Turning the Tide: Cross-Architecture Distillation for Diffusion Large Language Models introduces a new framework designed to shrink massive, high-performing diffusion large language models (dLLMs) into much smaller, efficient versions. While previous distillation methods focused on compressing models within the same architecture, this research addresses the more complex challenge of "cross-architecture" distillation, where the teacher and student models differ in their internal structure, attention mechanisms, and tokenizers.
Bridging the Gap Between Models
Distilling knowledge from a large teacher model to a small student model is difficult when the two models are fundamentally different. The researchers identified three primary barriers: the teacher’s reliability changes depending on the diffusion timestep, heavy masking during training makes it hard for the teacher to provide useful information, and different tokenizers make it mathematically difficult to align the models' outputs. The Tide framework solves these issues by integrating three specialized, modular components that work together to guide the student model's learning process.
How the Framework Works
The Tide framework uses three core components to manage the transfer of knowledge:
Tidal (Scheduling): This component acts as a "pacemaker" for the training process. It adjusts the distillation strength based on both the training progress and the diffusion timestep. By doing this, the student model learns from the teacher only when the teacher’s signals are most reliable, avoiding the noise that occurs at high masking levels.
CompDemo (Context): To solve the problem of limited context, this component enriches the teacher's input. It splits the masked tokens into subsets, allowing the teacher to see more information during its forward passes. This provides the student with clearer, more robust targets for learning.
Reverse Calm (Output Alignment): This component handles the challenge of mismatched tokenizers. It uses a technique that aligns text chunks between the two models rather than trying to map individual tokens directly. By reversing the direction of the loss calculation, it creates a more stable training process that filters out noise and prevents the gradient explosions that often occur when models are poorly aligned.
Notable Performance Gains
The researchers tested the framework by distilling 8B dense and 16B Mixture-of-Experts (MoE) teacher models into a compact 0.6B student model. The results showed that the distilled models significantly outperformed the non-distilled baseline across eight different benchmarks, including tasks in reasoning, knowledge, and commonsense. Most notably, the distilled models showed a major improvement in code generation, with HumanEval scores reaching 48.78 compared to 32.3 for the autoregressive baseline.
Key Takeaways
The study demonstrates that cross-architecture distillation is not only possible but highly effective for dLLMs. By using a modular approach, the researchers were able to tailor the distillation strategy to the specific needs of different model pipelines. The success of these experiments suggests that large, complex diffusion models can be successfully compressed into smaller, faster, and more deployable versions without sacrificing significant performance.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!