ML//Inference//distributional shift//basin of attraction

Regions of latent space where certain types of continuations are strongly preferred — the model "falls into" them based on context.

Regions of latent space where certain types of continuations are strongly preferred — the model "falls into" them based on context.

Pretraining created these basins: "after reasoning → coherent conclusion" is one basin, "after question → direct answer" is another, "after nonsense → unpredictable continuation" is another.

Random text as context doesn't help because it positions the model in a basin with no semantic structure — the model has no signal about what type of continuation to produce.

Structured reasoning activates a specific basin because during pretraining, that pattern co-occurred systematically with correct conclusions. It's intentional distributional shift toward a useful attractor.

The technical effect: reasoning tokens reduce entropy of the output distribution — fewer plausible tokens as continuations, and the ones that remain tend to be more correct.

Overthinking happens when excessive reasoning tokens push the model past the useful basin into a "complex reasoning" attractor that's counterproductive for simple problems.

Exposure bias is falling into a bad attractor — once errors accumulate, the context vector enters a basin the model never navigated during training.