R2VC is a modular architecture designed for evidence-grounded fact checking. It integrates a hybrid sparse+dense retrieval system over Wikipedia, a generator producing diverse verdict candidates, and an external NLI cross-encoder for candidate selection. A sequence-level calibrator estimates confidence and enables selective abstention. On the FEVER dataset, a 8B backbone utilizing R2VC demonstrated a 13.74% accuracy increase over baseline models.
Ablation studies revealed that verifier-based candidate selection and confidence calibration were the most significant contributors to performance gains. Removing candidate selection reduced FEVER accuracy to 76.24%, while removing calibration increased the Brier score to 0.161. Analysis of 250 errors identified retrieval failures, particularly those involving wrong-entity evidence, as the primary bottleneck.
These results indicate that modular fact-checking pipelines can substantially improve both predictive accuracy and confidence reliability in open-domain verification. The system’s modular design allows for targeted improvements and diagnostics. The architecture’s components operate independently, facilitating identification and mitigation of specific failure modes.
The research emphasizes the importance of addressing uncertainty and improving confidence estimation in large language models used for fact-checking. The findings contribute to building more robust and trustworthy automated verification systems. Source: https://arxiv.org/abs/2609.11955
