Writing reasoning into code comments may help models stretch their thinking budget

mertdumenci · x · 2026-07-24

The post points to a useful failure mode: letting a model insert reasoning directly into the output can help it spend more of its available “thinking budget” in a way that may not be captured by RL reward.

It also notes that Claude’s habit of narrating the session in code comments may be the same kind of behavior—an output style that leaks internal reasoning into the code path instead of keeping it separate.

Related event: Claude's Habit of Narrating in Code Comments Linked to Reasoning Flaws(2 posts)→

Original post →

More from coding & agent

coding & agent channel →