analiza aiqbench
This research demonstrates that Sparse AutoEncoders (SAEs) successfully localize morpho-syntactic information within large language models, revealing that parts-of-speech categories emerge from distributed groups of latents rather than atomic features. The findings indicate that these latent structures are stable on held-out data and exhibit distinct characteristics between open and closed class tags, offering new insights into how grammatical knowledge is encoded in neural representations.
Parts-of-Speech as Emergent Categories in SAE Latent Space
arXiv:2609.29362v1 Announce Type: new 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.
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.
Computation and Language (cs.CL)
arXiv:2609.29362 [cs.CL]
(or arXiv:2609.29362v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2609.29362 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 231 cuvinte
Subiecte conexe
Runway’s WorldPrompt and the Engineering of Real-Time Worlds
Latent Space · 25 sept. 2026
[AINews] Jev: a “System One Model” that only decides/classifies/routes/scores — >100x faster, >200x cheaper than small frontier LLMs
Latent Space · 16 sept. 2026
Latent reasoning architectures would likely undermine CoT, our strongest oversight tool
AI Alignment Forum · 23 sept. 2026
I Did This Without Your Knowledge
LessWrong · 17 oct. 2026