Provides mechanistic evidence that LLMs internalize 'vibes' (informal registers like slang) as language-agnostic abstractions that can be causally steered.
March 30, 2026
Original Paper
A Universal Vibe? Finding and Controlling Language-Agnostic Informal Register with SAEs
arXiv · 2603.26236
The Takeaway
Using Sparse Autoencoders (SAEs), researchers isolated a cross-linguistic core for informal register and demonstrated that activation steering transfers zero-shot to six unseen languages. This allows developers to control the pragmatic tone of a model across diverse languages without language-specific fine-tuning.
From the abstract
While multilingual language models successfully transfer factual and syntactic knowledge across languages, it remains unclear whether they process culture-specific pragmatic registers, such as slang, as isolated language-specific memorizations or as unified, abstract concepts. We study this by probing the internal representations of Gemma-2-9B-IT using Sparse Autoencoders (SAEs) across three typologically diverse source languages: English, Hebrew, and Russian. To definitively isolate pragmatic r