A 2026 arXiv preprint reports higher average scores for sparse model merging, while noting added training overhead and mixed task results.