Paris Giampouras

Optimization & Representation Learning · Generative Models · Adaptive & Agentic AI

Foundations and algorithms for reliable, adaptive AI.

I develop theory and algorithms for learning useful representations, aligning generative models, and adapting AI systems, drawing on optimization, probabilistic inference, and structured low-rank models.

Publications RAISE Lab Projects Contact

Paris Giampouras
University of Warwick · Meta FAIR

I am an Assistant Professor of Machine Learning and AI in the Department of Computer Science at the University of Warwick, where I am a member of the Foundations of AI and Machine Learning (FAM) division. I am also a part-time Research Scientist at FAIR, Meta AI within Meta Superintelligence Labs, and an Affiliated Researcher at the Archimedes AI Research Center.

Previously, I was a Research Faculty member at the Mathematical Institute for Data Science (MINDS) at Johns Hopkins University, where I held a Marie Skłodowska-Curie Global Fellowship from 2019 to 2022.

⚙️

Optimization & Structured Learning

Low-rank matrix factorization, tensor decompositions, nonconvex optimization, and Bayesian and variational inference, with algorithms and theoretical guarantees for efficient and robust learning.

✨

Generative Models & Alignment

Representation learning, post-training, and alignment of diffusion and flow-based models, with applications to inverse problems, controllable generation, and scientific imaging.

🤖

Continual Adaptation & AI Agents

Continual learning and parameter-efficient adaptation of LLMs, alongside AI research agents, test-time reasoning, and automated machine-learning experimentation.

My research connects the mathematical foundations of learning with the adaptation and alignment of modern AI models. I combine theoretical analysis, algorithm design, and empirical evaluation to understand how models acquire useful representations, adapt efficiently to new tasks, and retain previously learned knowledge. My work on low-rank matrix factorization, tensor decomposition, and optimization complements my research on generative models, robust probabilistic inference, and continual learning.

At Meta, my work includes contributions to the post-training pipeline for Muse Glimmer, an open-weight language model, and research on AI agents for automated machine-learning experimentation. My current interests include continual and parameter-efficient adaptation of LLMs, post-training of diffusion and flow models, and agents that use reasoning, feedback, and experimentation to improve their research decisions.

Examples of my work include:

Representation alignment: Align & Invert (NeurIPS 2026) applies alignment with pretrained visual encoders at inference time to improve image reconstruction with diffusion and flow models.
Causal abstraction and transportability: Distributionally Robust Causal Abstractions (ICML 2026) and Generalised Transportability via Causal Abstractions (2026) develop theory and algorithms for robust causal modelling and transferring causal conclusions across environments and populations.
Efficient foundation-model training: DES-LOC (ICLR 2026) develops adaptive optimizers that reduce communication costs while preserving convergence guarantees.
Continual learning: The Ideal Continual Learner (ICML 2023) provides a theoretical framework for learning sequential tasks without catastrophic forgetting.

Lab RAISE Lab

Workshop ReALM–GEN @ ICLR 2026

Seminar FAIS Warwick

RAISE Lab develops foundations and algorithms for Reliable, Adaptive, and Self-Improving AI. Our research brings together optimization and structured representations, generative-model alignment, continual adaptation, and AI research agents. Learn more about the lab.

I co-organized the ReALM–GEN workshop at ICLR 2026, bringing together researchers working on constrained and preference-aligned diffusion and flow-based generative models. Workshop material is available on the ICLR virtual platform.

I also co-organize the Foundations of AI Seminar series at Warwick. Prospective speakers are welcome to get in touch.

I welcome enquiries from prospective students and collaborators interested in optimization and structured learning, generative models, continual adaptation, or AI research agents. Please contact me to discuss potential research opportunities.

news

Oct 9, 2026	I’m delighted that our paper, Align & Invert, has been accepted to NeurIPS 2026! We use representation alignment at inference time to guide image reconstruction with diffusion and flow-based models. Congratulations to my PhD student, Loukas Sfountouris, on his first PhD paper, and many thanks to Giannis Daras for the collaboration!
Sep 11, 2026	I was delighted to give an invited talk, “Semantic Alignment for Robust Representations in Deep Generative Models,” at the MARS Annual Symposium at Lancaster University on 11 September 2026. Many thanks to the organizers for the invitation!
Aug 16, 2026	Our new paper, Generalised Transportability via Causal Abstractions, uses causal abstraction to study how causal conclusions can transfer across populations. We provide quantitative guarantees for approximate transportability, including settings without target data.
Jun 24, 2026	I was delighted to co-organize the BMVA Symposium on Continual AI: Learning to Adapt in a Continuous World, held in London on June 24, 2026! Check the website for the videos of the talks!
Mar 18, 2026	I am happy to be serving as an Area Chair for NeurIPS 2026!

selected publications

arXiv

AI Research Preference Models

Thomas Simon Foster, Bassel Al Omari, Tingchen Fu, and 30 more authors

arXiv preprint arXiv:2608.13940 2026

Bib PDF

@article{foster2026ai,
  title = {{AI} Research Preference Models},
  author = {Foster, Thomas Simon and Al Omari, Bassel and Fu, Tingchen and Mann, Thomas and Domond, Carl and Cipolina-Kun, Lucia and Gauri, Bhavul and Aghamelu, Muna and Goldie, Alexander D. and Helenowski, Eryk and Gagnon-Audet, Jean-Christophe and Pepe, Alberto and Nazir, Saba and Izcovich, Daniel and Levi, Noam and Hazra, Rishi and Hambardzumyan, Karen and Baldwin, Nicolas and Li, Xian and Josifoski, Martin and Giampouras, Paris and Jalili Sabet, Masoud and Sims, Anya and Momand, Hela and Shavrina, Tatiana and Magka, Despoina and Weston, Jason and Wang, Yulin and Goyal, Anirudh and Henriques, Jo{\~a}o and Bachrach, Yoram and McMilin, Emily and Foerster, Jakob Nicolaus},
  journal = {arXiv preprint arXiv:2608.13940},
  year = {2026},
  doi = {10.48550/arXiv.2608.13940},
  eprint = {2608.13940},
  archiveprefix = {arXiv},
}

arXiv

Generalised Transportability via Causal Abstractions

Yorgos Felekis, Paris Giampouras, Fabio Massimo Zennaro, and 1 more author

arXiv preprint arXiv:2608.15645 2026

Bib PDF

@article{felekis2026transportability,
  title = {Generalised Transportability via Causal Abstractions},
  author = {Felekis, Yorgos and Giampouras, Paris and Zennaro, Fabio Massimo and Damoulas, Theodoros},
  journal = {arXiv preprint arXiv:2608.15645},
  year = {2026},
  doi = {10.48550/arXiv.2608.15645},
  eprint = {2608.15645},
  archiveprefix = {arXiv},
}

ICML

Distributionally Robust Causal Abstractions

Yorgos Felekis, Theodoros Damoulas, and Paris Giampouras

2026

Bib PDF

@inproceedings{felekis2025distributionally,
  title = {Distributionally Robust Causal Abstractions},
  author = {Felekis, Yorgos and Damoulas, Theodoros and Giampouras, Paris},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year = {2026},
  volume = {306},
  series = {Proceedings of Machine Learning Research},
  pages = {29930--29980},
  publisher = {PMLR},
}

ICLR

DES-LOC: Desynced Low Communication Adaptive Optimizers for Foundation Models

Alex Iacob, Lorenzo Sani, Mher Safaryan, and 8 more authors

2026

Bib PDF

@inproceedings{iacob2025loc,
  title = {{DES-LOC}: Desynced Low Communication Adaptive Optimizers for Foundation Models},
  author = {Iacob, Alex and Sani, Lorenzo and Safaryan, Mher and Giampouras, Paris and Horv{\'a}th, Samuel and Jovanovi{\'c}, Andrej and Kurmanji, Meghdad and Aleksandrov, Preslav and Shen, William F. and Qiu, Xinchi and Lane, Nicholas D.},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=6N2qFixxYZ},
}

NeurIPS

Align & Invert: Solving Inverse Problems with Diffusion and Flow-based Models via Representation Alignment

Loukas Sfountouris, Giannis Daras, and Paris Giampouras

2026

Bib PDF

@inproceedings{sfountouris2025align,
  title = {Align \& Invert: Solving Inverse Problems with Diffusion and Flow-based Models via Representation Alignment},
  author = {Sfountouris, Loukas and Daras, Giannis and Giampouras, Paris},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2026},
}

arXiv

Rates of Convergence of Generalised Variational Inference Posteriors under Prior Misspecification

Terje Mildner, Paris Giampouras, and Theodoros Damoulas

arXiv preprint arXiv:2510.03109 2025

Bib PDF

@article{mildner2025rates,
  title = {Rates of Convergence of Generalised Variational Inference Posteriors under Prior Misspecification},
  author = {Mildner, Terje and Giampouras, Paris and Damoulas, Theodoros},
  journal = {arXiv preprint arXiv:2510.03109},
  year = {2025},
  doi = {10.48550/arXiv.2510.03109},
  eprint = {2510.03109},
  archiveprefix = {arXiv},
}

ICML

Guarantees of a Preconditioned Subgradient Algorithm for Overparameterized Asymmetric Low-rank Matrix Recovery

Paris Giampouras, HanQin Cai, and René Vidal

2025

Bib PDF

@inproceedings{giampouras2024guarantees,
  title = {Guarantees of a Preconditioned Subgradient Algorithm for Overparameterized Asymmetric Low-rank Matrix Recovery},
  author = {Giampouras, Paris and Cai, HanQin and Vidal, Ren{\'e}},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning},
  year = {2025},
  volume = {267},
  series = {Proceedings of Machine Learning Research},
  pages = {19458--19479},
  publisher = {PMLR},
}

ICML

Federated Generalised Variational Inference: A Robust Probabilistic Federated Learning Framework

Terje Mildner, Oliver Hamelijnck, Paris Giampouras, and 1 more author

2025

Bib PDF

@inproceedings{mildner2025federated,
  title = {Federated Generalised Variational Inference: A Robust Probabilistic Federated Learning Framework},
  author = {Mildner, Terje and Hamelijnck, Oliver and Giampouras, Paris and Damoulas, Theodoros},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning},
  year = {2025},
  volume = {267},
  series = {Proceedings of Machine Learning Research},
  pages = {44134--44174},
  publisher = {PMLR},
  note = {Spotlight presentation},
}

ICML

The Ideal Continual Learner: An Agent That Never Forgets

Liangzu Peng, Paris Giampouras, and René Vidal

2023

Bib PDF

@inproceedings{peng2023ideal,
  title = {The Ideal Continual Learner: An Agent That Never Forgets},
  author = {Peng, Liangzu and Giampouras, Paris and Vidal, Ren{\'e}},
  booktitle = {Proceedings of the 40th International Conference on Machine Learning},
  year = {2023},
  volume = {202},
  series = {Proceedings of Machine Learning Research},
  pages = {27585--27610},
  publisher = {PMLR},
}

ICLR
Implicit Bias of Projected Subgradient Method Gives Provable Robust Recovery of Subspaces of Unknown Codimension

Paris Giampouras, Benjamin Haeffele, and Rene Vidal

2022

Abs Bib

Robust subspace recovery (RSR) is the problem of learning a subspace from sample data points corrupted by outliers. Dual Principal Component Pursuit (DPCP) is a robust subspace recovery method that aims to find a basis for the orthogonal complement of the subspace by minimizing the sum of the distances of the points to the subspaces subject to orthogonality constraints on the basis. Prior work has shown that DPCP can provably recover the correct subspace in the presence of outliers as long as the true dimension of the subspace is known. In this paper, we show that if the orthogonality constraints –adopted in previous DPCP formulations– are relaxed and random initialization is used instead of spectral one, DPCP can provably recover a subspace of \emphunknown dimension. Specifically, we propose a very simple algorithm based on running multiple instances of a projected sub-gradient descent method (PSGM), with each problem instance seeking to find one vector in the null space of the subspace. We theoretically prove that under mild conditions this approach succeeds with high probability. In particular, we show that 1) all of the problem instances will converge to a vector in the nullspace of the subspace and 2) the ensemble of problem instance solutions will be sufficiently diverse to fully span the nullspace of the subspace thus also revealing its true unknown codimension. We provide empirical results that corroborate our theoretical results and showcase the remarkable implicit rank regularization behavior of the PSGM algorithm that allows us to perform RSR without knowing the subspace dimension
@inproceedings{giampouras2022implicit, title = {Implicit Bias of Projected Subgradient Method Gives Provable Robust Recovery of Subspaces of Unknown Codimension}, author = {Giampouras, Paris and Haeffele, Benjamin and Vidal, Rene}, booktitle = {International Conference on Learning Representations}, year = {2022}, url = {https://openreview.net/forum?id=vA7doMdgi75}, }
IEEE TSP
Block-Term Tensor Decomposition Model Selection and Computation: The Bayesian Way

Paris Giampouras, Athanasios A. Rontogiannis, and Eleftherios Kofidis

IEEE Transactions on Signal Processing 2022

Abs Bib

The so-called block-term decomposition (BTD) tensor model, especially in its rank- (Lr,Lr,1) version, has been recently receiving increasing attention due to its enhanced ability of representing systems and signals that are composed of blocks of rank higher than one, a scenario encountered in numerous and diverse applications. Uniqueness conditions and fitting methods have thus been thoroughly studied. Nevertheless, the challenging problem of estimating the BTD model structure, namely the number of block terms, R , and their individual ranks, Lr , has only recently started to attract significant attention, mainly through regularization-based approaches which entail the need to tune the regularization parameter(s). In this work, we build on ideas of sparse Bayesian learning (SBL) and put forward a fully automated Bayesian approach. Through a suitably crafted multi-level hierarchical probabilistic model, which gives rise to heavy-tailed prior distributions for the BTD factors, structured sparsity is jointly imposed. Ranks are then estimated from the numbers of blocks ( R ) and columns ( Lr ) of non-negligible energy. Approximate posterior inference is implemented, within the variational inference framework. The resulting iterative algorithm completely avoids hyperparameter tuning, which is a significant defect of regularization-based methods. Alternative probabilistic models are also explored and the connections with their regularization-based counterparts are brought to light with the aid of the associated maximum a-posteriori (MAP) estimators. We report simulation results with both synthetic and real-word data, which demonstrate the merits of the proposed method in terms of both rank estimation and model fitting as compared to state-of-the-art relevant methods.
@article{9735405, title = {Block-Term Tensor Decomposition Model Selection and Computation: The Bayesian Way}, author = {Giampouras, Paris and Rontogiannis, Athanasios A. and Kofidis, Eleftherios}, journal = {IEEE Transactions on Signal Processing}, year = {2022}, volume = {70}, pages = {1704--1717}, doi = {10.1109/TSP.2022.3159029}, }
ICML
Reverse Engineering \ell_p attacks: A block-sparse optimization approach with recovery guarantees

Darshan Thaker, Paris Giampouras, and René Vidal

2022

Abs Bib PDF

Deep neural network-based classifiers have been shown to be vulnerable to imperceptible perturbations to their input, such as \ell_p-bounded norm adversarial attacks. This has motivated the development of many defense methods, which are then broken by new attacks, and so on. This paper focuses on a different but related problem of reverse engineering adversarial attacks. Specifically, given an attacked signal, we study conditions under which one can determine the type of attack (\ell_1, \ell_2 or \ell_∞) and recover the clean signal. We pose this problem as a block-sparse recovery problem, where both the signal and the attack are assumed to lie in a union of subspaces that includes one subspace per class and one subspace per attack type. We derive geometric conditions on the subspaces under which any attacked signal can be decomposed as the sum of a clean signal plus an attack. In addition, by determining the subspaces that contain the signal and the attack, we can also classify the signal and determine the attack type. Experiments on digit and face classification demonstrate the effectiveness of the proposed approach.
@inproceedings{pmlr-v162-thaker22a, title = {Reverse Engineering $\ell_p$ attacks: A block-sparse optimization approach with recovery guarantees}, author = {Thaker, Darshan and Giampouras, Paris and Vidal, Ren{\'e}}, booktitle = {Proceedings of the 39th International Conference on Machine Learning}, editor = {Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le and Szepesvari, Csaba and Niu, Gang and Sabato, Sivan}, year = {2022}, volume = {162}, series = {Proceedings of Machine Learning Research}, pages = {21253--21271}, publisher = {PMLR}, note = {Darshan Thaker and Paris Giampouras contributed equally.}, }