Latent Reasoning
Latent Reasoning is an LLM Reasoning paradigm where the model performs computations within its internal hidden space, as opposed to reasoning at the token level in Chain-of-Thought Reasoning.
One example is Coconut (Chain of Continuous Thought). Instead of decoding the last hidden state into a token, the model feeds it straight back in as the next input embedding, so it can reason without being constrained to words (Hao et al., 2026).
It's worth mentioning that the major distinguishing factor between the total doom scenario and not in AI 2027 is our ability to interpret the reasoning of AI models. So I guess, as with everything in AI in 2025, we'll keep exploring this paradigm at our own risk.
References
Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. Training Large Language Models to Reason in a Continuous Latent Space. August 2026. arXiv:2412.06769, doi:10.48550/arXiv.2412.06769. ↩