Redwood Research proposes a metric for hidden reasoning

Redwood Research proposed a concrete way to measure how much of a model’s reasoning happens outside its written chain of thought — where it cannot be read or audited. The metric, which the authors call NLS (“natural-language-rooted node-separated”) depth, tracks the longest chain of computation a model can carry out without passing through a step that outputs discrete, human-readable tokens. A standard transformer that reasons step by step in plain text has low NLS depth by construction; an architecture built around looped or latent internal states, where reasoning proceeds silently and only a token is emitted at the end, would score far higher.

The authors — including Ryan Greenblatt, Nathan Sheffield, Alek Westover, Lukas Finnveden, Alexa Pan and Julian Stastny — argued NLS depth should function as a governance tool: something labs could compute and report before or during training, giving outside observers advance warning of an architectural shift toward less monitorable reasoning, rather than discovering the shift only once a released model’s chain of thought stopped reliably reflecting its actual computation. They wrote that “even some fairly small (hypothetical) latent reasoning architectures would have much larger NLS depth than the largest open-source CoT models.”

The paper was explicit about its trigger: it named OpenAI’s GPT-6 Astra, whose “recurrent depth” technique had drawn criticism for reducing how much of the model’s reasoning surfaces in an inspectable chain of thought, as an example of the kind of architectural change the metric is meant to catch early. It followed within days of OpenAI chief scientist Jakub Pachocki’s own essay warning that no lab had solved AI alignment and arguing for preserving chain-of-thought monitorability as capabilities scale — putting an independent alignment-research group and a frontier lab’s own leadership on record within the same week about the same narrowing channel for oversight, though neither has proposed a binding mechanism to require the reporting the paper calls for.