The research presents EvoTree, a staged framework designed to generate evolution trees from citation graphs. The initial stage utilizes a graph-aware encoder and distribution-based hierarchical clustering to establish a stable taxonomy backbone. Subsequent temporal fine-tuning then attaches marginal papers to the backbone, adhering to monotonic-path constraints. A final pass employs a large language model to label concepts, maintaining the established topology. The framework was evaluated on an annotated benchmark across 11 AI subfields. EvoTree demonstrated the highest NMI and citation-direction accuracy compared to baseline methods. Furthermore, it uniquely identified marginal papers and achieved the best concept purity on the benchmark.
Source: https://arxiv.org/abs/2609.09561