SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data

TL;DR

SynthSAEBench is a comprehensive benchmark toolkit for evaluating Sparse Autoencoders on large-scale, realistic synthetic data, enabling precise diagnosis of architectural improvements and failure modes.

Contribution

We introduce SynthSAEBench, a scalable synthetic data benchmark with a standardized model for detailed SAE evaluation and analysis.

Findings

Reproduces known LLM SAE phenomena

Identifies superposition noise exploitation as a failure mode

Highlights overfitting risks with more expressive encoders

Abstract

Improving Sparse Autoencoders (SAEs) requires benchmarks that can precisely validate architectural innovations. However, current SAE benchmarks on LLMs are often too noisy to differentiate architectural improvements, and current synthetic data experiments are too small-scale and unrealistic to provide meaningful comparisons. We introduce SynthSAEBench, a toolkit for generating large-scale synthetic data with realistic feature characteristics including correlation, hierarchy, and superposition, and a standardized benchmark model, SynthSAEBench-16k, enabling direct comparison of SAE architectures. Our benchmark reproduces several previously observed LLM SAE phenomena, including the disconnect between reconstruction and latent quality metrics, poor SAE probing results, and a precision-recall trade-off mediated by L0. We further use our benchmark to identify a new failure mode: Matching Pursuit SAEs exploit superposition noise to improve reconstruction without learning ground-truth features, suggesting that more expressive encoders can easily overfit. SynthSAEBench complements LLM benchmarks by providing ground-truth features and controlled ablations, enabling researchers to precisely diagnose SAE failure modes and validate architectural improvements before scaling to LLMs.