Awesome Papers for Sparse Auto-Encoder (SAE)

This list focuses on sparse auto-encoder (SAE) techniques in mechanistic interpretability. Another list focuses on understanding the internal mechanism of LLMs.

Paper/preprint/blog recommendation: please release a issue or contact me.

Papers

2025

• Probing the Vulnerability of Large Language Models to Polysemantic Interventions

  • [arxiv] [2025.5]

• SAE-SSV: Supervised Steering in Sparse Representation Spaces for Reliable Control of Language Models

  • [arxiv] [2025.5]

• Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoders

  • [arxiv] [2025.5]

• On the Biology of a Large Language Model

  • [Anthropic] [2025.3]

• A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models

  • [SAE survey] [2025.3]

• SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability

  • [arxiv] [2025.3]

• Sparse Autoencoders Do Not Find Canonical Units of Analysis

  • [ICLR 2025] [2025.2]

• Interpreting and Steering LLMs with Mutual Information-based Explanations on Sparse Autoencoders

  • [arxiv] [2025.2]

• Scaling Sparse Feature Circuits For Studying In-Context Learning

  • [openreview] [2025.1]

• AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders

  • [ICML 2025] [2025.1]

• Enhancing Automated Interpretability with Output-Centric Feature Descriptions

  • [arxiv] [2025.1]

2024

• BatchTopK Sparse Autoencoders

  • [arxiv] [2024.12]

• Stage-Wise Model Diffing

  • [Anthropic] [2024.12]

• Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models

  • [ICLR 2025] [2024.11]

• Improving Steering Vectors by Targeting Sparse Autoencoder Features

  • [arxiv] [2024.11]

• Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks

  • [arxiv] [2024.11]

• Applying sparse autoencoders to unlearn knowledge in language models

  • [arxiv] [2024.10]

• Evaluating feature steering: A case study in mitigating social biases

  • [Anthropic] [2024.10]

• Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders

  • [arxiv] [2024.10]

• Efficient Training of Sparse Autoencoders for Large Language Models via Layer Groups

  • [arxiv] [2024.10]

• Sparse Autoencoders Reveal Temporal Difference Learning in Large Language Models

  • [ICLR 2025] [2024.10]

• Sparse Crosscoders for Cross-Layer Features and Model Diffing

  • [Anthropic] [2024.10]

• Scaling Automatic Neuron Description

  • [transluce] [2024.10]

• Evaluating Open-Source Sparse Autoencoders on Disentangling Factual Knowledge in GPT-2 Small

  • [arxiv] [2024.9]

• Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2

  • [Deepmind] [2024.8]

• Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders

  • [Deepmind] [2024.8]

• Self-explaining SAE features

  • [Lesswrong blog] [2024.8]

• SAEs (usually) Transfer Between Base and Chat Models

  • [AI alignment forum blog] [2024.7]

• BatchTopK: A Simple Improvement for TopK-SAEs

  • [Lesswrong blog] [2024.7]

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

  • [ICML MI workshop] [2024.7]

• Interpreting Attention Layer Outputs with Sparse Autoencoders

  • [arxiv] [2024.6]

• Transcoders Find Interpretable LLM Feature Circuits

  • [arxiv] [2024.6]

• Scaling and evaluating sparse autoencoders

  • [OpenAI] [2024.6]

• Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning

  • [NeurIPS 2024] [2024.5]

• Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control

  • [arxiv] [2024.5]

• Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization

  • [arxiv] [2024.5]

• Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet

  • [Anthropic] [2024.5]

• How to use and interpret activation patching

  • [arxiv] [2024.4]

• Improving Dictionary Learning with Gated Sparse Autoencoders

  • [Deepmind] [2024.4]

• Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models

  • [arxiv] [2024.3]

• RAVEL: Evaluating Interpretability Methods on Disentangling Language Model Representations

  • [ACL 2024] [2024.2]

• Addressing Feature Suppression in SAEs

  • [lesswrong blog] [2024.2]

• Examining Language Model Performance with Reconstructed Activations using Sparse Autoencoders

  • [lesswrong blog] [2024.2]

• Dictionary Learning Improves Patch-Free Circuit Discovery in Mechanistic Interpretability: A Case Study on Othello-GPT

  • [arxiv] [2024.2]

• Open Source Sparse Autoencoders for all Residual Stream Layers of GPT2-Small

  • [alignmentforum blog] [2024.2]

2023

• Steering Llama 2 via Contrastive Activation Addition

  • [ACL 2024] [2023.12]

• Codebook Features: Sparse and Discrete Interpretability for Neural Networks

  • [arxiv] [2023.10]

• Attribution patching outperforms automated circuit discovery

  • [BlackboxNLP 2024] [2023.10]

• Towards Monosemanticity: Decomposing Language Models With Dictionary Learning

  • [Anthropic] [2023.10]

• Sparse Autoencoders Find Highly Interpretable Features in Language Models

  • [ICLR 2024] [2023.9]

• Steering Language Models With Activation Engineering

  • [arxiv] [2023.8]

• Inference-Time Intervention: Eliciting Truthful Answers from a Language Model

  • [NeurIPS 2023] [2023.6]

• Language models can explain neurons in language models

  • [OpenAI] [2023.5]

• Distributed Representations: Composition & Superposition

  • [Anthropic] [2023.5]

• Privileged Bases in the Transformer Residual Stream

  • [Anthropic] [2023.3]

• Attribution Patching: Activation Patching At Industrial Scale

  • [Neel Nanda Blog] [2023.2]

2022

• Causal Scrubbing: a method for rigorously testing interpretability hypotheses

  • [AI alignment forum] [2022.12]

• Taking features out of superposition with sparse autoencoders

  • [AI alignment forum] [2022.12]

• Engineering Monosemanticity in Toy Models

  • [arxiv] [2022.11]

• Polysemanticity and Capacity in Neural Networks

  • [arxiv] [2022.9]

• Toy Models of Superposition

  • [Anthropic] [2022.9]