chiprook
← AI
AIOctober 24, 2026, 05:02

Audit finds answer leakage in OpenBookQA, SciQ and QASC annotations

MV2 (Munim, Inc.) measured how often the gold answer appears verbatim in the annotations of three science benchmarks. Leakage scales with how answer-shaped the annotation is: 6.36% in OpenBookQA, 91.57% in SciQ and 87.64% in QASC, versus 0.59–4.49% for distractors.

Audit finds answer leakage in OpenBookQA, SciQ and QASC annotations
#OpenBookQA#SciQ#QASC
Read next
AI

Benchmark: DeepSeek-R1 Missed Data Leakage in ML Pipeline Audit

AI

Audit finds hidden prompt tokens in cheap LLM APIs

AI

Audit finds 25 defects in LLM benchmark metrics and gold labels

AI

Audit of 110 AI usage metering tools finds 45+ billing bugs