XNews

Tencent Hunyuan's EvolveScaler: Benchmark for Information Evolution in AI

Tencent Hunyuan's EvolveScaler: Benchmark for Information Evolution in AI

Tencent Hunyuan introduced EvolveScaler, a benchmark based on a 40-day RPG log where the world state changes after reading. It tests an AI's ability to handle information evolution—retractions, corrections, and backfills—through 117 prototypes, 159 question operators, and up to ~1,200 events per sample. Frontier models scored a median avg@5 of just 11.3 on the hardest tier, while training on the dataset improved performance by +5.25 on average across 8 out-of-distribution benchmarks.

Source posts