Symbolic Augmentation Boosts Neural Fact-Checkers' Robustness

Large language models (LLMs) often exhibit a critical flaw when summarizing scientific texts: they hallucinate numbers and units, potentially inverting the meaning of a scientific claim. Researchers have identified a specific "canonical-equivalence blind spot" in neural fact-checkers, where models fail to recognize physically equivalent quantities expressed in different units or forms (e.g., Celsius vs. Kelvin). While a fine-tuned ModernBERT encoder achieved 89.9% macro-F1 on a new 1500-item benchmark for typed-quantity error detection, its accuracy plummeted to 36.5% on canonical-equivalent rewrites. To address this, a novel training-time framework called Symbolic Augmentation has been developed. This method generates label-preserving augmented training data by running modules of a symbolic verifier in reverse. This augmentation dramatically improves robustness to canonical-equivalence, boosting accuracy to 98.2%, and even slightly enhances in-distribution accuracy. The augmented encoder now matches closed-frontier LLMs without additional inference cost and shows improved transferability to external benchmarks. This work highlights that integrating symbolic and learned components through training-time augmentation is a highly effective strategy for improving the reliability of AI fact-checking.