Are Single-Token Sparse Autoencoder Features Causally Necessary? Layer-Depth and SAE-Family Effects

2026-07-25 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

The study investigates whether Sparse Autoencoder (SAE) features, used for interpreting and steering large language models, maintain stable causal roles across different SAE families. Analyzing 3.9M features from six models (GPT2-Small, Gemma-2-2B/9B, Gemma-3-1B, Llama-3.1-8B, DeepSeek-R1-8B) and three SAE families (GemmaScope/res-jb, LlamaScope, BatchTopK), researchers found single-token features cluster 4.7$\times$ tighter in decoder space and concentrate in early layers (Layer 0 in GPT2-Small; L0–L4 in Gemma). Ablating these features caused Benjamini-Hochberg-significant logit reductions in 178 of 208 full-layer conditions. Crucially, cross-family causal differences exceeded within-family scale effects: GemmaScope and BatchTopK features remained causally anchored, while LlamaScope features were locally redundant, with target token rank recovering 96–98% of the time. This indicates interpretability claims are sensitive to training methodology.

Key takeaway

For AI Scientists and Machine Learning Engineers working on LLM interpretability or steering, you must account for the specific Sparse Autoencoder family used. Your feature's causal necessity cannot be assumed portable across different SAE families, even on the same base model. Therefore, re-run ablation checks when deploying new SAE families and prioritize direct causal profiling over inferring behavior from family names. Incorporate per-feature causal necessity into your SAE evaluation benchmarks to ensure robust interpretability.

Key insights

A feature's causal necessity is real but not portable across Sparse Autoencoder families, depending on training methodology.

Principles

• Single-token features are geometrically and causally distinct.
• Causal necessity varies by layer depth and semantic category.
• SAE training methodology shapes feature detection and causal importance.

Method

The study used zero-ablation interventions, replacing residual-stream activations and measuring logit reductions for top-activating tokens, validated with Mann-Whitney U tests and BH correction.

In practice

• Re-verify causal claims when switching SAE families.
• Profile a checkpoint's causal behavior directly.
• Add per-feature causal necessity to SAE evaluation benchmarks.

Topics

• Sparse Autoencoders
• LLM Interpretability
• Causal Necessity
• Feature Ablation
• SAE Families
• Training Methodology

Code references

• jbloomAus/SAELens

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects — Sparse Autoencoder features' causal effects are complex and context-dependent, rarely providing stable, one-dimensional steering directions.
• From Geometric Recovery to Causal Validation: A Reproducible Audit of Sparse Autoencoder Features, from Superposition Geometry to Causal Inertness — High correlational recovery in Sparse Autoencoders often masks causally inert features, necessitating deeper validation.
• SAEExplainer: Interpreting SAE Features with Activation-Guided Preference Optimization — SAEExplainer uses activation-guided preference optimization for self-correcting, iterative SAE feature interpretation, reducing hallucinations.
• Evaluating the Interpretability of Sparse Autoencoders with Concept Annotations — A new framework evaluates Sparse Autoencoder interpretability by aligning latents with human concepts using synthetic data and perturbation scores.
• Look But Don't Touch with Sparse Autoencoders for Unlearning in Diffusion Models — Sparse autoencoders excel at concept detection in diffusion models, but direct latent intervention for unlearning causes artifacts; detection-based replacement is superior.
• Towards Isolated Interventions via Almost Orthogonal Features in Language Models — Orthogonality regularization of language model features enables isolated causal interventions while preserving performance.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.