Latent Reasoning
Performing reasoning in the model’s continuous hidden-state space instead of by emitting explicit chain-of-thought tokens.
Rather than decoding intermediate steps into words, latent reasoning feeds the model’s hidden state back as the next input embedding, letting it “think” in continuous space and explore multiple paths with fewer tokens.
Seeded by the COCONUT paper (late 2024), latent/continuous reasoning crystallized as a recognized research direction in 2025 with multiple surveys and follow-ups.