# Sparse Autoencoders Trained on the Same Data Learn Different Features

**Authors:** Gonçalo Paulo, Nora Belrose  
**arXiv:** [2501.16615](https://arxiv.org/abs/2501.16615)  
**Date:** 2025-01-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2501.16615v2/

## Key points

- Sparse Autoencoders (SAEs) used in Mechanistic Interpretability to reveal human-interpretable features in large language models (LLMs)
- SAEs trained with different random seeds for weight initialization identify distinct feature sets
- ReLU SAEs with L1 sparsity loss show greater stability across seeds, while TopK activation function SAEs are more seed-dependent
- Features uncovered by SAEs should be viewed as a practical decomposition of activation space, not an exhaustive list of truly utilized features
- Standard SAE designs may not capture the hierarchical structure inherent in human concepts adequately
- Debate on whether neural networks rely solely on linear representations or nonlinear features also play a crucial role
- Study highlights limitations and complexities in interpreting feature extraction through sparse autoencoders in large language models

## Summary

, , , , 

In the field of Mechanistic Interpretability, Sparse Autoencoders (SAEs) have been widely used to uncover human-interpretable features in the activations of large language models (LLMs). However, a recent study has revealed that SAEs trained on the same model and data, but with different random seeds for weight initialization, identify distinct sets of features. This phenomenon was observed across multiple layers of various LLMs, datasets, and SAE architectures. While ReLU SAEs trained with the L1 sparsity loss exhibited greater stability across seeds, those using the TopK activation function were more seed-dependent even when controlling for sparsity levels. These findings suggest that the set of features uncovered by an SAE should be seen as a practical decomposition of activation space rather than an exhaustive list of features truly utilized by the model. Furthermore, there is evidence suggesting that standard SAE designs may not adequately capture the hierarchical structure inherent in human concepts. Some studies have raised questions about whether neural networks solely rely on linear representations or if nonlinear features also play a crucial role. Overall, this study sheds light on the limitations and complexities associated with interpreting feature extraction through sparse autoencoders in large language models. The findings emphasize the need for a nuanced understanding of how these models operate and highlight the importance of considering alternative approaches to uncovering meaningful insights from neural network activations.