FAST

“FAST [35] tokenization relies on a BPE training procedure... making the tokenization data-dependent and potentially unstable across domains or control ranges. Their ablation in Table 4a shows FFAD achieving 98.0% success compared to FAST's 96.9%”