Domain Agnostic Features]{\centering Domain Agnostic Features for Robust Novelty Assessment

Assessing the degree of novelty in scientific publications remains a challenging task in research evaluation. While large language models enable fine-grained semantic analysis of texts, reliable and architecture independent methods for the quantitative assessment of novelty are still under development. This study refines previously introduced quantitative metrics for measuring novelty in scientific articles and aims to identify a stable set of domain-agnostic characteristics that can support consistent novelty classification. Based on transformer-derived signals, we extract a core set of features designed to capture semantic deviation in scientific abstracts. Using this feature set, two classifiers drawn from fundamentally different model families are trained and evaluated in order to examine whether the identified characteristics remain informative regardless of the underlying classification approach. The results show that the proposed feature set consistently captures meaningful signals of novelty. Classifiers trained on these features demonstrate comparable performance across model families, indicating that the extracted semantic patterns reflect properties of the data rather than artifacts of a specific architecture. The proposed analytical tool can support human reviewers in assessing the likelihood that a manuscript contains substantial novelty, while not intended to replace expert judgment. Further improvements in classification performance may be achieved through a broader exploration of hyperparameter configurations and by increasing the dimensionality of the feature representation.