Geometric Factual Recall in Transformers

Hugging Face Daily Papers Papers

Summary

This paper introduces a theoretical framework for geometric factual recall in transformers, demonstrating that embeddings can encode relational structure via linear superpositions while MLPs act as selectors. It provides empirical and theoretical evidence that this mechanism allows for efficient memorization of facts and multi-hop queries.

How do transformer language models memorize factual associations? A common view casts internal weight matrices as associative memories over pairs of embeddings, requiring parameter counts that scale linearly with the number of facts. We develop a theoretical and empirical account of an alternative, geometric form of memorization in which learned embeddings encode relational structure directly, and the MLP plays a qualitatively different role. In a controlled setting where a single-layer transformer must memorize random bijections from subjects to a shared attribute set, we prove that a logarithmic embedding dimension suffices: subject embeddings encode linear superpositions of their associated attribute vectors, and a small MLP acts as a relation-conditioned selector that extracts the relevant attribute via ReLU gating, and not as an associative key-value mapping. We extend these results to the multi-hop setting -- chains of relational queries such as ``Who is the mother of the wife of x?'' -- providing constructions with and without chain-of-thought that exhibit a provable capacity-depth tradeoff, complemented by a matching information-theoretic lower bound. Empirically, gradient descent discovers solutions with precisely the predicted structure. Once trained, the MLP transfers zero-shot to entirely new bijections when subject embeddings are appropriately re-initialized, revealing that it has learned a generic selection mechanism rather than memorized any particular set of facts.
Original Article
View Cached Full Text

Cached at: 05/13/26, 04:10 AM

Paper page - Geometric Factual Recall in Transformers

Source: https://huggingface.co/papers/2605.12426

Abstract

Transformer language models use geometric memorization where embeddings encode linear superpositions of attributes and MLPs act as relation-conditioned selectors rather than associative key-value mappings.

How dotransformer language modelsmemorize factual associations? A common view castsinternal weight matricesasassociative memoriesover pairs of embeddings, requiring parameter counts that scale linearly with the number of facts. We develop a theoretical and empirical account of an alternative, geometric form of memorization in which learned embeddings encode relational structure directly, and theMLPplays a qualitatively different role. In a controlled setting where a single-layer transformer must memorize random bijections from subjects to a shared attribute set, we prove that a logarithmicembedding dimensionsuffices: subject embeddings encodelinear superpositionsof their associated attribute vectors, and a smallMLPacts as a relation-conditioned selector that extracts the relevant attribute viaReLU gating, and not as an associative key-value mapping. We extend these results to the multi-hop setting -- chains ofrelational queriessuch as ``Who is the mother of the wife of x?‘’ -- providing constructions with and withoutchain-of-thoughtthat exhibit a provable capacity-depth tradeoff, complemented by a matchinginformation-theoretic lower bound. Empirically,gradient descentdiscovers solutions with precisely the predicted structure. Once trained, theMLPtransfers zero-shot to entirely new bijections when subject embeddings are appropriately re-initialized, revealing that it has learned a generic selection mechanism rather than memorized any particular set of facts.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2605\.12426

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.12426 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.12426 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.12426 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

A mathematical theory of balancing relational generalization and memorization

arXiv cs.LG

This paper introduces a novel task, transitive inference with exceptions, and analytically characterizes how neural network models (kernel ridge regression) balance relational generalization and memorization. The theory is validated in pretrained language models, showing systematic mistakes predicted by the theory.

Transformers Learn the Mestre-Nagao Heuristic

arXiv cs.LG

This paper trains a two-layer transformer encoder to classify rational elliptic curves by rank from Frobenius traces, achieving >99% accuracy. Mechanistic interpretability reveals the model learns the Mestre-Nagao heuristic and concentrates attention on prime positions, demonstrating that transformers can learn number-theoretic algorithms.