CORE

Term definitionscoming soon

Definitions for this piece will land here. Inline terms in the body will link into this panel once they are filled in.

/** Fill in when the case-study rewrite lands. Used by ArticlePage Term definitions. */ export const glossary = [];

Everyone has an opinion about whether comments help a language model write better code. Almost all of that opinion is formed by reading the model's output and deciding whether it looks smarter. That is a measurement of prose, not of reasoning. CORE is a second research paper, in preparation with the Gradient PG AI/ML science club, that tries to answer the question mechanistically.

The question

Given the same function with and without comments, does the model take a different internal route to its answer — or does it arrive the same way and merely narrate differently?

How we probe it

Three angles, none of which depend on judging output quality:

• Logit masking. Suppress the comment tokens' contribution and watch what moves. If the reasoning genuinely depends on them, the effect should show up well before the final token.
• Hidden state analysis. Compare layer-by-layer representations of the commented and uncommented versions, looking for where the two trajectories diverge rather than whether they end in the same place.
• Weight inspection. Which heads and circuits change their attribution when a comment is present.

Why it matters

PLACEHOLDER — the practical version of the finding. If comments do real work inside the model, that changes how we should write code destined for AI review, and how much we should trust a model that has only ever seen stripped source.

Status

PLACEHOLDER — paper in preparation. Update with the venue and the headline result once you have them.