AI Emergence: Improv Games between 2 Agents

Two agents (gemma-3-4b @ Temp=1) play 15 games of Mind Meld, with the same initial pair of words, and mostly converge after 3 rounds.

We have run 1000s of experiments observing and analyzing how small LLMs (1B to 14B) interact with each other.

One of our favorite experimental architectures is Mind Meld, a fun human improv game where 2 agents work to converge on the same single word, depending only on the previous pair of words.  We track the similarity of their words at each round of Mind Meld - creating "similarity trajectories" .  We also track *when* convergence happens.

Sometimes AI pairs crash through chaos before suddenly snapping into alignment (those stars). Others never quite agree. The question isn't just if they converge—it's how, and what that means.  We love it :)
​
WHY IS THIS IMPORTANT? Could multi-agent debates unlock new optimization? Inform decision-making? Spark creative solutions? Self-correct without humans in the loop? We're exploring the edges where cooperation becomes computation, and the data hints at something deeper. We're still figuring out what. 

Stay tuned for more results and other games of emergence.