Compute Optimal Tokenization Study

Researchers trained 1,300 models to establish compression-aware neural scaling laws, demonstrating that bytes-per-token metrics should replace token counts for compute allocation decisions to optimize efficiency across multilingual datasets.