Reproduce “LiRA: A Multi-Agent Framework for Reliable and Readable Literature Review Generation” 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
Gregory Hok Tjoan Go, Khang Ly, Anders Søgaard, Amin Tabatabaei, Maarten de Rijke, Xinyi Chen
LiRA divides review writing among agents that plan outlines, draft subsections from selected references, edit assembled reviews, and return weak components for revision. It uses writing guidelines without task-specific fine-tuning. On SciReviewGen, LiRA scores 0.76 on CQF1, versus 0.63 for AutoSurvey. CQF1 estimates citation support using an LLM judge and reference abstracts; the result applies to tested writing setups, not independent literature search, screening, or risk-of-bias assessment.
Key insights
- Decomposing complex literature review writing into specialized, interacting LLM agents that emulate human workflows leads to superior qualitative and quantitative outcomes compared to monolithic LLM approaches.
- Explicitly incorporating citation grounding and an iterative review process is crucial for enhancing factual accuracy and significantly mitigating hallucinations in LLM-generated scientific text.
- Automated systems can achieve high readability and structural coherence in long-form scientific writing, producing generated articles that are both usable and trustworthy for researchers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.