Google DeepMind published Co-Scientist in Nature, a multi-agent AI system built with Gemini that generates and refines scientific hypotheses through structured debate and iteration.

The system addresses what the company calls a critical bottleneck in scientific discovery: the challenge of connecting disparate facts to formulate breakthrough hypotheses in an era of information overload.

Co-Scientist operates through six specialized agents organized into three phases. The generation and proximity agents propose initial hypotheses and map research spaces. The reflection and ranking agents act as virtual peer reviewers, critically evaluating ideas through simulated scientific debates. The evolution and meta-review agents continuously refine top-ranked hypotheses and synthesize insights.

A supervisor agent coordinates the coalition, breaking down research goals and enabling parallel exploration of multiple research avenues simultaneously.

Tournament of ideas validates hypotheses

The system's core innovation is its "tournament of ideas" — an approach inspired by AlphaGo that uses scientific debates rather than game-playing to rank hypotheses. The majority of computational resources focus on verification, cross-checking claims against scientific literature and specialized databases including ChEMBL and UniProt.

Google DeepMind is making Co-Scientist available through Hypothesis Generation, an experimental tool developed across Google DeepMind, Google Research, Google Cloud and Google Labs. Researchers can register interest at labs.google/science, with rollout beginning in the coming weeks.

The system has already been tested on challenging problems including antimicrobial resistance, plant immunity and liver fibrosis. Early research collaborations are exploring integration with specialized models like AlphaFold for enhanced capabilities.

"We believe AI can help dramatically accelerate the pace of breakthroughs by serving as a dedicated partner in the generation and refinement of breakthrough scientific hypotheses," the Co-Scientist team said.

The company positions Co-Scientist as one of the first reliable multi-agent systems for structured scientific thinking, designed to deliver tangible results in novel hypothesis generation for complex scientific problems.