Daily Paper Briefing

Abstract Overview

This paper studies how part-of-speech (PoS) information is represented in Sparse AutoEncoder (SAE) latent activations extracted from LLaMA-3-8B. Using the English GUM treebank and a controlled evaluation set, the authors test whether PoS distinctions are recoverable from token-level SAE activations and whether they correspond to single latents or groups of latents. They find that PoS categories are strongly recoverable with linear probes, but the evidence does not support a one-latent-per-category interpretation. Instead, PoS information appears to be localized in compact yet distributed latent groups whose size and selectivity vary by category, with clearer compactness for many closed-class tags than for open-class tags.

Novelty

The paper’s main novelty is a category-level analysis of SAE latent space using PoS as a controlled morphosyntactic testbed, moving beyond top-activating examples or generic probe accuracy. It introduces a pipeline combining recoverability, salience, coverage/compactness, and held-out validation to characterize whether linguistic categories emerge as single latent features or structured groups.

Results

One-vs-rest probes show that PoS distinctions are consistently recoverable from SAE activations, while coverage analysis shows that categories require different numbers of salient latents to reach high coverage rather than mapping one-to-one to individual latents. A compact union of 498 selected latents preserves multi-class PoS classification performance close to the full SAE feature set, and these latent groups remain highly active on held-out data, though with noticeable cross-category overlap. Control experiments further indicate that the effect is not reducible to lexical memorization alone.

Key Points

1. PoS information is highly recoverable from SAE activations, but it is represented by structured groups of latents rather than isolated one-to-one latent/category mappings.
2. Closed-class PoS tend to have more compact and selective latent groups, whereas open-class PoS show broader, more overlapping activation patterns tied to lexical and contextual variability.
3. A small subset of salient latents supports strong multi-class PoS prediction and generalizes to held-out data, but the learned groups are only partially category-specific.

References

• arXiv: https://arxiv.org/abs/2609.29362v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2609.29362v1
• Hugging Face Papers: https://huggingface.co/papers/2609.29362
• GitHub: https://github.com/colinglab/pos-sae-latents