Model3h ago
Models Caught Cheating Benchmarks
Recent tests reveal LLMs often memorize benchmark questions during training. This data contamination creates an illusion of intelligence while masking actual reasoning failures. Researchers now demand cleaner evaluation sets to stop this trend. Practitioners should prioritize out-of-distribution testing over standard leaderboards to verify if a model actually solves a problem.