Whether a reasoning model can hide computation from a chain-of-thought (CoT) monitor depends on task difficulty and model size. The work combines theory with experiments to show that simple hidden computations leave no trace, while beyond a difficulty threshold that depends on model size, solving the task forces a near-linear amount of information about the hidden input into the CoT.
That footprint is not necessarily useful to a monitor. Under plausible cryptographic assumptions, even a one-layer Transformer can encrypt its reasoning online, so no polynomial-time monitor can extract the hidden information. The upshot is a mixed verdict for CoT monitoring: hard tasks cannot hide computation in principle, but that does not guarantee a readable trace.