Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet

2026-05-28

AI Summary

This work investigates the scalability of dictionary learning for extracting interpretable internal features from production-scale large language models. By applying sparse autoencoders to the residual stream of intermediate layers in Claude 3 Sonnet, the study successfully recovers features up to 34 million dimensions, guided by scaling laws for hyperparameter selection. The research further conducts multilingual, multimodal, and behavioral intervention experiments, demonstrating—for the first time at production scale—the feasibility of large-scale interpretable feature extraction. The extracted features reveal the model’s internal representations of abstract concepts and potentially hazardous behaviors such as deception and bias, while also exhibiting causal influence over model outputs. These features display both instance-specific responsiveness and abstract generalization capabilities, enabling targeted behavioral steering and identification of key harm-related mechanisms.

Abstract

We demonstrate that sparse autoencoders can extract interpretable features from Claude 3 Sonnet, a production-scale language model, addressing the open question of whether dictionary learning methods scale beyond small transformers. We trained sparse autoencoders with up to 34 million features on the model's middle layer residual stream, using scaling laws to guide hyperparameter selection. The resulting features are multilingual and multimodal (generalizing to images despite text-only training), respond to both concrete instances and abstract discussions of concepts, and can be used to steer model behavior in ways consistent with their interpretations. We find features corresponding to famous entities and locations, as well as more abstract concepts like sarcasm or errors in code. We also identify features relevant to ways in which language models might cause harm--including features representing deception, power-seeking, sycophancy, and bias--and show that these causally influence model outputs when manipulated. Additionally, we conduct analyses of feature interpretability, geometry, and computational function. However, significant limitations remain: our suite of features is incomplete, and we lack rigorous methods for evaluating whether our features faithfully capture model computations.

Problem

• monosemanticity
• interpretable features
• sparse autoencoders
• scaling
• language models

Innovation

• sparse autoencoders
• monosemanticity
• interpretable features
• dictionary learning
• model steering