🤖 AI Summary
This work addresses a critical limitation in current reference-free faithfulness evaluation, which focuses solely on precision while neglecting recall—the coverage of relevant factual content—thereby encouraging models to adopt a “less is more” bias. We propose the first coverage-aware faithfulness evaluation framework, enabling precise quantification of recall in domains with complete factual annotations, such as Formula 1 race reports and NOAA weather forecasts. Our analysis reveals severe under-coverage in mainstream models, with even the best achieving recall below 50% and ranking last by F1 score. We introduce a unified scoring metric and a reference-free Verifier-Guided Generation approach. Evaluated on 7,253 multilingual decision-making instances, our metric achieves a system-level Spearman correlation of 1.0. The benchmark, annotations, metrics, baselines, and an interactive demo are publicly released.
📝 Abstract
Reference-free faithfulness metrics verify each atomic claim a model makes against ground truth, and are increasingly used to evaluate grounded generation. We show they share a blind spot: they measure only precision -- are the stated claims supported? -- and therefore reward abstention, since a model can score near-perfect faithfulness by saying almost nothing. We make this measurable using Formula 1 telemetry, a domain where strategic ground truth is derived deterministically and, crucially, completely: for each decision we know the full set of facts that mattered. This completeness -- absent in open-domain faithfulness benchmarks -- lets us measure recall (coverage of the relevant facts) exactly, alongside precision. On a multilingual (EN/ES/PT) benchmark of 7,253 decision instances spanning 150 races, the most precise frontier model covers under half of the relevant facts and ranks last by F1, so requiring coverage reorders the systems; the same effect reappears in a second complete-oracle domain (NOAA weather forecasts). A prompt ablation shows the low coverage is not an under-prompting artifact: explicitly asking models to be thorough does not close the gap. We pair faithfulness with coverage into a single score, validate the metric (controlled perturbation; agreement across a model-free regex extractor and a cross-family LLM extractor, system-level Spearman 1.0), and give a verifier-guided generation method that improves precision and recall without references. We release the benchmark, structured annotations, metric, baselines, and an interactive demo.