Reproduce “PEARL: Auditable Repair for Scientific Reasoning Graph Extraction” 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
Bohan Su, Pengze Li, Yuchen Lu, Xi Chen
PEARL repairs scientific reasoning graphs for reuse through structural cleanup and evidence-matched local corrections, preserving supported graph parts and re-evaluating edits before acceptance. On 350 outputs from five ARCHE archives, strict passes requiring both CG and REA to equal 1.0 rose from 0/350 to 300/350. Results are limited to this schema and text-only model archives.
Key insights
- Decoupling structural recovery from semantic validation, combined with a strict acceptance gate and auditable repairs, is essential for transforming noisy LLM outputs into reliable scientific reasoning graphs.
- The primary weakness of raw LLM outputs in scientific reasoning graph extraction (SRGE) lies in the correctness of reasoning edges and terminal root structures (semantic validity) rather than just content coverage.
- Localized, judge-feedback-guided repair, which avoids full graph regeneration, preserves original model output where appropriate and enables fine-grained auditability, critical for scientific trustworthiness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.