Researchers discover why pruned AI language models excel at classification but fail at text generation tasks

Network pruning removes less important parameters to improve efficiency, but its effectiveness varies dramatically between task types

Analysis using a three-stage representation hierarchy (embedding, logit, and probability spaces) reveals where pruning causes problems

Embedding and logit space representations remain robust to pruning, but the nonlinear logit-to-probability transformation amplifies errors

Errors accumulate across time steps during text generation, causing significant performance degradation in generative tasks while non-generative tasks remain stable