LLM Conversations Form Attractor States in Latent Space During Mixed Play

We can actually pin this to a number of concrete behavioral changes, such as model copying each other traits such as metacommentary:

If you're interested in more details, please check out the paper at https://arxiv.org/abs/2606.30571. We'll also be at ICML next weeks discussing this in workshops, come say hi!

All of this comes out to 40 pages of hard work, all thanks to @ko175041 who is leading this study!

But here is what happens when talking to other models:

The arrows in the second figure show the 'movement' in latent space towards the partner in the conversation, notice the pull towards Haiku

Here are cleaner zoom-ins of just single model pairs:

@jonasgeiping @ko175041 Very interesting! Did you compare reasoning vs. non-reasoning models?

In our unscaffolded free-discussion setting, we find quite different semantic-space dynamics between reasoning and non-reasoning agents.

https://arxiv.org/abs/2605.17885