Researchers let an AI agent think partly without words: faster and cheaper

Anyone who asks a modern AI chatbot a difficult question often sees it "think" first. The language model then writes out its reasoning step by step. It does this in tokens: small pieces of text, such as a short word or part of a longer word. The word "sunglasses", for example, can be cut into "sun" and "glasses". Writing it out helps: with intermediate steps the model makes fewer mistakes. But it also costs time and computing power, because each token has to be calculated separately.

Researchers from Stanford University, the University of Illinois and the AI company Together AI have found a way to skip part of that writing out. Their system is called HybridAgent. They published the work on 13 October 2026 on arXiv, a website where researchers share their papers before other experts have reviewed them. The code is publicly available on GitHub.

Thinking in numbers

Internally, a language model does not calculate with words but with long series of numbers. Only at the end of each step does it convert those numbers into a token. In so-called latent reasoning, the model skips that conversion. It passes its internal numbers directly on to the next thinking step. Compare it to doing mental arithmetic instead of putting every intermediate step on paper.

That idea already existed, but it had a drawback. An AI agent is an AI that carries out tasks itself, such as looking something up on the internet or using a calculation program. Such tools are controlled with text. A model that only thinks "in its head" cannot do that.

Switching between head and paper

HybridAgent combines the two ways. A small extra component, which the researchers call HybridLink, lets the model switch between silent thinking and writing. If it needs to look something up, use a tool or give an answer, it switches back to text. What it has worked out earlier, it keeps in a working memory, so that it keeps track from step to step.

The model learns this in two rounds. First it is given examples of how to switch. Then it learns through practice, with a reward for good answers, when silent thinking pays off and when it is better to write something out or take an action.

The silent thoughts are not completely invisible either. According to the researchers, they can be translated back into words, so that people can roughly follow what the model is doing.

The results

The researchers built HybridAgent on the open language models Qwen3 from the Chinese company Alibaba, in three sizes. They tested it on eleven tests: maths problems, questions about science and medicine, and tasks in which the agent has to search the internet.

According to the researchers, HybridAgent performs on average 10.4 per cent better than the systems they compared it with. It uses 35 to 55 per cent fewer tokens, is 1.8 to 2.6 times faster and is 2.7 to 3.7 times cheaper to use.

These are first results, measured by the researchers themselves. The cost saving is a calculation based on the computing costs per token. Because the code is public, other researchers can now rebuild and test the method themselves.

What does this mean for the road to the singularity?

AI agents are increasingly doing real work, such as searching for solutions to research problems in physics. How much they can do depends strongly on how much computing power and electricity each thinking step costs. An agent that can think twice as much with the same chips can try twice as many ideas.

Another encouraging point is that the method comes closer to our own thinking. People do not think everything out in words either, and only pick up pen and paper when it is needed. The fact that AI researchers can now build such a way of working into models shows how fast the techniques behind AI themselves are improving. Every step that makes AI more efficient brings smart help for science and everyday work a little closer.

What is the technological singularity? Read the explanation →