Generalizability of Abstract-Based Summarization Across Biomedical Datasets

Determine whether the performance pattern favoring abstract-only input over full-text input in biomedical lay summarization holds on the PLOS dataset, rather than being specific to the eLife dataset evaluated in prior work.

Background

Prior research compared abstract-only and full-text fine-tuning for biomedical lay summarization using the eLife dataset and reported trade-offs among relevance, readability, factuality, and inference time. Because that comparison was limited to a single eLife split, the paper identifies an unresolved question concerning whether the same performance pattern generalizes to the PLOS dataset.

Resolving this question would clarify whether conclusions about the relative effectiveness of abstract-based and full-text biomedical lay summarization are dataset-specific or apply across the two principal BioLaySumm datasets, which differ in article length, authorship, and lay-summary characteristics.

References

However, the comparison was limited to a single eLife split, so it remains unclear whether the same pattern holds on PLOS.

— Abstracts or full text? Evaluating input scope and computational efficiency in biomedical lay summarization  (2610.21690 - Evgin et al., 15 Oct 2026) in Related Work, subsection “Abstract-based strategies”

Full-text reading is supported but has not yet been evaluated at scale.

— Optimist Deep Research: An Iterative, State-Centric Agent for Autonomous Literature Research  (2610.23299 - Panapitiya, 16 Oct 2026) in Section 11, Discussion and Limitations, paragraph “Abstract-only evidence”