Stanford researchers have built an AI agent that thinks in silence and switches back to words for key decisions, web searches and code. Across 11 tests, it beat rival agents’ scores by 10.4% on average. Against agents that think in words, it ran about twice as fast and cost up to 3.7 times less.
Most AI agents think out loud. They work through a task in steps, searching the web or running code along the way. Before each action, an agent types out every thought, word by word. It is like solving a sum while saying each step aloud: carry the two, add the four. Each written word costs time and computing power.
Earlier AI models that thought silently moved faster but could not use tools, and people could not easily read their thinking.
Stanford built the agent, HybridAgent, with Together AI, Meta and the University of Illinois Urbana-Champaign, and gave it two modes. In silent mode, each step stays inside the model as a string of numbers. That string feeds straight into the next step without becoming a word, the way a person does a sum in their head. In text mode, the agent writes as usual, and at every step it decides whether to stay in its current mode or switch.
The final round of training rewarded correct answers, proper formatting and use of search and code tools, but offered no reward for going silent. Even so, the largest version moved more and more of its work into silence while getting better at using tools. On its own, it saved words for tool use and the key steps people want to check.
People can still read the silent steps, by matching each one to the closest word in the model’s vocabulary. The result looks messy at times but follows what the model is doing.
Silent thinking also held up on hard tasks. On GAIA, a test of real-world assistant tasks, the largest version scored 52.4, while Kimi K3, a larger model, scored 41.7. OpenAI’s GPT-5.6 Luna, also larger, scored 37.9, though both rivals stayed ahead on some other tests.
The paper is a preprint, a study shared before peer review, and its models are small. The gains also depend on hidden steps that stay clean. When the researchers mixed random noise into those steps, the largest version’s score on a math test fell from 88.2 to 70.2.
