Parts-of-Speech as Emergent Categories in SAE Latent Space
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.
Community
Hi all! This is our new paper on how Sparse Autoencoders encode parts of speech, which was accepted at EMNLP 2026 🎉
SAE latents are often described as monosemantic features, so we tested whether linguistic categories like POS map onto individual latents. Using SAE activations from LLaMA-3-8B on the UD English GUM treebank, we found that:
- POS is highly recoverable from SAE activations, and not just through lexical memorisation.
- There's no one-latent-per-tag mapping. Each POS relies on a compact group of latents, and just 498 of them (out of ~130k) are enough for multi-class POS classification.
- Closed classes (DET, PRON, CCONJ) use small, focused groups, while open classes (NOUN, ADJ, ADV) spread over broader ones.
- These groups are stable on held-out data but overlap across related categories, so they also pick up surface form and context.
In short: SAEs capture POS in a distributed, category-dependent way rather than through atomic grammatical features.
Code:
https://github.com/colinglab/pos-sae-latents
Data: https://huggingface.co/datasets/colinglab/UD_English-GUM-Latents_Meta-Llama-3-8B_L30
Curious to hear what people think about the paper and wether this would hold up under causal interventions or with other SAE architectures!
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper