ML//inference//semantic continuity

Semantic continuity is a quality of generated text, the degree to which a model maintains thematic and register coherence across a generation.

Semantic continuity is a quality of generated text, the degree to which a model maintains thematic and register coherence across a generation.

It is subtler than logical coherence: it asks whether the text reads as if written by the same author, in the same context, with the same intent. It is also distinct from factual accuracy, since a model can be semantically continuous while hallucinating confidently.

Base models (before alignment) had it sporadically; models trained with RLHF and DPO have it almost always. Alignment delivered it as a side effect: the explicit goal was helpfulness, but training on human preferences also regularized register and tone.

It breaks down when the prompt pushes the model into a region of the latent space it cannot sustain.

It also breaks down when the temperature is too high and sampling drifts off-distribution. Aggressive sampling (high temperature, high top-p) breaks it first; conservative sampling preserves it at the cost of diversity.