Reproduce “Attribution-Guided Distillation of Matryoshka Sparse Autoencoders” on OpenResearch

About this paper

Cristina P. Martin-Linares, Jonathan P. Ling

Distilled Matryoshka Sparse Autoencoders (DMSAEs) introduce an attribution-guided iterative distillation pipeline to identify and preserve a stable core of latent features in Sparse Autoencoders, enhancing feature consistency and transferability. The approach demonstrates improved performance on interpretability benchmarks such as SAEBench, particularly reducing feature absorption and increasing RAVEL scores.

Key insights

• Attribution-guided distillation successfully identifies a compact and stable core of latent features that persist across multiple training cycles, demonstrating effective feature consistency.
• The utility of the distilled core is empirically validated as models consistently rely on these features, unlike randomly initialized cores which tend towards zero activity.
• The stability and transferability of core latents are dependent on factors like sparsity target and attribution threshold, with intermediate sparsity levels yielding the most robust feature reuse.

Summary by alphaXiv.