PEARL: Auditable Repair for Scientific Reasoning Graph Extraction

Source

Source name: arXiv
Source kind: paper
Authors / organization: Bohan Su, Pengze Li, Yuchen Lu, Xi Chen
Venue: arXiv: cs.AI
Published: 2026-07-20
Collected: 2026-07-21
URL: http://arxiv.org/abs/2607.17917v1
PDF: https://arxiv.org/pdf/2607.17917v1
DOI: UNKNOWN

Plain-Language Brief

The paper proposes a new framework called PEARL for extracting reliable scientific reasoning graphs from Large Language Models (LLMs). The LLM outputs often contain errors such as malformed syntax, incorrectly oriented roots, and weak source anchors. PEARL aims to turn these noisy responses into auditable reasoning graphs that are more semantically valid. It uses matched evidence-grounded judge feedback to repair the graphs, while preserving an audit trail. The framework was tested on five model archives from ARCHE, a benchmark for latent reasoning-chain extraction, and improved the strict gate passes from 0/350 to 300/350.

What the Document Says

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 contain errors such as malformed syntax, drifting edge labels, incorrectly oriented roots, and weak source anchors.
PEARL (Peircean Extraction via Abstraction and Repair Layer) is a training-free framework that turns noisy LLM graph responses into auditable reasoning graphs and repairs them towards strict semantic validity.
PEARL materializes explicit graph content under a closed Peircean schema.
PEARL 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, PEARL raises strict gate passes from 0/350 for the LLM baseline to 300/350.
The average REA improves from 0.339 to 0.906 with the use of PEARL.

Key Concepts Mentioned

Scientific Reasoning Graph Extraction
LLMs
PEARL
Peircean schema
evidence-grounded judge feedback
strict semantic validity
ARCHE
strict gate passes
REA

Claims Made by the Document

PEARL can improve the reliability of graph responses produced by LLMs.
PEARL raises strict gate passes from 0/350 to 300/350 on five model archives from ARCHE.

Evidence Used

On five 70-paper model archives from ARCHE, 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.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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 .