Prompt Compression in Diffusion Large Language Models: Evaluating LLMLingua-2 on LLaDA

In three linesStudy of prompt compression on LLaDA, an 8B-parameter DLLM, using LLMLingua-2. Evaluation on GSM8K, DUC2004, ShareGPT at 2× compression ratio shows semantic preservation does not ensure stability in diffusion models: mathematical reasoning degrades substantially while summarization remains robust. Autoregressive compression methods do not transfer uniformly to DLLMs.

Summary generated by Claude — human-verified