PEARL: Auditable Repair for Scientific Reasoning Graph Extraction
2026-07-20 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors address the problem of turning messy scientific explanation graphs generated by large language models (LLMs) into clear, reliable reasoning graphs. They introduce PEARL, a method that fixes errors in the graphs without needing extra training, by following strict logical rules and using evidence-based feedback. Tested on a benchmark of scientific papers, PEARL greatly improved the quality and correctness of the reasoning graphs. This helps create trustworthy and easy-to-check reasoning paths for AI tools used in scientific research.
Scientific Reasoning Graph ExtractionLarge Language ModelsPeircean SchemaGraph RepairLatent Reasoning ChainsEvidence-grounded FeedbackSemantic ValidityResearch-agent WorkflowsAI ScientistAudit Trail
Authors
Bohan Su, Pengze Li, Yuchen Lu, Xi Chen
Abstract
Scientific Reasoning Graph Extraction (SRGE) aims to recover explicit links among observations, evidence, intermediate claims, and paper-level conclusions. LLMs can produce graph-like scientific explanations, but their outputs often mix malformed syntax, drifting edge labels, incorrectly oriented roots, and weak source anchors. We propose PEARL (Peircean Extraction via Abstraction and Repair Layer), a training-free framework that turns noisy LLM graph responses into auditable reasoning graphs and repairs them toward strict semantic validity. PEARL first materializes explicit graph content under a closed Peircean schema, then uses matched evidence-grounded judge feedback to repair rejected edge types, local inference steps, and terminal roots while preserving an audit trail. On five 70-paper model archives from ARCHE, a benchmark for latent reasoning-chain extraction, PEARL raises strict gate passes from 0/350 for the LLM baseline to 300/350, with average REA improving from 0.339 to 0.906. The graphs provide a reliability layer for research-agent and AI scientist workflows that need inspectable reasoning traces rather than unconstrained graph regeneration. Code and audit artifacts are available at https://github.com/BohanSu/auditable-repair-reasoning-graphs/tree/300-350_workshop .