Instruction-Guided Autoregressive Neural Network Parameter Generation

Soro Bedionita 1, Bruno Andreis 1, Song Chong 1, Sung Ju Hwang 1, 2] · 2025

Read Instruction-Guided Autoregressive Neural Network Parameter Generation by Soro Bedionita 1, Bruno Andreis 1, Song Chong 1, Sung Ju Hwang 1, 2] (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-2504-02012.txt.

Description

Soro Bedionita, Bruno Andreis, Song Chong, and Sung Ju Hwang present a way to build neural network weights from written instructions in a 2025 paper. Instead of training a fresh model for each task, their system writes the weights a task needs.

Older methods based on diffusion struggle with large models and can produce layers that do not fit together well. The new system, called IGPG, builds weights step by step in order, guided by the task, the data, and the model shape. Tests on several image data sets gave results as good as or better than top methods. Gains were clearest for large models, where speed and scale matter most.

Catalog record

CREATORSoro Bedionita 1, Bruno Andreis 1, Song Chong 1, Sung Ju Hwang 1, 2]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-02012

Browse the full library · About the Tower of Babel