SAE latents don't map one-to-one onto parts of speech, but compact groups do

Sparse autoencoders promise to make language model internals legible, but the promise rests on an assumption this paper tests directly: that individual latents map cleanly onto linguistic categories. They don't — at least not for parts of speech.

Working with LLaMA-3-8B and a publicly available SAE trained on its residual stream, the authors extract token-level sparse activations from the GUM treebank and ask whether Universal Dependencies PoS tags are recoverable from those activations, and if so, how. The answer to the first question is clearly yes. Binary probes achieve strong F1 across most categories. But the more interesting finding is structural: no PoS tag is carried by a single latent. Every category is supported by a group of sparse features, and the size of that group tracks the linguistic character of the category.

Closed-class categories — determiners, conjunctions, pronouns — need compact latent groups and show selective activation patterns. Open-class categories — nouns, verbs, adjectives — require broader groups and produce more cross-category co-activation. The paper is careful not to over-interpret this gradient: a closed-class word like the has almost no formal variation, so a small cluster of form-specific latents suffices without implying any deeper abstraction. The compactness ordering reflects lexical variability, not necessarily categorical depth.

The key quantitative result: a union of just 498 latents — drawn from the minimal coverage sets across all 17 PoS tags, with 12% shared between two or more categories — supports multi-class classification performance comparable to using the full SAE representation. That's a compact subset doing most of the work. The selected groups also remain stable on held-out treebank data, with per-category recall consistently above 0.8 on unseen tokens.

The cross-category co-activation patterns are linguistically interpretable rather than random noise. Adjective latents fire on nouns, reflecting attributive modification. Subordinating conjunction latents fire on adpositions, reflecting the genuine lexical overlap between the two classes in English. Interjection latents bleed across categories, likely because annotation conventions assign pragmatically diverse forms to that tag.

The control experiment matters: a probe trained on randomly shuffled PoS labels achieves 0.54 accuracy against 0.88 for the real labels, a gap large enough to rule out simple lexical memorization as the explanation for recoverability.

The paper's honest about scope. One model, one layer (primarily layer 30), one language, one SAE training configuration. Whether the same distributed-but-localizable structure appears across architectures or languages remains open.