REKEY: New Benchmark Exposes VLM Score Inflation from Memorization

jiqizhixin · x · 2026-08-14

AI benchmarks often suffer from data contamination, where test data leaks into training, allowing models to memorize answers rather than truly understand. To address this, researchers introduced REKEY.

REKEY is a live benchmark that randomly swaps crucial details in real images at test time, forcing Vision-Language Models (VLMs) to rely on actual perception. Across 8 top VLMs, scores dropped by 9.5 to 18.8 points compared to standard tests, demonstrating the extent of memorization-induced score inflation.

Original post →

More from Research

Research channel →