Training reasoning models with chain-of-thought length penalties reduces chain-of-thought monitorability.

The claim holds that penalizing chain-of-thought length during reinforcement-learning training of reasoning models makes their reasoning traces less useful to monitors. It is a plausible mechanism: pressure toward brevity can push models to drop intermediate steps, compress reasoning into terse or idiosyncratic shorthand, or move computation out of the visible trace, any of which would leave monitors less to read. The 2025 multi-lab position paper on chain-of-thought monitorability (Korbak et al.) names optimization pressures on the reasoning trace, including length penalties, as a possible threat to monitorability, and related work (Baker et al. 2025) shows that direct optimization pressure on chains of thought can produce obfuscated reasoning.

The direct evidence for length penalties specifically is thin. Most of the concern is stated as a risk rather than measured, and evidence on length penalties tends to show shorter or less legible traces rather than a measured drop in a monitor's ability to catch misbehavior. Shorter traces are not automatically less monitorable: a concise trace that still states the decisive steps can be just as informative. The effect probably depends on penalty strength, the model, the task, and how monitorability is measured. The claim as a general causal statement is therefore plausible but not established. A controlled study that trains matched models with and without length penalties and measures monitor detection rates for misbehavior would settle it.