![[2508.04605] Multitask Learning with Stochastic Interpolants](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2508.04605] Multitask Learning with Stochastic Interpolants
Summary
This article presents a new framework for multitask learning using stochastic interpolants, enhancing generative models' capabilities across various tasks without specific training.
Why It Matters
The proposed framework broadens the applicability of generative models by enabling them to learn from multiple probability distributions simultaneously. This innovation could significantly reduce the need for specialized models, making AI more efficient and versatile in handling diverse tasks.
Key Takeaways
- Introduces a framework for multitask learning using stochastic interpolants.
- Generalizes time dynamics of flow and diffusion models for broader applications.
- Demonstrates zero-shot efficacy in tasks like conditional generation and inpainting.
- Offers a theoretical perspective that unifies existing generative models.
- Potentially reduces reliance on task-specific models, enhancing efficiency.
Computer Science > Machine Learning arXiv:2508.04605 (cs) [Submitted on 6 Aug 2025 (v1), last revised 25 Feb 2026 (this version, v4)] Title:Multitask Learning with Stochastic Interpolants Authors:Hugo Negrel, Florentin Coeurdoux, Michael S. Albergo, Eric Vanden-Eijnden View a PDF of the paper titled Multitask Learning with Stochastic Interpolants, by Hugo Negrel and 3 other authors View PDF HTML (experimental) Abstract:We propose a framework for learning maps between probability distributions that broadly generalizes the time dynamics of flow and diffusion models. To enable this, we generalize stochastic interpolants by replacing the scalar time variable with vectors, matrices, or linear operators, allowing us to bridge probability distributions across multiple dimensional spaces. This approach enables the construction of versatile generative models capable of fulfilling multiple tasks without task-specific training. Our operator-based interpolants not only provide a unifying theoretical perspective for existing generative models but also extend their capabilities. Through numerical experiments, we demonstrate the zero-shot efficacy of our method on conditional generation and inpainting, fine-tuning and posterior sampling, and multiscale modeling, suggesting its potential as a generic task-agnostic alternative to specialized models. Subjects: Machine Learning (cs.LG); Dynamical Systems (math.DS) Cite as: arXiv:2508.04605 [cs.LG] (or arXiv:2508.04605v4 [cs.LG] for this ve...
