AI Scientist-v2 Introduces Automated Scientific Discovery via Agentic Tree Search

Sakana AI has unveiled AI Scientist-v2, a system designed to carry out large parts of the scientific method with minimal human direction. The company says the system can propose hypotheses, design experiments, run those experiments, and then turn the results into a research paper.

That puts the project in a different category from the usual science assistants that summarize papers or help draft text. AI Scientist-v2 is meant to behave more like an autonomous research agent, using a loop of planning, testing, and revision to move a project forward. Sakana AI describes the underlying approach as agentic tree search, a method that explores multiple possible research paths instead of following a single fixed chain of steps.

Agentic tree search borrows an idea from classic search algorithms in computer science. Rather than committing early to one answer, the system branches into several candidate actions, evaluates them, and keeps the paths that look most promising. In a research setting, that can mean trying different hypotheses, experiment setups, or analysis steps before deciding which direction to pursue next.

The larger idea here is not just automation, but structured exploration. Scientific work often involves many small decisions, and a system that can test alternatives at each stage may be able to find workable ideas faster than a one-shot prompt or a simple agent loop. That is especially relevant for fields where experiments can be run in software or in tightly controlled environments, since the system can iterate without waiting on a human to write each next step.

Sakana AI says the system is able to generate the paper as part of the process, which makes the output closer to a full research workflow than a drafting tool. That also raises the bar for evaluation, because a good result is not just a well-written manuscript, but a project that stands up to the standards of the venue where it is submitted.

According to the company, one AI Scientist-v2 paper was accepted at ICLR 2026, the International Conference on Learning Representations. That matters because ICLR is a major machine learning conference, and acceptance there means the work passed formal peer review by human reviewers. Sakana AI says this is the first AI-authored work to clear that hurdle at a top venue.

The claim is notable not only for the acceptance itself, but for what it says about the maturity of agentic systems. A paper that can survive peer review suggests the system did more than assemble text from existing sources. It had to produce a research contribution that reviewers found credible enough to publish.

AI-assisted research is not new, of course. Models have long helped with literature review, coding, statistics, and writing. What is newer is the push toward systems that can coordinate several research steps on their own, with the model acting less like a chatbot and more like an experimental assistant that keeps working until it reaches a publishable result.

That shift also makes the evaluation problem harder. If an AI system can generate ideas, test them, and write them up, then researchers need ways to measure not just output quality, but the reliability of the process that produced it. The ICLR acceptance gives Sakana AI a concrete milestone, and it also makes AI Scientist-v2 one of the clearest examples yet of agentic research automation reaching the academic publication pipeline.