Convergent Evolution: How Different Language Models Learn Similar Number Representations

• URL: http://arxiv.org/abs/2604.20817v1
• Date: Wed, 22 Apr 2026 17:45:27 GMT
• Title: Convergent Evolution: How Different Language Models Learn Similar Number Representations
• Abstract summary: Language models trained on natural text learn to represent numbers using periodic features with dominant periods at $T=2, 5, 10$.<n>Transformers, Linear RNNs, LSTMs, and classical word embeddings all learn features that have period-$T$ spikes in the Fourier domain.<n>Only some learn geometrically separable features that can be used to linearly classify a number mod-$T$.
• Score: 53.8769088563044
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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.

Related papers

• Learning Coherent Representations: A Topological Approach to Interpretability [0.0]
  Deep neural networks learn representations where individual features often lack interpretable meaning.<n>We introduce coherence, a geometric property inspired by neural coding in the brain, where neurons like grid cells respond to contiguous regions of state space.
  arXiv  Detail & Related papers  (2026-06-01T20:05:58Z)
• Vocabulary embeddings organize linguistic structure early in language model training [3.2661767443292646]
  Large language models (LLMs) work by manipulating the geometry of input embedding vectors over multiple layers.<n>Here, we ask: how are the input vocabulary representations of language models structured, and how does this structure evolve over training?<n>We run a suite of experiments that correlate the geometric structure of the input embeddings and output embeddings of two open-source models with semantic, syntactic, and frequency-based metrics over the course of training.
  arXiv  Detail & Related papers  (2025-10-08T23:26:22Z)
• Why Can't Transformers Learn Multiplication? Reverse-Engineering Reveals Long-Range Dependency Pitfalls [54.57326125204404]
  Language models are increasingly capable, yet still fail at a seemingly simple task of multi-digit multiplication.<n>We study why, by reverse-engineering a model that successfully learns multiplication via emphimplicit chain-of-thought'
  arXiv  Detail & Related papers  (2025-09-30T19:03:26Z)
• Learning Compositional Functions with Transformers from Easy-to-Hard Data [63.96562216704653]
  We study the learnability of the $k$-fold composition task, which requires computing an interleaved composition of $k$ input permutations and $k$ hidden permutations.<n>We show that this function class can be efficiently learned, with runtime and sample in $k$, by gradient descent on an $O(log k)$-depth transformer.
  arXiv  Detail & Related papers  (2025-05-29T17:22:00Z)
• In-Context Language Learning: Architectures and Algorithms [73.93205821154605]
  We study ICL through the lens of a new family of model problems we term in context language learning (ICLL) We evaluate a diverse set of neural sequence models on regular ICLL tasks.
  arXiv  Detail & Related papers  (2024-01-23T18:59:21Z)
• Do Neural Networks Trained with Topological Features Learn Different Internal Representations? [1.418465438044804]
  We investigate whether a model trained with topological features learns internal representations of data that are fundamentally different than those learned by a model trained with the original raw data. We find that structurally, the hidden representations of models trained and evaluated on topological features differ substantially compared to those trained and evaluated on the corresponding raw data. We conjecture that this means that neural networks trained on raw data may extract some limited topological features in the process of making predictions.
  arXiv  Detail & Related papers  (2022-11-14T19:19:04Z)
• Benchmarking Compositionality with Formal Languages [64.09083307778951]
  We investigate whether large neural models in NLP can acquire the ability tocombining primitive concepts into larger novel combinations while learning from data. By randomly sampling over many transducers, we explore which of their properties contribute to learnability of a compositional relation by a neural network. We find that the models either learn the relations completely or not at all. The key is transition coverage, setting a soft learnability limit at 400 examples per transition.
  arXiv  Detail & Related papers  (2022-08-17T10:03:18Z)
• Linear Connectivity Reveals Generalization Strategies [54.947772002394736]
  Some pairs of finetuned models have large barriers of increasing loss on the linear paths between them. We find distinct clusters of models which are linearly connected on the test loss surface, but are disconnected from models outside the cluster. Our work demonstrates how the geometry of the loss surface can guide models towards different functions.
  arXiv  Detail & Related papers  (2022-05-24T23:43:02Z)
• A Deep Structural Model for Analyzing Correlated Multivariate Time Series [11.009809732645888]
  We present a deep learning structural time series model which can handle correlated multivariate time series input. The model explicitly learns/extracts the trend, seasonality, and event components. We compare our model with several state-of-the-art methods through a comprehensive set of experiments on a variety of time series data sets.
  arXiv  Detail & Related papers  (2020-01-02T18:48:29Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.