SCALAR: Benchmarking SAE Interaction Sparsity in Toy LLMs

The recent introduction of SCALAR (Sparse Connectivity Assessment of Latent Activation Relationships) provides a new benchmark for evaluating interaction sparsity in sparse autoencoders (SAEs), a crucial aspect of mechanistic interpretability in neural networks. Traditional methods have focused on individual SAE performance without addressing how features interact across layers. SCALAR aims to fill this gap by comparing various SAE models, including TopK SAEs, Jacobian SAEs (JSAEs), and the newly proposed Staircase SAEs. The findings reveal that Staircase SAEs significantly enhance relative sparsity, outperforming TopK SAEs by 59.67% in feedforward layers and 63.15% in transformer blocks. JSAEs also show some improvement, but they struggle with transformer blocks. This research underscores the importance of interaction sparsity in SAEs, paving the way for more efficient neural network designs and deeper insights into their functioning.