# Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks

arXiv: https://arxiv.org/abs/2411.18895 · HTML: https://mlsanity.com/paper/2411.18895/

## Key facts

- Cited by 15 papers in the map (data through 14 Oct 2026).
- Placed in Sparse autoencoder training and evaluation, one of 13 directions in Mechanistic interpretability.
- First posted on arXiv 28 Nov 2024.

## Placed in

- Mechanistic interpretability / [Sparse autoencoder training and evaluation](https://mlsanity.com/mechinterp/sparse-autoencoder-training-and-evaluation/)

## Built on by (most cited first, 15 of 15)

- [Learning Multi-Level Features with Matryoshka Sparse Autoencoders](https://mlsanity.com/paper/2503.17547/) (2025)
- [How does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding](https://mlsanity.com/paper/2507.22928/) (2025)
- [Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words](https://mlsanity.com/paper/2501.06254/) (2025)
- [Evaluating Adversarial Robustness of Concept Representations in Sparse Autoencoders](https://mlsanity.com/paper/2505.16004/) (2025)
- [SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data](https://mlsanity.com/paper/2602.14687/) (2026)
- [AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features](https://mlsanity.com/paper/2510.00404/) (2025)
- [Ensembling Sparse Autoencoders](https://mlsanity.com/paper/2505.16077/) (2025)
- [Concept-Guided Interpretability via Neural Chunking](https://mlsanity.com/paper/2505.11576/) (2025)
- [Aligned Training: A Parameter-Free Method to Improve Feature Quality and Stability of Sparse Autoencoders (SAE)](https://mlsanity.com/paper/2605.18629/) (2026)
- [ReSAE: Residualized Sparse Autoencoders for Multi-Layer Transformer Interventions](https://mlsanity.com/paper/2605.27819/) (2026)
- [Discovering Chunks in Neural Embeddings for Interpretability](https://mlsanity.com/paper/2502.01803/) (2025)
- [GPT and Prejudice: A Sparse Approach to Understanding Learned Representations in Large Language Models](https://mlsanity.com/paper/2510.01252/) (2025)
- [Feature Evolution and Migration during Vision Transformer Training](https://mlsanity.com/paper/2608.20134/) (2026)
- [When Pruning Meets Interpretability: Preserving Sparse Autoencoder Robustness in LLMs](https://mlsanity.com/paper/2608.25941/) (2026)
- [Clinical Concept Centers in LLMs](https://mlsanity.com/paper/2610.02829/) (2026)

## Abstract (arXiv)

Sparse Autoencoders (SAEs) are an interpretability technique aimed at decomposing neural network activations into interpretable units. However, a major bottleneck for SAE development has been the lack of high-quality performance metrics, with prior work largely relying on unsupervised proxies. In this work, we introduce a family of evaluations based on SHIFT, a downstream task from Marks et al. (Sparse Feature Circuits, 2024) in which spurious cues are removed from a classifier by ablating SAE features judged to be task-irrelevant by a human annotator. We adapt SHIFT into an automated metric of SAE quality; this involves replacing the human annotator with an LLM. Additionally, we introduce the Targeted Probe Perturbation (TPP) metric that quantifies an SAE's ability to disentangle similar concepts, effectively scaling SHIFT to a wider range of datasets. We apply both SHIFT and TPP to multiple open-source models, demonstrating that these metrics effectively differentiate between various SAE training hyperparameters and architectures.