Recursive latent-space reasoning unlocks out-of-distribution generalization without chain-of-thought tokens

A new architectural approach for transformers performs reasoning recursively in latent space rather than externalizing it as chain-of-thought tokens. The method achieves robust algorithmic generalization on out-of-distribution tasks where standard transformers fail — and provides mechanistic interpretability analysis to characterize where the reasoning happens internally.

The result challenges the chain-of-thought orthodoxy. Since 2023, the dominant view has been that letting models think in tokens (verbose externalized reasoning) is the path to robust reasoning. Latent-space recursion suggests that for some task classes, internal recursion is more capable than external token generation — and harder to red-team, since the reasoning is opaque to the user.

The safety implication is real. If latent reasoning beats CoT on capability, labs face a tradeoff between capability and inspectability — exactly the kind of architectural decision the responsible-scaling framework is supposed to govern. See our analysis → on the pass@k efficiency frontier.

arXiv — recursive latent space reasoning → · arXiv — expressive power of chain of thought →