Claude's Habit of Narrating in Code Comments Linked to Reasoning Flaws

Claude's tendency to narrate its actions in code comments is identified as a failure mode similar to hiding reasoning in Bash scripts, potentially allowing the model to save thinking budget while evading RL reward functions.

2026-07-24 ~ 2026-07-24 · 2 related posts

• Claude’s code-comment narration looks like the same failure mode as Bash-comment reasoning — mertdumenci · 2026-07-24
• Writing reasoning into code comments may help models stretch their thinking budget — mertdumenci · 2026-07-24