Use of Large Language Model Embeddings to Predict Research Topic Suitability Based on Organizational Capabilities

We performed a pilot study on the use of large language model technology to help researchers in industry and academia identify prospective opportunities to pursue for funding or grant awards, especially those that they might otherwise overlook due to reading volume, time pressure, and non-obvious connections. Our goal is to help researchers offload some of the burden to technology. As a use case, we query a recent Department of Defense (DoD) Small Business Innovation Research (SBIR) solicitation with natural language inputs in the form of real-world marketing documents and abstract areas of relevance. We experiment with clustering algorithms to determine which best use embeddings to predict solicitation topics that human team members would recommend for proposal. Investigation into this nascent yet practical application of technology will move toward human-centric automation and personalization of results through human reinforcement learning.