Symbolic Augmentation boosts neural fact-checker robustness on scientific text

Researchers have developed a new method called Symbolic Augmentation to improve the accuracy of neural fact-checkers, particularly in handling scientific text. These models often struggle with numbers and units, leading to errors that can subtly alter scientific claims. The new approach addresses a specific blind spot where canonical-equivalent quantities, like different temperature scales, cause accuracy to collapse. By generating augmented training data that preserves labels, Symbolic Augmentation significantly boosts robustness and even improves in-distribution accuracy, matching closed-frontier LLMs without increased inference cost. AI

IMPACT Enhances the reliability of LLMs in scientific contexts, reducing hallucinations and improving the accuracy of claim verification.

RANK_REASON This is a research paper detailing a new method for improving LLM fact-checking capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• Krippendorff's alpha
• ModernBERT
• ScienceCast
• SciFact-Open
• Symbolic Augmentation

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →