Reproduce “CForce: Boosting Parallel Decoding for dLLMs via Consistency Forcing” on OpenResearch

About this paper

Yuji Ren, Chenkai Xu, Zhuocheng Gong, Jianguo Li, Zhijie Deng

Researchers developed CForce, a self-distillation method for diffusion large language models (dLLMs) that enhances parallel decoding efficiency by enforcing consistency between early and late generation stages. It achieved up to 78.3% higher Tokens Per Forward pass and 42.82% faster wall-clock throughput on non-edit dLLMs, while improving average accuracy by 0.84 percentage points and TPF by 30.8% on edit-capable dLLMs.

Key insights

• Later, more context-rich decoding states (especially with Token-to-Token refinement in edit-capable dLLMs) can provide effective supervisory signals for improving the reliability of earlier, low-context predictions.
• Self-distillation, where a dLLM learns from its own evolving decoding trajectories using stop-gradient targets from later states, proves effective in improving training-inference alignment.
• Combining a forward KL divergence with a confidence-weighted reverse KL divergence (CAD) effectively balances distributional drift control and mode-seeking sharpening, preventing premature collapse while boosting parallelism.

Summary by alphaXiv.