mlsanityData to 14 Oct 2026

Parts-of-Speech as Emergent Categories in SAE Latent Space

arXiv 2609.29362 ↗First posted 24 Sep 2026cs.CLcited by 0

On the map: Mechanistic interpretability → SAE features and their applications

Abstract (arXiv)

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.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in SAE features and their applications, one of 13 directions in Mechanistic interpretability. Builds on 14 of its 21 references that are in the map. First posted on arXiv 24 Sep 2026.

Builds on

14 of 21 references.

Closest ancestors · 3

Same direction, newest first

  1. 2025SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability12 Mar
  2. 2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probing23 Feb
  3. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Models15 Sep

Other references · 11

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2019BERT Rediscovers the Classical NLP Pipelinecited by 1,059
  3. 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 876
  4. 2020A Primer in BERTology: What we know about how BERT workscited by 815
  5. 2019Designing and Interpreting Probes with Control Taskscited by 640
  6. 2018What you can cram into a single vector: Probing sentence embeddings for linguistic propertiescited by 457
  7. 2019What do you learn from context? Probing for sentence structure in contextualized word representationscited by 438
  8. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  9. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  10. 2024Not All Language Model Features Are One-Dimensionally Linearcited by 212
  11. 2020Information-Theoretic Probing for Linguistic Structurecited by 207

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025Dense SAE Latents Are Features, Not Bugs
  2. 2025A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models
  3. 2025Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoders
  4. 2026Descriptive Collision in Sparse Autoencoder Auto-Interpretability: When One Explanation Describes Many Features
  5. 2025Large Language Models Share Representations of Latent Grammatical Concepts Across Typologically Diverse Languages
  6. 2025Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders
  7. 2024Quantifying Feature Space Universality Across Large Language Models via Sparse Autoencoders
  8. 2026Persistent Sparse Autoencoders: Learning Feature-Specific Timescales in Language Model Representations