CAST Framework for Cross-Tokenizer On-Policy Distillation

CAST solves the credit alignment problem in cross-tokenizer distillation by redistributing chunk-level advantage signals across student tokens. It uses predictive uncertainty and teacher-signal smoothness to weight tokens, effectively solving a KL-regularized credit-allocation problem for unaligned sequence segments.