CoWeaver matches humans and AI agents for stronger scientific collaboration

LLM-based agents excel at writing, coding, and retrieval but fail to form strong collaborations in scientific communities due to the bidirectional, dynamic nature of matching and the need for interpretability. CoWeaver, proposed by researchers from multiple institutions, solves this with a learnable algorithm that matches scientists and agents by filling capability gaps. It uses a two-stage ranking to filter candidates and maintains uncertainty-aware capability estimates to explore newcomers, combining UCB exploration with greedy selection.

In experiments, CoWeaver's hybrid exploration-greedy mechanism outperformed the pure greedy approach on 6 out of 20 tasks and performed on par on the rest. It beat all baselines on matching quality and efficiency. By enabling explainable, dynamic team formation, CoWeaver could unlock more productive human-AI collaboration in research, helping teams capitalize on complementary skills while remaining transparent.