A new arXiv paper, INTACT, proposes a way to get latent reasoning from a frozen language model. Instead of generating natural-language chains of thought, the method uses a small external encoder to produce compact continuous "soft-token" representations. These replace the verbose intermediate steps that models normally decode one token at a time.

The key benefit is efficiency: because the soft tokens are compact and continuous, the model needs to decode far fewer intermediate tokens autoregressively. The approach is designed to work on a frozen backbone, meaning the underlying language model does not need to be fine-tuned or modified.

The abstract is brief, so details on how the external encoder is trained or how the soft tokens are integrated are not yet available. But the core idea is clear: shift reasoning into a compressed latent space to save computation while keeping the base model intact.