Reproduce “Unstructured Evidence Attribution for Long Context Query Focused Summarization” 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

Dustin Wright, Zain Muhammad Mujahid, Lu Wang, Isabelle Augenstein, David Jurgens

Researchers from the University of Copenhagen and the University of Michigan developed a method for unstructured evidence attribution in long-context query-focused summarization, utilizing a cost-effective synthetic dataset called SUnsET. Their approach enhanced LLM summary quality, increased evidence precision, and reduced positional bias in evidence extraction.

Key insights

  • Unstructured evidence attribution offers superior precision and relevance compared to fixed-granularity methods, enhancing the trustworthiness of LLM summaries.
  • High-quality synthetic datasets can be generated cost-effectively (SUnsET at ~$200 vs. ~$237k manual) to train LLMs for complex, fine-grained tasks.
  • Training with shuffled document sections significantly reduces the "lost-in-the-middle" positional bias in LLMs, improving evidence retrieval from all parts of a long context.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.