Convergent Evolution: How Different Language Models Learn Similar Number Representations

Hugging Face Daily Papers Papers

Summary

Study reveals that diverse language-model architectures independently evolve similar periodic Fourier features for representing numbers, with only some achieving geometric separability for modular arithmetic.

Language models trained on natural text learn to represent numbers using periodic features with dominant periods at T=2, 5, 10. In this paper, we identify a two-tiered hierarchy of these features: while Transformers, Linear RNNs, LSTMs, and classical word embeddings trained in different ways all learn features that have period-T spikes in the Fourier domain, only some learn geometrically separable features that can be used to linearly classify a number mod-T. To explain this incongruity, we prove that Fourier domain sparsity is necessary but not sufficient for mod-T geometric separability. Empirically, we investigate when model training yields geometrically separable features, finding that the data, architecture, optimizer, and tokenizer all play key roles. In particular, we identify two different routes through which models can acquire geometrically separable features: they can learn them from complementary co-occurrence signals in general language data, including text-number co-occurrence and cross-number interaction, or from multi-token (but not single-token) addition problems. Overall, our results highlight the phenomenon of convergent evolution in feature learning: A diverse range of models learn similar features from different training signals.
Original Article
View Cached Full Text

Cached at: 04/23/26, 03:35 AM

Paper page - Convergent Evolution: How Different Language Models Learn Similar Number Representations

Source: https://huggingface.co/papers/2604.20817

Abstract

Transformers and other language models exhibit periodic numerical representations in their Fourier domains, with some models developing geometrically separable features for linear classification of numbers modulo T, though Fourier sparsity alone is insufficient for this separability.

Language models trained on natural text learn to represent numbers using periodic features with dominant periods at T=2, 5, 10. In this paper, we identify a two-tiered hierarchy of these features: whileTransformers,Linear RNNs,LSTMs, and classicalword embeddingstrained in different ways all learn features that haveperiod-T spikesin theFourier domain, only some learn geometrically separable features that can be used to linearly classify a numbermod-T. To explain this incongruity, we prove thatFourier domain sparsityis necessary but not sufficient formod-Tgeometric separability. Empirically, we investigate when model training yields geometrically separable features, finding that the data, architecture, optimizer, and tokenizer all play key roles. In particular, we identify two different routes through which models can acquire geometrically separable features: they can learn them from complementaryco-occurrence signalsin general language data, includingtext-number co-occurrenceandcross-number interaction, or from multi-token (but not single-token) addition problems. Overall, our results highlight the phenomenon ofconvergent evolutionin feature learning: A diverse range of models learn similar features from different training signals.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2604\.20817

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/2604.20817 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2604.20817 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2604.20817 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

Modular Cognitive Architecture Emerges in Large Language Models

arXiv cs.AI

This paper investigates whether modular cognitive architecture emerges in large language models, finding that LLMs develop specialized neural networks mirroring human brain organization across cognitive domains, suggesting modularity is a fundamental property of intelligent systems.