage mechanisms. In this paper, we introduce an MLP construction framework that improves over previous constructions in three areas: it 1) works for all but a measure-zero set of feasible input-output pairs, 2) achieves asymptotically optimal parameter efficiency matching information-theoretic bounds for some embeddings, and 3) maintains usability within Transformers for factual recall. Through our improvements, we 1) discover a metric on value embeddings that characterizes facts-per-parameter scaling for both constructed and gradient-descent-trained MLPs, 2) identify a simple encoder-decoder mechanism that empirically matches gradient-descent MLP facts-per-parameter asymptotics across all the inputs and outputs we test, and 3) uncover a fundamental tradeoff between an MLP's fact-storage capacity and its usability within Transformers. Finally, we demonstrate a proof-of-concept application of fact-storing MLPs: modular fact editing on one-layer Transformers by \textit{replacing entire MLPs at once}.
format Preprint
id arxiv_https___arxiv_org_abs_2512_00207
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Constructing Efficient Fact-Storing MLPs for Transformers
Dugan, Owen
Garcia, Roberto
Junkins, Ronny
Liu, Jerry
Zinsley, Dylan
Eyuboglu, Sabri
Rudra, Atri
Ré, Chris
Machine Learning
Artificial Intelligence
The success of large language models (LLMs) can be attributed in part to their ability to efficiently store factual knowledge as key-value mappings within their MLP parameters. Recent work has proposed explicit weight constructions to build such fact-storing MLPs, providing an improved understanding of LLM fact storage mechanisms. In this paper, we introduce an MLP construction framework that improves over previous constructions in three areas: it 1) works for all but a measure-zero set of feasible input-output pairs, 2) achieves asymptotically optimal parameter efficiency matching information-theoretic bounds for some embeddings, and 3) maintains usability within Transformers for factual recall. Through our improvements, we 1) discover a metric on value embeddings that characterizes facts-per-parameter scaling for both constructed and gradient-descent-trained MLPs, 2) identify a simple encoder-decoder mechanism that empirically matches gradient-descent MLP facts-per-parameter asymptotics across all the inputs and outputs we test, and 3) uncover a fundamental tradeoff between an MLP's fact-storage capacity and its usability within Transformers. Finally, we demonstrate a proof-of-concept application of fact-storing MLPs: modular fact editing on one-layer Transformers by \textit{replacing entire MLPs at once}.
title Constructing Efficient Fact-Storing MLPs for Transformers
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2512.00207