Ground truth is a process, not a dataset
Amazon Science
Amazon's AGI group developed a new evaluation method for AI-generated research reports, discovering that traditional static benchmarks fail when assessing complex factuality claims that require cross-document synthesis. Expert evaluators achieved only 60.8% accuracy on known answers using standard labeling, but improved to 90.9% accuracy when placed in an auditing role that compared competing evidence. The audit-then-score protocol treats ground truth as a dynamic process where models can challenge benchmark answers with evidence, fundamentally changing how AI evaluation works and enabling DeepFact-Eval to reach 83.4% accuracy on their benchmark.
Why it matters
Automatically fact-checking long, AI-generated research reports poses new challenges — including benchmarking.
