Definition
Plain language
A sudden reorganization of a model's internal state that flips its answer between iterations.
As stated in the literature
In iterative/recurrent inference, a discrete event where a token's hidden state moves between qualitatively distinct attractor basins, causing a large change in the argmax output.
Also called: basin shifts
Why it matters: Basin shifts explain how small additional reasoning steps can produce qualitatively different answers, which matters for the stability and reproducibility of iterative inference.
For example, after the seventh refinement step a token might suddenly flip from 'yes' to 'no' as its hidden state crosses into a different attractor basin.
Heard on the show
“Aviss calls these moments *basin shifts*.”Episode 032 — A Sticky-Note for Every Layer: Letting Transformers Remember What They Were Just Thinking