AI Scholar: Learning the Scientific Taste of Elite Researchers to Predict Future Discovery Trajectories

Abstract

As artificial intelligence transitions from an analytical tool to an autonomous agent of scientific discovery, a critical gap remains: current generative ideation systems produce generic, domain-level proposals that lack the idiosyncratic intuition, methodological preferences, and risk tolerance of human experts. We propose \emph{AI Scholar}, a novel computational framework designed to operationalize and learn ``scientific taste.'' By ingesting the lifetime bibliometric trajectories of 10,000 elite, independent principal investigators, our system constructs temporal latent representations of individual research styles. Unlike existing evaluation benchmarks, AI Scholar utilizes these taste embeddings to condition a multi-agent generative pipeline, autonomously predicting and synthesizing the precise future research proposals these scientists are likely to pursue. This paper outlines the theoretical foundations, the novel Trajectory-Conditioned Ideation methodology, and the time-partitioned evaluation design required to validate this massive-scale predictive framework, ultimately exploring the scaling laws of personalized scientific discovery.

Keywords