Prune, Update and Trim: Robust Structured Pruning for Large Language Models

Researchers introduced Putri, a post-training structured pruning method for large language models.

Putri reduces pruning damage by updating remaining feed-forward weights, pruning layers sequentially, and removing individual attention heads.

The authors report strong results across multiple models, sparsity levels, and datasets.

The approach could make LLMs cheaper and more practical to run, especially for long-context inference and limited hardware.