CForce Boosts Parallel Decoding for Diffusion LLMs

Researchers have developed a new technique called Consistency Forcing (CForce) to enhance the performance of diffusion large language models (dLLMs), particularly in parallel decoding scenarios. dLLMs are designed to accelerate language generation by predicting multiple masked tokens simultaneously, but this can lead to unreliable predictions in early denoising stages, causing errors to propagate. CForce addresses this by using a distillation method that forces the mask predictions made in earlier stages to align more closely with the more accurate predictions from later stages. The method trains the dLLM on self-rollout trajectories, improving the alignment between training and inference. It employs a Confidence Adaptive KL Divergence as a distillation objective, which combines the benefits of both forward and reverse KL divergence. This approach is applicable to both mask-to-token and edit-capable decoding, where later token-to-token refinements provide additional supervision. Experiments with LLaDA models demonstrate that CForce significantly improves the speed-quality trade-off, especially when using aggressive parallel decoding strategies.