cross-architecture

TIDE distills diffusion LLMs across architectures

TIDE distills diffusion LLMs across architectures, adding noise-aware weighting and tokenizer-aware objectives to improve a 0.6B student.