Dynamic Model Merging Made Slim

Guodong Du, Wanyu Lin · 2026

Read Dynamic Model Merging Made Slim by Guodong Du, Wanyu Lin (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-18904.txt.

Description

Guodong Du and Wanyu Lin slim down dynamic model merging, which combines fine-tuned models without joint training or task data. Dynamic methods keep task-specific capacity but must split limited storage between shared and expert weights. Their DiDi-Merging method assigns low-rank capacity to both parts through differentiable rank choice, then refines parameters without data. On an eight-task vision benchmark the small setup keeps 98.7 percent accuracy at just 1.24 times one model's storage. Dropping refinement costs over four points, and results assume the task is known.

Catalog record

CREATORGuodong Du, Wanyu Lin
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE97735
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.18904

Browse the full library · About the Tower of Babel