AI scientists are building multi-agent systems that assist, not replace, human researchers, with Nature papers introducing Robin and Co-Scientist as collaborative tools designed to accelerate discovery through human-AI teamwork.

These systems use multi-agent setups to aid drug repurposing and disease research, while still requiring human input at various stages to validate and guide findings.

The broader aim is to ground AI reasoning in human-driven goals and supervision, as large language models help navigate literature and data but cannot conduct research independently.

A tournament-of-ideas approach drives hypothesis verification and refinement, using debates and pairwise comparisons to rank hypotheses and cross-check claims against literature and data with supporting tools like web search, ChEMBL, UniProt, and AlphaFold.

Impact examples include a drug candidate blocking most of a fibrosis-driven response in lab tests, faster literature interpretation for aging research, and collaboration across top labs to generate ideas.

Automated evaluations show that increasing test-time compute consistently improves the quality of generated hypotheses over time, indicating AI-augmented science can accelerate workflows.

In drug repurposing, Co-Scientist proposed dozens of candidates for AML, with human experts narrowing to a few and several showing positive results in early tests, underscoring the need for validation.

In another AML-focused study, Co-Scientist helped identify approved drugs that could be repurposed, with researchers selecting several candidates and some showing promise in cell-based assays.

Co-Scientist uses specialized agents and an Elo-rating-like self-assessment to rank hypotheses, having suggested multiple cancer drug candidates for AML and guiding lab testing toward promising options.

Practical applications include using LiDAR to look around corners, as explored in related research.

Additional highlights cover diverse topics like tardigrade resilience and pristine Antarctic ice records relevant to Solar System history.

Key contributions center on a scalable multi-agent architecture and a tournament evolution process that fosters self-improving hypothesis generation.