Discovering Millions of Interpretable Features with Sparse Autoencoders

2026-06-25

AI Summary

This work addresses the challenge of interpreting highly entangled internal representations in language models and the scarcity of scalable, open-source sparse autoencoders (SAEs). The authors systematically train hierarchical SAEs across multiple scales of the Qwen3 instruction-tuned model family, targeting key circuit locations—including residual streams, MLP outputs, and attention outputs—and release the first large-scale SAE suite spanning diverse model sizes and architectural positions. Through both activation-level and model-level evaluations, they characterize the trade-off between sparsity and reconstruction fidelity and demonstrate the causal manipulability of learned features in tasks such as refusal behavior. This resource provides a high-fidelity, practical toolkit for mechanistic interpretability research.

Abstract

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 \textbf{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

Problem

• sparse autoencoders
• interpretable features
• instruction-tuned language models
• feature steering
• model interpretability

Innovation

• Sparse Autoencoders
• Interpretable Features
• Instruction-Tuned Language Models
• Feature Steering
• Model Mechanisms