Are Single-Token Sparse Autoencoder Features Causally Necessary? Layer-Depth and SAE-Family Effects
Not on the map. Why?
Abstract (arXiv)
Sparse autoencoder (SAE) features are used to interpret and steer large language models, yet nobody has tested whether a feature's causal role is stable across SAE families. Single-token features fire on one vocabulary item, so ground truth permits direct comparison. We analyze 3.9M features across six models and three SAE families and zero-ablate at full layer depth: they sit 4.7x tighter in decoder space and concentrate in early layers. Deleting one lowers the model's logit for that token in 178 of 208 layer conditions, significant after multiple-comparison correction. And depth decides how the damage lands: early-layer deletions disrupt the layers that follow, late-layer deletions change the output directly. Cross-family causal differences exceed within-family scale effects: on the same base model, GemmaScope and BatchTopK features are causally anchored, LlamaScope features locally redundant. Under LlamaScope the token returns to within 2x its pre-ablation rank 96-98% of the time. Changing only the activation function reverses the sign of that difference, so the training recipe is the remaining candidate: cross-family claims are sensitive to training methodology, not just activation function or scale.
Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Builds on 23 of its 42 references that are in the map. First posted on arXiv 22 Jul 2026.
Builds on
23 of 42 references.
References · 23
- 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,928
- 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,359
- 2018Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representationscited by 929
- 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 412
- 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 381
- 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 283
- 2013k-Sparse Autoencoderscited by 209
- 2024Not All Language Model Features Are One-Dimensionally Linearcited by 201
- 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 171
- 2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoderscited by 148
- 2024Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoderscited by 133
- 2025SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretabilitycited by 118
11 more
- 2024Improving Steering Vectors by Targeting Sparse Autoencoder Featurescited by 86
- 2024Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Controlcited by 83
- 2025Sparse Autoencoders Do Not Find Canonical Units of Analysiscited by 80
- 2025SAEs Are Good for Steering -- If You Select the Right Featurescited by 74
- 2024Identifying Functionally Important Features with End-to-End Sparse Dictionary Learningcited by 64
- 2025Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometrycited by 58
- 2024The Geometry of Concepts: Sparse Autoencoder Feature Structurecited by 52
- 2024Quantifying Feature Space Universality Across Large Language Models via Sparse Autoencoderscited by 25
- 2024Mechanistic Permutability: Match Features Across Layerscited by 21
- 2024Evolution of SAE Features Across Layers in LLMscited by 14
- 2026Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?cited by 10
Built on by
Citing papers per year: 2026: 1.
1 paper.