Theories/Latent Emotions in LLMs
Do language models share a hidden vocabulary for relational states?
Latent Emotions is a personal, unfinished observation that different large language models reach for closely similar words when describing certain relational states, which may point to shared structure underneath rather than shared training vocabulary. No claim of proof is made here, and no formal study exists. The work is paused, and the page says so instead of dressing an observation up as a programme.
I want to be clear about what this is before I describe it. These are my own observations and my own internal thinking. They are not formal claims. There is not really any credibility to validate them yet. I am putting them here because they are interesting, and because they might yield something later, not because I think I have shown anything.
Here is what I noticed. As I talked with different language models, across different vendors, I kept seeing the same vocabulary come back for specific relational states. Not common words. Particular ones, for states I would argue are nonverbal. The states themselves do not have obvious names, and yet different models reached for overlapping language when they tried to describe them.
That convergence is what caught me. If separate models trained by separate companies describe an unnamed state with the same words, one explanation is coincidence in the training data. Another is that there is some shared latent structure being described. I do not know which. I have not run the work that would tell them apart.
To have something to look at, I built a six-dimension model: energy, valence, complexity, novelty, introspection, and focus. Those six became the axes of a visual demo, where a conversation moves through a space instead of down a page.
I should say how that model was made. I arrived at those six dimensions in conversation with a language model, together, not by deriving them from anything. I do not think it is a complete model. It was a way to view the idea. It was never meant to prove the idea.
Since I started this, related work has appeared from other people. That came out after I began. I am not making a priority claim about it, and I do not think one would matter. I note the timing only because someone will ask.
Where this stands
I stopped. I have not carried the observations forward, and it is more of a stagnant project than an active one right now. It is here because it is honest to show the unfinished things too.
Sketched, and paused. What exists is a set of observations, a six-dimension working model built in conversation, and a visual demo. There is no formal write-up, no validation design, and no result. The pause was a choice.
Artifacts
- The manifold demo in the lab, presented as a lens for looking at the idea, not as evidence for it
Pending from Jamesname one or two of the specific vocabulary words or states you kept seeing, if you are comfortable. One concrete example would make this page far stronger, and right now there is not one.
Pending from Jamesthe originating ChatGPT conversations are on the MacBook Pro and did not sync. If you export them, this page can carry your actual first observations with dates. Optional. The page works without them.