{
  "abstract": "Variational autoencoders (VAEs) provide a probabilistic framework for learning latent representations and generating data through approximate posterior inference and probabilistic decoding. Although conventional VAE research primarily focuses on improving the representation, inference, generation, and optimization capabilities of a given model, a complementary question concerns how the VAE itself can be constructed, adapted, and optimized according to the characteristics of a task. This paper proposes the concept of Meta-Variational Autoencoders (Meta-VAEs), a framework in which a meta-level mechanism learns to generate, select, configure, and adapt the mechanisms underlying VAE architectures and their training processes. Rather than treating a VAE as a fixed generative architecture, Meta-VAE treats its architectural structure, latent-space configuration, prior distributions, approximate posterior families, encoder–decoder organization, objective functions, and optimization procedures as potential objects of meta-level decision-making.\n\nThe proposed framework formalizes a hierarchy consisting of a base VAE, a VAE configuration space, a meta-controller, and a task-dependent evaluation mechanism. The controller selects or generates a candidate VAE configuration based on task characteristics, resource constraints, uncertainty, and previous evaluation results. The resulting VAE is then trained or adapted and evaluated, with feedback used to improve subsequent configuration decisions. A general objective is formulated to balance task performance, computational cost, adaptation cost, and model complexity while respecting application-specific constraints. The framework also considers recursive extensions in which the mechanisms responsible for constructing VAEs become objects of higher-level optimization. Relationships to hierarchical latent-variable models, meta-learning, neural architecture search, hyperparameter optimization, and variational inference are discussed to clarify the proposed framework's scope. Finally, an experimental research agenda is developed to evaluate whether meta-level VAE construction and adaptation can improve performance, efficiency, robustness, and transfer across tasks. This paper establishes a conceptual and mathematical framework rather than claiming empirically demonstrated improvements.",
  "abstract_shown": true,
  "affiliations": [],
  "article_type": null,
  "arxiv": null,
  "author_total": 1,
  "authors": [
    {
      "name": "Jincheng Zhang",
      "orcid": "0009-0005-1860-0009"
    }
  ],
  "bibtex": "https://research.draconitic.com/paper/bb2bv9qzehncm.bib",
  "cited_by": 0,
  "doi": "10.5281/zenodo.23529981",
  "fulltext_url": null,
  "id": "bb2bv9qzehncm",
  "issns": [],
  "landing_url": "https://zenodo.org/records/23529981",
  "language": null,
  "licence": "CC-BY-4.0",
  "licence_label": "CC BY 4.0",
  "licence_source": "article",
  "licence_url": "https://creativecommons.org/licenses/by/4.0/",
  "metadata_licence": null,
  "native_subjects": [],
  "openalex": null,
  "pdf_url": "https://zenodo.org/api/records/23529981/files/Meta-Variational%20Autoencoders.pdf/content",
  "pmcid": null,
  "pmid": null,
  "published": "2026-10-21",
  "publisher": null,
  "ris": "https://research.draconitic.com/paper/bb2bv9qzehncm.ris",
  "sources": [
    {
      "first_seen": "2026-11-04T12:21:15Z",
      "native_id": "23529981",
      "source": "zenodo",
      "url": "https://zenodo.org/records/23529981"
    }
  ],
  "subjects": [],
  "title": "Meta-Variational Autoencoders: A Framework for Learning to Construct, Adapt, and Optimize Variational Autoencoders",
  "title_text": "Meta-Variational Autoencoders: A Framework for Learning to Construct, Adapt, and Optimize Variational Autoencoders",
  "type": "preprint",
  "updated_at": "2026-11-04T12:21:15Z",
  "url": "https://research.draconitic.com/paper/bb2bv9qzehncm",
  "venue": "Zenodo",
  "versions": []
}