Measuring Progress in Dictionary Learning for Language Model Interpretability with Board Game Models

TL;DR

This paper introduces a new approach to evaluate and improve interpretable feature extraction in language models by using board game data, providing supervised metrics and a novel training technique.

Contribution

It proposes a supervised evaluation framework for autoencoders in language models using chess and Othello data, and introduces p-annealing to enhance interpretability learning.

Findings

p-annealing improves SAE performance on interpretability metrics

Supervised metrics effectively evaluate interpretable features in game transcripts

Framework enables measuring progress in disentangling features in language models

Abstract

What latent features are encoded in language model (LM) representations? Recent work on training sparse autoencoders (SAEs) to disentangle interpretable features in LM representations has shown significant promise. However, evaluating the quality of these SAEs is difficult because we lack a ground-truth collection of interpretable features that we expect good SAEs to recover. We thus propose to measure progress in interpretable dictionary learning by working in the setting of LMs trained on chess and Othello transcripts. These settings carry natural collections of interpretable features -- for example, "there is a knight on F3" -- which we leverage into $\textit{supervised}$ metrics for SAE quality. To guide progress in interpretable dictionary learning, we introduce a new SAE training technique, $\textit{p-annealing}$, which improves performance on prior unsupervised metrics as well as our new metrics.