Turning the TIDE: Cross-Architecture Distillation for Diffusion Large Language Models

TL;DR

This paper introduces TIDE, a novel framework for cross-architecture distillation of diffusion large language models, enabling knowledge transfer between models with different architectures, attention mechanisms, and tokenizers.

Contribution

TIDE is the first framework for cross-architecture distillation of dLLMs, incorporating three modular components to improve knowledge transfer and model performance.

Findings

Distilling 8B dense and 16B MoE teachers into a 0.6B student improves benchmark scores.

The approach yields a 48.78 HumanEval score, surpassing the baseline.

Outperforms baseline by an average of 1.53 points across eight benchmarks.

Abstract

Diffusion large language models (dLLMs) offer parallel decoding and bidirectional context, but state-of-the-art dLLMs require billions of parameters for competitive performance. While existing distillation methods for dLLMs reduce inference steps within a single architecture, none address cross-architecture knowledge transfer, in which the teacher and student differ in architecture, attention mechanism, and tokenizer. We present TIDE, the first framework for cross-architecture dLLM distillation, comprising three modular components: (1) TIDAL, which jointly modulates distillation strength across training progress and diffusion timestep to account for the teacher's noise-dependent reliability; (2) CompDemo, which enriches the teacher's context via complementary mask splitting to improve predictions under heavy masking; and (3) Reverse CALM, a cross-tokenizer objective that inverts chunk-level likelihood matching, yielding bounded gradients and dual-end noise filtering. Distilling 8B dense and 16B MoE teachers into a 0.6B student via two heterogeneous pipelines outperforms the baseline by an average of 1.53 points across eight benchmarks, yielding notable gains in code generation, where HumanEval scores reach 48.78 compared to 32.3 for the AR baseline.