The Year in AI Papers
The year/Independent research

Paper 2602.14111

Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?

Anton Korznikov, Andrey Galichin, Alexey Dontsov, Oleg Rogov, Ivan Oseledets, Elena Tutubalina

Published
Feb 2026
Research lab
Independent
Citations
7
GitHub
Not linked

01 In brief

Summary

This paper evaluates whether Sparse Autoencoders (SAEs) recover meaningful features from neural networks.

In synthetic experiments with known ground-truth features, SAEs achieved 71% explained variance but recovered only 9% of true features, showing a disconnect between reconstruction fidelity and feature recovery.

On real LLM activations, the authors introduced three baselines that freeze SAE components to random values: Frozen Decoder, Soft-Frozen Decoder, and Frozen Encoder.

Across multiple SAE architectures (BatchTopK, JumpReLU, ReLU) and models (Gemma-2-2B, Llama-3-8B), these baselines matched fully-trained SAEs on interpretability (0.87 vs 0.90), sparse probing (0.69 vs 0.72), and causal editing (0.73 vs 0.72).

The results suggest that current SAEs do not reliably decompose model internals, and the authors propose their baselines as sanity checks for future SAE evaluation.

02 From the paper

Abstract

Sparse Autoencoders (SAEs) have emerged as a promising tool for interpreting neural networks by decomposing their activations into sparse sets of human-interpretable features. Recent work has introduced multiple SAE variants and successfully scaled them to frontier models. Despite much excitement, a growing number of negative results in downstream tasks casts doubt on whether SAEs recover meaningful features. To directly investigate this, we perform two complementary evaluations. On a synthetic setup with known ground-truth features, we demonstrate that SAEs recover only $9\%$ of true features despite achieving $71\%$ explained variance, showing that they fail at their core task even when reconstruction is strong. To evaluate SAEs on real activations, we introduce three baselines that constrain SAE feature directions or their activation patterns to random values. Through extensive experiments across multiple SAE architectures, we show that our baselines match fully-trained SAEs in interpretability (0.87 vs 0.90), sparse probing (0.69 vs 0.72), and causal editing (0.73 vs 0.72). Together, these results suggest that SAEs in their current state do not reliably decompose models' internal mechanisms.