Structure activates higher-quality training distributions

Type: note · Status: seedling

LLMs are autoregressive — they produce text that continues the pattern in context. When the context contains sections like ## Evidence and ## Reasoning, the model's output will resemble the training data that had similar structure: scientific papers, legal analyses, peer-reviewed arguments. These documents are, on average, higher quality for reasoning purposes than the bulk of internet text.

The structure acts as a distribution selector. A free-form prompt might draw from blog posts, forum comments, or opinion pieces. A Toulmin-shaped template steers the model toward the subset of its training data where authors were already doing rigorous argumentation. We assume — reasonably — that scientific papers and formal arguments have better epistemic value for our purposes than unstructured web text.

This argument is independent of failure-mode transfer. Even if LLMs had no human-like failure modes at all, the distribution-selection effect would still apply: structured context activates structured training data, which tends to be higher quality. And it's independent of readability for humans — the quality improvement happens in the generation process itself, before any human reads the output.


Relevant Notes:

Topics: