Reproduce “Document Attribution: Examining Citation Relationships using Large Language Models” 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
Vipula Rawte, Ryan A. Rossi, Franck Dernoncourt, Nedim Lipka
Researchers at Adobe Inc. developed methods to enhance the attribution of Large Language Model (LLM) outputs, where one approach, framing attribution as zero-shot textual entailment, improved F1 scores by up to 2.4% over baselines on out-of-distribution datasets. The study also found that specific attention layers within smaller LLMs could provide strong signals for attribution, achieving F1 scores up to 66.93.
Key insights
- A straightforward zero-shot textual entailment method can effectively enhance attribution for LLMs without requiring task-specific fine-tuning.
- Certain internal attention layers within LLMs contain valuable signals for tracing information provenance and improving attribution performance.
- The proposed methods advance the trustworthiness and transparency of LLM-generated content by providing verifiable links to sources.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.