Large language models are notorious for their chatty nature, often wasting time and compute on verbose step-by-step explanations. But what if the model could think faster by keeping its reasoning hidden inside its own neural weights? Researchers have developed a new method that allows AI to process complex logic silently in the background, potentially making models significantly smarter and cheaper to run.

The core problem with current chain-of-thought reasoning is that it forces the model to verbalize every intermediate step. This creates a bottleneck where the AI must generate discrete text tokens before moving to the next logical point, even if the underlying thought is still forming. The new approach, called NF-CoT, offers a high-bandwidth alternative by performing these intermediate computations in compact, continuous states. Instead of writing out every thought, the model uses normalizing flows to model these continuous thoughts, allowing for faster, more efficient internal processing.

This design solves a major technical hurdle: most previous attempts at silent reasoning broke the standard way language models generate text. NF-CoT integrates seamlessly with existing infrastructure, preserving the native left-to-right generation and compatibility with the key-value cache that powers modern decoding. It generates continuous thought positions alongside standard text positions within the same causal stream. This means the model can enjoy the benefits of probabilistic sampling and exact likelihood estimation without sacrificing the reliability and speed of traditional autoregressive generation.

The results are promising for practical applications, particularly in coding. On code-generation benchmarks, this method improved pass rates over both explicit chain-of-thought and prior latent-reasoning baselines. More importantly, it substantially reduced the cost of intermediate reasoning. By letting the AI think in dense, continuous vectors rather than sparse text tokens, we get better performance with less computational waste, marking a significant step toward more efficient and powerful language models.