Parts-of-Speech as Emergent Categories in SAE Latent Space
Alessandro Bondielli, Lucia Passaro, Serena Auriemma, Alessandro Lenci
University of Pisa
Published September 24, 2026 · on the feed September 25, 2026 · ▲ 13 · arXiv 2609.29362
The abstract
Sparse AutoEncoders (SAEs) offer a promising way to inspect language model representations, but it is still unclear what kind of linguistic structure their latents expose. We use part-of-speech (PoS) categories as a controlled test case to study whether morpho-syntactic information is encoded by individual latents or by structured groups of features. We find that PoS distinctions are highly recoverable from SAE activations, but do not align with one-to-one latent / category mappings. This recoverability is not reducible to lexical memorisation, and Open and Closed PoS classes differ substantially. Categories are supported by compact groups of sparse latents, with substantial variation across tags. These groups remain stable on held-out data, while also showing overlap between related categories. Our results show that SAEs localise morpho-syntactic information in a distributed and category-dependent form rather than through atomic grammatical features.
The Brief
1:48A short podcast summary of the paper. Written and voiced by AI.
Transcript
Parts-of-Speech as Emergent Categories in SAE Latent Space. By Alessandro Bondielli and three others, from the University of Pisa. Published on the twenty-fourth of September. Its code repository is public and real: about thirty source files.
Sparse autoencoders are a promising way to look inside language model representations, but it has stayed unclear what kind of linguistic structure their latents actually expose. This study uses part-of-speech categories as a controlled test case, and asks a precise question. Is morpho-syntactic information carried by individual latents, one per category, or by structured groups of features? The answer it reports is the second. Sparse autoencoders localise this information in a distributed, category-dependent form, not through atomic grammatical features.
Part-of-speech distinctions turn out to be highly recoverable from SAE activations. They simply do not line up with one-to-one mappings between a latent and a category. That recoverability is not reducible to lexical memorisation, so the signal is not just the model remembering which words belong where. Instead, each category is supported by a compact group of sparse latents. Those groups remain stable on held-out data, and related categories show overlap between their groups.
The picture is not uniform. The supporting groups vary substantially from tag to tag, and Open and Closed part-of-speech classes differ substantially from each other. The test case is also deliberately narrow: parts of speech were chosen because they are controlled, so the result speaks to morpho-syntactic information rather than to linguistic structure in general.
On this evidence, looking for the one latent that means noun is the wrong search. Looking for a small team of latents per category is the right one. Grammar is present in the sparse code. It is spread across groups whose shape depends on the category being asked about.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A clean dictionary of concepts is the promise of these autoencoders. Grammar is a fair place to test it.
You are looking inside a language model with a sparse autoencoder, a tool that rewrites each hidden state as a few active latents out of many. People hope each latent is one clean concept.
The road ahead
- 1 · The problem ~7 min A sparse autoencoder turns a model's hidden state into a few active latents, but it is still unclear how it stores grammar like parts of speech.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Each part-of-speech tag is carried by a small group of latents, not one latent, and open-class tags need broader groups than closed ones.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Probe the latents for each tag, rule out word memory with a control, find each tag's latent group, then test those groups on held-out data.
- 6 · The experiment ~6 min You’ll guess: Inside a sparse autoencoder on a language model, how are part-of-speech categories carried by its latents?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.