Latent reasoning models might be a good thing?

Latent reasoning models (LRMs) are an emerging AI paradigm that performs chain-of-thought reasoning entirely in the model's latent space, bypassing the language model head that maps to token distributions. Instead, they feed activations directly back into the model as input embeddings for the next token position. This approach, popularized by Meta's Coconut paper and improved upon by CODI, compresses entire thoughts into single tokens rather than spreading them across multiple text tokens. Currently, the best public LRMs are only at GPT-2 scale and specialized for narrow tasks, but the author argues they could be a game-changer for AI safety and interpretability at scale.

The core argument is that as models scale to transformative AI, text-based chain-of-thought reasoning becomes increasingly unreliable for safety monitoring. Because the same weights are used for both thinking and output, optimizing output text can indirectly corrupt the reasoning process. In contrast, LRMs' compressed thought tokens are more isolated from output optimization and potentially easier to interpret—each token represents a complete thought rather than a fragment of neuralese spread across multiple tokens. While polysemantic tokens (encoding multiple ideas) remain a concern, the author suggests that interpretability tools like sparse autoencoders could extract more information from a single compressed thought token than from several tokens of neuralese. The main caveat: this argument assumes LRMs can be scaled effectively and that compressed tokens are indeed more interpretable than distributed ones—both of which remain unproven at scale.