Reproduce “TLDR: Extreme Summarization of Scientific Documents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Isabel Cachola, Kyle Lo, Arman Cohan, Daniel S. Weld
This research from the Allen Institute for AI defined and introduced a new task of extreme summarization for scientific documents, TLDR generation, and released SCITLDR, a multi-target dataset for this purpose. Their proposed CATTS learning strategy, leveraging paper titles, improved the informativeness of generated ultra-short summaries while maintaining factual correctness.
Key insights
- Scientific TLDRs are characterized by extreme compression ratios (238.1) and moderate abstractiveness, typically containing 2-4 key information "nuggets" (e.g., Area, Contribution, Results).
- There is significant lexical variability (e.g., 15.0% 1-gram Jaccard index) between different human-written TLDRs for the same paper (e.g., author-written vs. peer-reviewer derived), underscoring the need for multi-target evaluation.
- Leveraging paper titles as an auxiliary training signal is an effective strategy for improving TLDR generation models, particularly in data-scarce specialized domains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.