CoWeaver Matches Humans and AI for Scientific Collaboration

Large Language Model (LLM)-based agents demonstrate proficiency in tasks like article writing, coding, and information retrieval. However, their ability to form robust collaborations within the scientific community is limited by the dynamic, bidirectional nature of such partnerships and the high demand for decision interpretability. To address this, a new algorithm called CoWeaver has been proposed. CoWeaver is a bidirectional, learnable, and explainable matching engine designed to facilitate strong collaborations between human scientists and AI agents. Its core mechanism involves identifying and filling capability gaps between candidates and requesters. The system employs a two-stage ranking process to filter potential collaborators effectively. Furthermore, CoWeaver incorporates an exploration strategy for newcomers by maintaining uncertainty-aware capability estimates, which are updated based on requester feedback. Comparative analysis shows that CoWeaver's combined exploration (UCB) and greedy selection mechanism outperforms a purely greedy approach in 6 out of 20 tasks and performs comparably in others, demonstrating its superior matching quality and efficiency against baselines.