Semantic Convergence: Investigating Shared Representations Across Scaled LLMs

TL;DR

This paper examines whether large language models of different sizes develop similar internal features, finding that models converge on comparable concepts especially in middle layers, supporting the idea of feature universality across models.

Contribution

It introduces a novel method using SAE and alignment techniques to compare internal features across scaled LLMs, demonstrating significant overlap in middle layers.

Findings

Middle layers show strongest feature overlap.

Early and late layers have less similarity.

Semantic subspaces interact similarly across models.

Abstract

We investigate feature universality in Gemma-2 language models (Gemma-2-2B and Gemma-2-9B), asking whether models with a four-fold difference in scale still converge on comparable internal concepts. Using the Sparse Autoencoder (SAE) dictionary-learning pipeline, we utilize SAEs on each model's residual-stream activations, align the resulting monosemantic features via activation correlation, and compare the matched feature spaces with SVCCA and RSA. Middle layers yield the strongest overlap, while early and late layers show far less similarity. Preliminary experiments extend the analysis from single tokens to multi-token subspaces, showing that semantically similar subspaces interact similarly with language models. These results strengthen the case that large language models carve the world into broadly similar, interpretable features despite size differences, reinforcing universality as a foundation for cross-model interpretability.