Block-OBS-GS: Exact Per-Block Joint Brain Surgery with Gauss–Seidel Refinement for LLM Pruning
To appear in Conference on Neural Information Processing Systems (NeurIPS), Dec. 2026.
Authors: Yuwen Huang and Xiang Pan
Block-OBS-GS improves the reconstruction of large language models after post-training pruning. It jointly adjusts the weights retained within each block, then uses a Gauss–Seidel sweep to refine the solution across blocks. The paper analyzes the fixed-mask reconstruction problem and evaluates the trade-off between model quality and pruning time.
