Abstract:
This work explores methods for modifying, transferring, and restructuring internal representations inside large models. The paper introduces a set of techniques for “model surgery,” enabling controlled edits to latent spaces, representation transfer between models, and structural adjustments without full retraining. The goal is to provide practical tools for understanding and manipulating internal model behavior at a finer granularity.
Paper: https://doi.org/10.5281/zenodo.19467270
Code: https://github.com/HeavenFYouMissed/model-surgery-paper
[link] [comments]