How much a model can think in silence

One safety hope for today's models is almost embarrassingly simple: if a model has to reason its way through a hard problem, it has to think out loud, writing the steps into the chain of thought where a monitor can read them. The worry is that a clever model might do the whole thing silently, in the arithmetic of its own activations, and only write down an innocent-looking answer. A DeepMind safety team has now turned that worry into a measurable quantity they call opaque serial depth: the length of the longest chain of sequential computation a network can perform without spilling a single intermediate step into visible text.

The architecture that hides less is the one you can trust to show its work.

The useful surprise is how it scales. For a standard Transformer, this hidden-thinking budget grows only with the model's layer count and barely at all — logarithmically — with the length of the prompt. Feeding it a longer context, in other words, buys almost no extra room to reason in secret; past some point the model has to externalize, and the reasoning surfaces in the chain of thought. The team also ships an open-source tool that computes this ceiling for any network, and it turns intuition inside out: a mixture-of-experts model, often assumed the more capable design, scores a lower opaque depth than a comparable dense one, because routing between experts cuts the serial dependencies.

The catch, which the authors are careful about, is that this is a ceiling, not a reading: it bounds what a model could hide, not what any real model does — an actual Transformer may sit far below the line. But the sharper point is what happens when you change the architecture. Swap the Transformer for a recurrent design and the same tool shows the hidden budget growing linearly with length, wide open. A follow-up applying the tool to a diffusion language model found it could naively hide vastly more than the equivalent autoregressive one. So the monitorability that safety cases lean on turns out to be a property of the architecture, not of the model's good behavior — and the efficient, recurrent, latent-reasoning designs several labs are chasing are exactly the ones that would quietly erase it.