Information Boundaries for Group-Robust LLM Pruning
August 5, 2026
The study identifies why compression statistics fail to maintain group robustness during LLM pruning. Implementing a coarse depth allocation can reduce worst-group perplexity inflation by up to 20.9% compared to balanced uniform allocation.
HOW THIS AFFECTS YOU
●
researcherYou can use these information interface models to prevent performance degradation in specific subgroups during pruning.