LLMs can tell when you're dictating. They can tell when another LLM is writing. They can tell what kind of keyboard layout you're using over enough text by the mistakes you make. What other 'modes' are LLMs detecting? Can we isolate 'mode' as a latent variable in LLM interpretation of text?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-modes-in-llm-2026,
author = {Holtzman, Ari},
title = {'Modes' in LLM Reading},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/RjfuHaexNEys0O727a4M}
}Please sign in to comment on this idea.
I’ve also been thinking about whether LLMs implicitly encode sociolinguistic signals (slang, region, etc.) as separable latent variables in hidden representations ; I'd love to discuss if this idea is open to volunteers!