What's this about?
BDH-CQ is a reasoning model that combines in-context learning with a recurrent latent memory that updates continuously as it reads input at inference time — so new information keeps reshaping the model's internal state rather than just sitting in a fixed context window. It was the runaway top paper on HuggingFace today (249 upvotes).
What exactly is "recurrent latent reasoning"?
Standard transformers process a fixed chunk of text at once and "reason" only within that window. A recurrent latent approach maintains a running internal state — like a working memory — that it updates step by step as more input arrives. "Latent" means the reasoning happens in the model's internal representations, not as visible text. It's an attempt to give models a more persistent, evolving train of thought.
Why it matters
Continuously-updating internal memory is a promising route past the fixed-context-window limits of today's LLMs — relevant to long tasks, streaming input, and agents that need to keep reasoning over time.