Generative Adaptation of Dynamics to Environmental Shifts via Weight-space Diffusion

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Li, Ruikun, Wang, Huandong, Ding, Jingtao, Yuan, Yuan, Liao, Qingmin, Li, Yong
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866915977449439232
author Li, Ruikun
Wang, Huandong
Ding, Jingtao
Yuan, Yuan
Liao, Qingmin
Li, Yong
author_facet Li, Ruikun
Wang, Huandong
Ding, Jingtao
Yuan, Yuan
Liao, Qingmin
Li, Yong
contents Data-driven dynamics prediction often fails under environmental shifts, while traditional fine-tuning remains computationally prohibitive for hardware-constrained or data-scarce applications. We propose DynaDiff, a generative meta-learning framework that transitions the paradigm from gradient-based tuning or modulation to direct weight-space generation. Specifically, we first abstract expert weights as novel weight graphs, utilizing multi-head attention to explicitly capture topological coupling within weights. Subsequently, we design a functional loss to ensure that the generated models achieve consistency with expert models in physical behavior. Finally, we develop a dynamics-informed prompter that extracts cross-domain physical and spectral features from observation sequences to condition the diffusion model. Experiments demonstrate that DynaDiff boosts average prediction accuracy by 10.78% over competitive baselines. Furthermore, by pre-constructing a model zoo of expert predictors, we amortize the fine-tuning overhead into a one-time offline cost, significantly boosting deployment efficiency in new environments.
format Preprint
id arxiv_https___arxiv_org_abs_2505_13919
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Generative Adaptation of Dynamics to Environmental Shifts via Weight-space Diffusion
Li, Ruikun
Wang, Huandong
Ding, Jingtao
Yuan, Yuan
Liao, Qingmin
Li, Yong
Computational Engineering, Finance, and Science
Data-driven dynamics prediction often fails under environmental shifts, while traditional fine-tuning remains computationally prohibitive for hardware-constrained or data-scarce applications. We propose DynaDiff, a generative meta-learning framework that transitions the paradigm from gradient-based tuning or modulation to direct weight-space generation. Specifically, we first abstract expert weights as novel weight graphs, utilizing multi-head attention to explicitly capture topological coupling within weights. Subsequently, we design a functional loss to ensure that the generated models achieve consistency with expert models in physical behavior. Finally, we develop a dynamics-informed prompter that extracts cross-domain physical and spectral features from observation sequences to condition the diffusion model. Experiments demonstrate that DynaDiff boosts average prediction accuracy by 10.78% over competitive baselines. Furthermore, by pre-constructing a model zoo of expert predictors, we amortize the fine-tuning overhead into a one-time offline cost, significantly boosting deployment efficiency in new environments.
title Generative Adaptation of Dynamics to Environmental Shifts via Weight-space Diffusion
topic Computational Engineering, Finance, and Science
url https://arxiv.org/abs/2505.13919