Tencent Hunyuan Introduces EvolveScaler for Evaluating AI Information Evolution

Tencent Hunyuan released EvolveScaler, a benchmark using a 40-day RPG log to test an AI's ability to handle information that evolves through retractions, corrections, and backfills. The hardest tier of the 117-prototype benchmark saw median accuracy drop to 11.3% across 14 frontier models, but training on the data improved average performance by 5.25 points on 8 out-of-distribution benchmarks.