To See a World in a Spark of Neuron: Disentangling Multi-task Interference for Training-free Model Merging

Zitao Fang, Guodong Du, Shuyang Yu, Yifei Guo, Yiwei Zhang, Yiyao Cao, Jing Li, Ho-Kin Tang, Sim Kuan Goh · 2025

Read To See a World in a Spark of Neuron: Disentangling Multi-task Interference for Training-free Model Merging by Zitao Fang, Guodong Du, Shuyang Yu, Yifei Guo, Yiwei Zhang, Yiyao Cao, Jing Li, Ho-Kin Tang, Sim Kuan Goh (2025) 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-2503-05320.txt.

Description

Zitao Fang and eight co-authors take on clashes that come when staff join more than one tuned model into one. Fine tuning lifts task scores but often hurts skill on new tasks, and plain joins can make scores worse. Their frame, NeuroMerging, works at the level of nerve cells. It does not treat the model as a black box.

The team splits task traits into two nerve cell spaces. One space rules how the model takes in new data. The other rules how it fits a task. Joining in those spaces cuts clashes with no more train time. In tests on speech tasks and sight tasks, NeuroMerging beat old ways. The team says match of nerve cell acts is a good path for joins of what models know.

Catalog record

CREATORZitao Fang, Guodong Du, Shuyang Yu, Yifei Guo, Yiwei Zhang, Yiyao Cao, Jing Li, Ho-Kin Tang, Sim Kuan Goh
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE61 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-05320

Browse the full library · About the Tower of Babel