TAPIOCA: Why Task- Aware Pruning Improves OOD model Capability

The paper introduces TAPIOCA, a method for task-aware layer pruning that enhances out-of-distribution (OOD) performance in large language models without benefiting in-distribution (ID) accuracy. The authors provide a geometric framework explaining how OOD inputs distort task-adapted representations and demonstrate that pruning specific layers can realign OOD inputs with the model's adapted geometry, leading to improved accuracy. This approach is significant for practitioners as it offers a targeted strategy to enhance model robustness in real-world applications where OOD data is prevalent.