Hidden messaging is easier to learn than reasoning through hidden text

A model that hides a number in ordinary-looking text has not necessarily learned to reason through that hidden representation. Learning Steganography Is Easy, Learning Steganographic Reasoning Is Hard tests that distinction. The researchers compare concealed messaging, openly illegible reasoning and reasoning embedded inside apparently unrelated text.

Three capabilities that should not be conflated

Steganographic messaging encodes information already given to the model. Encoded reasoning performs a computation in an illegible format without trying to appear innocuous. Steganographic reasoning combines computation with concealment: the model must use the hidden trace as a scratchpad while producing cover text.
The study deliberately elicits these behaviors in the model's output, with its ordinary chain of thought disabled. That removes limited control over the normal reasoning channel as a confound. It also means the experiments are not observations of naturally occurring concealment inside a production model's reasoning trace.
The concern shares an oversight question with opaque-reasoning coverage, but the mechanisms differ. Latent computation can leave fewer readable traces; this paper asks whether readable-looking output can carry a different computation. Neither case permits a monitor to assume that visible text fully explains the work producing an answer.

Most cover tasks make concealed computation difficult

The experiments use in-context learning, reinforcement learning and supervised fine-tuning. Across most task-agnostic schemes, models acquire messaging and encoded reasoning more readily than steganographic reasoning.
Successful supervised-fine-tuning runs require 2.4 to 20 times as many steps for concealed reasoning as for messaging. Several model–task combinations never reach the target within the tested budgets. In-context and reinforcement-learning experiments also fail to elicit the behavior reliably under the task-agnostic schemes.
These failures do not show impossibility. They describe specific models, tasks, encoding schemes and training budgets. The authors' explanation that simultaneous cover-text production and hidden computation strain the same forward pass remains a hypothesis, not an experimentally isolated cause.

A convenient cover task changes the result

A repetitive knapsack task leaves arbitrary choices that can carry information without conflicting with the main answer. In that setting, the researchers elicit steganographic reasoning under all three methods.
They also distinguish writing correct intermediate values from actually computing through the trace. Side-task accuracy must exceed the model's no-chain-of-thought baseline; otherwise it may have computed internally and merely written down the result. Corrupting an intermediate encoded value provides another check on whether later computation uses it.
For oversight, the difficult cases and the successful exception belong together. Evidence that a model can conceal a message is insufficient evidence that it can perform hidden multi-step reasoning. Conversely, failure on one cover task does not rule out success where the surface format makes encoding easier.
Covertness depends on a particular model monitor and suspiciousness threshold. It is not demonstrated invisibility to every reviewer. The study narrows a monitoring risk into capabilities that can be tested separately, while leaving open how often comparable behavior would arise under ordinary deployment training.