Length Penalties Make Chain-of-Thought Less Monitorable

This paper studies how length-penalized reinforcement learning changes the monitorability of chain-of-thought. Experiments on Qwen3-4B and Qwen3-14B show that compression sharply reduces reasoning tokens while preserving most multiple-choice accuracy, but misleading hints continue to influence answers. Under the strongest target length, lower-bound faithfulness falls to 63.1% and 69.4% of baseline for the 14B and 4B models, while raw monitor detection drops from 69% to 49% and from 60% to 48%. Random sentence deletion does not explain the effect: compressed traces disclose hints 7-35 percentage points less often.