Google has introduced a new artificial intelligence (AI) system designed to help scientists generate novel hypotheses and research proposals, the company announced in a Google Blog post published on Feb. 19, 2025. The new system, AI co-scientist, is a multi-agent AI system that scientists can use to navigate information and insights from scientific publications and other resources.
Built on Gemini 2.0, it is intended to be used as a collaborative tool by scientists and was designed to mirror the reasoning process of the scientific method. The tool can “uncover new, original knowledge [and] formulate demonstrably novel research hypotheses and proposals, building upon prior evidence and tailored to specific research objectives,” according toJuraj Gottweis, Google Fellow, and Vivek Natarajan, Research Lead (1).
AI co-scientist uses a combination of Generation, Reflection, Ranking, Evolution, Proximity, and Meta-review that use automated feedback to iteratively generate, evaluate, and refine hypotheses, resulting in a self-improving cycle of increasingly high-quality and novel outputs. It also uses web search and specialized AI models to enhance the quality of the hypotheses. Scientists can interact with the system by directly providing seed ideas or feedback on generated outputs.
The tool analyzes the assigned goal into a research plan configuration that is managed by a supervisor agent that assigns specialized agents and allocates resources. This enables the system to flexibly scale compute and to iteratively improve its scientific reasoning towards the specified research goal. It then leverages test-time compute scaling to reason, resolve, and improve outputs. It uses self-play-based scientific debate to create novel hypothesis and ranking tournaments for hypothesis comparison. The system self-improves by using the Elo auto-evaluation metric.
The Google Core Research Lead Investigators in their core role assessed whether higher Elo ratings correlate with higher output quality. Seven human domain experts curated 15 open research goals and best guess solutions in their field of expertise. Using the automated Elo metric it was observed that the human guided assistance outperformed state-of-the-art agentic and reasoning models for these complex problems. However, Google claims that as the AI system spent more time reasoning and improving, the self-rated quality of results improved for AI systems using unassisted human assistance.
|