mlsanityData to 14 Oct 2026

Discovering Millions of Interpretable Features with Sparse Autoencoders

arXiv 2606.26620 ↗First posted 25 Jun 2026cs.LG cs.AIcited by 0

On the map: Mechanistic interpretability → SAE features and their applications

Abstract (arXiv)

Sparse autoencoders (SAEs) have emerged as a powerful tool for decomposing superposed language model representations into sparse and interpretable features. However, training SAEs is computationally expensive, and available open-source SAE models remain limited. In this work, we introduce Qwen3-Instruct SAE, a comprehensive suite of SAEs trained on the Qwen3 instruction-tuned model family, covering Qwen3-1.7B, Qwen3-4B, and Qwen3-8B. For Qwen3-1.7B and Qwen3-4B, we train layer-wise SAEs at three key activation sites: residual streams, MLP outputs, and attention outputs. For Qwen3-8B, we train SAEs on a subset of residual stream layers. We systematically evaluate these SAEs using both activation-level reconstruction metrics and model-level recovery metrics, revealing distinct sparsity--fidelity trade-offs across layers and components. Finally, we demonstrate the utility of Qwen3-Instruct SAE through a refusal-steering case study, showing that selected SAE features can causally steer instruction-tuned Qwen3 models toward refusal behavior. Our release provides a practical resource for studying sparse representations, feature-level mechanisms, and behavioral interventions in instruction-tuned language models

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in SAE features and their applications, one of 13 directions in Mechanistic interpretability. Builds on 35 of its 53 references that are in the map. First posted on arXiv 25 Jun 2026.

Builds on

35 of 53 references.

Closest ancestors · 7

Same direction, newest first

  1. 2025SAEs Are Good for Steering -- If You Select the Right Features26 May
  2. 2025SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability12 Mar
  3. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders28 Jan
  4. 2024Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders27 Oct
  5. 2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders22 Sep
  6. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders19 Jul
  7. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Models15 Sep

Other references · 28

  1. 2013Efficient Estimation of Word Representations in Vector Spacecited by 4,375
  2. 2024HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusalcited by 1,429
  3. 2020Transformer Feed-Forward Layers Are Key-Value Memoriescited by 1,336
  4. 2024The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scalecited by 1,160
  5. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
  6. 2023Steering Language Models With Activation Engineeringcited by 922
  7. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 724
  8. 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 712
  9. 2024WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMscited by 496
  10. 2023SmoothLLM: Defending Large Language Models Against Jailbreaking Attackscited by 469
  11. 2023Catastrophic Jailbreak of Open-source LLMs via Exploiting Generationcited by 451
  12. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
16 more
  1. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  2. 2023Towards Best Practices of Activation Patching in Language Models: Metrics and Methodscited by 359
  3. 2023Finding Neurons in a Haystack: Case Studies with Sparse Probingcited by 334
  4. 2025Layer by Layer: Uncovering Hidden Representations in Language Modelscited by 311
  5. 2019Analyzing the Structure of Attention in a Transformer Language Modelcited by 243
  6. 2023LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70Bcited by 195
  7. 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 143
  8. 2024Safety Layers in Aligned Large Language Models: The Key to LLM Securitycited by 134
  9. 2024Not All Layers of LLMs Are Necessary During Inferencecited by 96
  10. 2024Steering Language Model Refusal with Sparse Autoencoderscited by 95
  11. 2025rStar2-Agent: Agentic Reasoning Technical Reportcited by 54
  12. 2024Mathematical Models of Computation in Superpositioncited by 35
  13. 2023Dynamical versus Bayesian Phase Transitions in a Toy Model of Superpositioncited by 29
  14. 2026Controllable LLM Reasoning via Sparse Autoencoder-Based Steeringcited by 15
  15. 2025Marco-o1 v2: Towards Widening The Distillation Bottleneck for Reasoning Modelscited by 13
  16. 2025Feature-Guided SAE Steering for Refusal-Rate Control using Contrasting Promptscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models
  2. 2024Improving Dictionary Learning with Gated Sparse Autoencoders
  3. 2026Are Sparse Autoencoder Benchmarks Reliable?
  4. 2025Revisiting End-To-End Sparse Autoencoder Training: A Short Finetune Is All You Need
  5. 2025Sparse Autoencoders Trained on the Same Data Learn Different Features
  6. 2025FaithfulSAE: Towards Capturing Faithful Features with Sparse Autoencoders without External Dataset Dependencies
  7. 2024Scaling and evaluating sparse autoencoders
  8. 2025On the Limits of Sparse Autoencoders: A Theoretical Framework and Reweighted Remedy