Paper proposes Evidence Chain Evaluation to improve selective fact-checking in LLMs

A new paper on arXiv proposes Evidence Chain Evaluation (ECE), a selective fact-checking framework designed to address a reliability problem in large language model (LLM) fact-checking systems. The core issue is that forced binary decisions can lead systems to issue confident verdicts even when supporting evidence is weak, sparse, or internally inconsistent.

The ECE framework permits abstention through an uncertain verdict, allowing the system to avoid making a definitive true/false decision when evidence quality is insufficient. The evaluated system is a tool-using verification agent that gathers evidence via web search, scholarly search, and executable checks, then returns a structured verdict with confidence scores and source-level metadata.

On the ECE-Bench benchmark, the system achieves 91.6% standard accuracy, 93.7% coverage, and 97.8% selective accuracy on answered claims. While ECE does not outperform the strongest retrieval baseline on aggregate calibration metrics such as Expected Calibration Error, Brier score, or AURC, it demonstrates a clear selective-prediction trade-off: high accuracy is maintained on answered claims while deferring 6 of 95 cases.

The deferred cases are concentrated in lower-reliability evidence settings, with 5 of 6 deferrals occurring at source level L4, suggesting that abstention functions as a safety mechanism for handling epistemically weak evidence.

The authors provide code for ECE at a public GitHub repository, enabling reproducibility and further research.