ML//Inference//semantic continuity

How well a model maintains thematic and register coherence across a generation.

How well a model maintains thematic and register coherence across a generation.

Not logical coherence; something subtler: whether the text reads as if written by the same author, in the same context, with the same intent.

Base models (pre-alignment) had it sporadically; RLHF and DPO trained models have it almost always.

Breaks down when the prompt pushes the model into a region of the latent space it can't sustain, or when temperature is too high and sampling drifts off-distribution.

Distinct from factual accuracy; a model can be semantically continuous while hallucinating confidently.