There is something almost magical about the way human beings can solve a problem without explaining every step. You might look at a tangled knot of traffic and instantly know the best route, or read a sentence in your native language and understand it without pausing to diagram the grammar. For a long time, researchers assumed that AI systems needed the same kind of step-by-step self-talk to reason well — a technique known as chain-of-thought reasoning. The idea is intuitive enough: if you want a model to answer a tricky question, you ask it to “show its work,” generating a long trail of intermediate words before it lands on a final answer. And this approach has worked, helping AI models crack puzzles, solve math problems, and follow complex instructions. But it comes with a quiet cost. Every word an AI produces, every so-called token, takes real computing power. The more evidence the model has to spell out, the slower and more expensive the answer becomes. It is a bit like insisting that someone write a full essay every time they need to decide whether to take an umbrella.
A new experimental system, however, suggests that AI may not always need to think out loud. The system, called BDH-CQ — the DH stands for Dragon Hatchling — offers a different way of reasoning, one that keeps its intermediate thinking mostly hidden. Instead of spelling out each step of its logic, the model works through a problem internally, using a small, fixed-size memory that it updates as it goes. That may sound technical, but the intuition is familiar. Imagine you are learning a new recipe. One way to cook is to read the instructions aloud again and again, narrating each chop and stir as you go. Another way is to internalize the technique so well that you just know, somehow, when the sauce is ready — without saying a word. BDH-CQ does something closer to the second approach. It was built to solve visual reasoning puzzles in the style of the Abstraction and Reasoning Corpus, the kind of tasks that involve spotting patterns in colored grids and applying abstract rules to new arrangements. The model does not need to generate a long chain of words before it gives its answer. It just, in a sense, thinks.
This might sound almost like a magic trick, but the researchers who built it — led by a team that includes scientist Stamirowska — wanted to test whether all that language was actually necessary. When most AI models are trained, they are shown examples of how to solve problems, often with the reasoning written out. Those examples help the model learn to follow a similar pattern when it encounters something new. But BDH-CQ does things differently. Instead of dumping each new example into an endlessly growing context, the model uses each example to quietly update a compressed memory. That memory stays the same size no matter how many examples it has seen. This may not sound revolutionary, but it means the model is not constantly looking backward, trying to compare the current problem against a huge pile of remembered examples. It is learning to capture the essence of a problem in a compact form, and then applying that understanding to new cases. In tests, the model was able to solve some new reasoning puzzles without ever writing out its intermediate steps. It handled tasks that involved rotating or moving shapes, correctly recognizing when a pattern had been shifted and transformed. It was less confident when dealing with color changes or with combinations of rules, and it sometimes stumbled when multiple rules had to be layered on top of each other. But for a small model, it showed a surprisingly nimble kind of intelligence.
The potential payoff here is about more than just novelty. In the world of artificial intelligence, words are expensive. Each token a model generates costs electricity, time, and money, and that cost scales with the length of the reasoning process. A model that can solve a puzzle with a short answer, rather than a sprawling chain of thought, could be significantly cheaper to run. It could also be faster, responding to challenging questions without making users wait while it narrates its way to a conclusion. The researchers believe that BDH-CQ’s approach of reasoning internally could help keep computing costs down, especially for problems where the intermediate steps are not all that interesting to spell out. Think about how you solve a simple puzzle like a Sudoku square. You might not announce every possibility you considered and discarded; you just see the answer, or at least the next move. The same kind of unspoken reasoning could be happening inside a model like BDH-CQ. It is not thinking in words so much as it is thinking in patterns, transformations, and abstract relationships. This challenges the assumption that language is the natural medium for all machine reasoning. Just because a model can explain something in words does not mean it needs to use words to compute it — and, more importantly, a model that stays quiet may still be doing something genuinely thoughtful.
Of course, this quieter kind of reasoning has its own set of problems. For one thing, it is harder to inspect. When an AI model writes out its chain of thought, researchers and users can peer inside and try to understand why it made a particular decision. They can trace the logic, spot errors, and perhaps even correct the model when it goes wrong. With BDH-CQ, that window is closed. The model arrives at an answer without leaving behind a trail of tokens, which means there is no easy way to see what it was thinking. It becomes a kind of black box, and black boxes can be unsettling, especially in high-stakes situations. But someone like Deng, a researcher involved in the study, points out that written-down reasoning may not be as transparent as it seems. A model’s chain of thought might not faithfully reflect the actual steps it took to reach an answer; it could be a post-hoc story, invented after the fact to sound plausible. Human beings do this all the time. We often make decisions based on intuition or emotion and then construct a rational-sounding explanation for why we chose what we did. There is no guarantee an AI is being any more honest when it writes out a chain of reasoning. So while BDH-CQ’s silence makes it harder to audit, the written chains of other models may offer only an illusion of understanding.
There are also limits to how far this specific system can be taken. BDH-CQ was built specifically for ARC-style problems, which are a narrow slice of the many kinds of reasoning AI might be asked to do. It is not a general-purpose model that can write an essay, chat about history, or plan a vacation. Comparing it to large, broad-based systems is a bit like comparing a sports car engine to a commercial airliner; both involve machinery, but they are built for different jobs. That is one reason researchers like Jonas Geiping at the ELLIS Institute Tübingen and the Max Planck Institute, while calling the approach “neat,” are cautious about drawing too many conclusions. He noted that the model can take on test problems without retraining, which is genuinely impressive, but more work is needed to separate the effects of the model’s design from the ways it was trained. Could any model learn to reason more quietly if it were given the right training regimen? Or is BDH-CQ successful because of the particular tricks that went into its architecture? These questions are still open. And yet, even with these caveats, the study points toward a deeper and more philosophical question: Does AI really need to put all of its reasoning into words?
Human beings have long treated language as the universal currency of thought. If you can’t say it, the saying goes, you don’t know it. But that is not really true. We often know far more than we can put into words. We can recognize a friend’s face, ride a bicycle, or sense that a joke is funny, all without being able to explain the underlying computations. Many of our best ideas arrive fully formed, as if from nowhere, and it is only afterward that we shape them into sentences. Language, as Deng puts it, is useful for communicating reasoning, but it does not need to be the most efficient representation for every intermediate computation. BDH-CQ offers a glimpse of what that might look like inside an AI. It is not a fully formed alternative to chain-of-thought reasoning, and it may never replace the kind of explicit, step-by-step thinking that helps models handle complex tasks. But it suggests that not all intelligence needs to be verbal. There might be a whole spectrum of ways to think, from the loud, sequential narration of problem-solving to the quiet, compressed flashes of insight that arrive in a single moment. Perhaps the best AI systems of the future will learn to move back and forth along that spectrum, using words when they help and staying silent when they do not. For now, BDH-CQ is a reminder that sometimes the most powerful thing a mind can do is think without saying a word.
