Sparse autoencoders trained as a stack, one per taxonomy level, read out unseen concepts through their hypernyms and gain 31.3 points on abstract-concept accuracy
AspireAI community · arXiv (Guo, Yang, Morrison, Blanco & Surdeanu)
TaxoSAE treats a taxonomy as a stack of sparse autoencoders, one per level, trained on a frozen LLM's last-token hidden state, with a taxonomic alignment loss that requires each concept's slot to fire whenever one of its hyponyms does. It needs explicit data only for terminal concepts. On two factual QA taxonomies, one over biographical relations and one over animals, the paper reports readout accuracy on abstract concepts never seen in training is 31.3 percentage points above the strongest baseline, with activations concentrated into 1.0 to 1.7 slots per abstract concept. Control improves too: editing a concept's feature steers the model to another concept with 25.5 points higher success, and suppressing a concept removes only that concept's answers, leaving 0 to 2 percent accuracy on it against 91 to 100 percent on other concepts. The generalisation claim is the one worth the attention: with a hyponym bound but the concept itself unseen, the unseen concept still activates its hypernym's slot, which the baselines rarely do, for up to 37.9 points higher hypernym readout accuracy. The baselines are three-sparse-autoencoder stacks that differ only in how they assign a slot to each element, so the comparison isolates the taxonomic structure rather than capacity.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.