Benchmark RadarAI BENCHMARK PROFILE
ConflictBench
ConflictBench is a benchmark with ConflictScore metric to quantify how well models acknowledge conflicting evidence in grounding documents, decomposing responses into claims and labeling them against documents.
- Released
- 2026-06-24
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Existing factuality metrics ignore coexisting contradictions; ConflictScore provides a nuanced measure over ConflictBench, offering a corrective feedback mechanism for improving truthfulness.
Motivation
Existing metrics for factuality and faithfulness evaluate whether an answer is supported or contradicted by its grounding documents, but they fail to capture when both supporting and contradicting evidence coexist.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.