Length Penalties Make Chain-of-Thought Less Monitorable

Bryce Little

To curb overthinking and reduce inference costs, researchers now train reasoning models with penalties on chain of thought length. We find that these penalties degrade monitorabili…