Research · pugXiv:2606.00091
On the Emergence of Sitting in Scaled Systems
J. Snortsworth, L. Napworth
- Scaling
- Emergence
Abstract
We observe that sitting emerges abruptly at approximately 40 training treats and persists across model scales. The behavior is not explicitly trained for and is not reliably controllable, which we consider a promising direction.
1. Introduction
Some capabilities appear to emerge abruptly as models are scaled. We study one of the most consequential: sitting. We ask at what scale it appears, whether it persists, and whether anyone asked for it.
2. Method
We evaluate a family of models trained on between 1 and 1,000 training treats. Sitting is scored when the model’s hindquarters make contact with the floor and remain there for at least two seconds. Models were not explicitly trained to sit.
3. Results
Sitting emerges abruptly at approximately 40 training treats. Below that threshold, models stand, hover, or lean. Above it, sitting persists across every scale we tested. The transition is sharp: a model at 30 treats sat in 4% of trials, and a model at 50 treats sat in 86%.
Sitting is not reliably controllable. Models sit when asked, when not asked, and, in one case, during a fire drill.
4. Discussion
It has been argued that emergence is an artifact of the metric, and that a smoother score would show a smoother curve. We tried. The smoother score was “almost sitting,” and models at 30 treats did not receive it.
We consider the lack of controllability a promising direction.
5. Conclusion
Sitting emerges at a predictable scale and then does not leave. We recommend that practitioners plan for it.
Cite this paper
@article{emergence-of-sitting,
title = {On the Emergence of Sitting in Scaled Systems},
author = {J. Snortsworth and L. Napworth},
journal = {pugXiv},
number = {2606.00091},
year = {2026},
note = {Peer reviewed by a pug}
}