Model Interpretability

Parts-of-Speech Categories Emerge in the SAE Latent Space

EMNLP 2026 research shows POS in LLaMA-3-8B SAEs is recoverable via compact latent groups, not one latent per …