Publications

You can also find my articles on my Google Scholar profile.

Pythagoras-Prover: Advancing Efficient Formal Proving via Augmented Lean Formalisation Permalink

arxiv preprint 2606.12594, 2026

We introduce Pythagoras-Prover, a compute-efficient open-source family of Lean theorem provers built for practical compute budgets. The family spans two generation paradigms: autoregressive models at 4B and 32B parameters, and a first proof-of-concept diffusion-based prover (4B) that iteratively refines Lean proofs at inference time. Empirically, Pythagoras-Prover-4B surpasses DeepSeek-Prover-V2-671B at pass@32 on MiniF2F-Test (86.1% vs 82.4%) with ~167x fewer parameters, while Pythagoras-Prover-32B sets the open-source state of the art at 93.0% on MiniF2F-Test and solves 93 of 672 PutnamBench problems. Read more

Joshua Ong Jun Leang, Zheng Zhao, Mihaela Cătălina Stoian, Qiyuan Xu, Haonan Li, Wenda Li, Shay B. Cohen, Eleonora Giunchiglia. Pythagoras-Prover: Advancing Efficient Formal Proving via Augmented Lean Formalisation. arXiv preprint 2606.12594, 2026

Can I Have Your Order? Monte-Carlo Tree Search for Slot Filling Ordering in Diffusion Language Models Permalink

In The Proceedings of ICML, 2026

While plan-and-infill decoding in Masked Diffusion Models (MDMs) shows promise for mathematical and code reasoning, performance remains highly sensitive to slot infilling order, often yielding substantial output variance. We introduce McDiffuSE, a framework that formulates slot selection as decision making and optimises infilling orders through Monte Carlo Tree Search. Read more

Joshua Ong Jun Leang, Yu Zhao, Mihaela C. Stoian, Wenda Li, Shay B. Cohen, Eleonora Giunchiglia. Can I Have Your Order? Monte-Carlo Tree Search for Slot Filling Ordering in Diffusion Language Models. In Proceedings of International Conference on Machine Learning (ICML), 2026.

A Survey on Deep Learning Approaches for Tabular Data Generation: Utility, Alignment, Fidelity, Privacy, Diversity, and Beyond Permalink

In Transactions on Machine Learning Research, 2026

In this survey, we review deep generative modelling approaches for tabular data from the perspective of five types of requirements: utility of the synthetic data, alignment of the synthetic data with domain-specific knowledge, statistical fidelity of the synthetic data distribution compared to the real data distribution, privacy-preserving capabilities, and sampling diversity. We group the approaches along two levels of granularity: (i) based on the requirements they address and (ii) according to the underlying model they utilise. Read more

Mihaela C. Stoian, Eleonora Giunchiglia, Thomas Lukasiewicz. A Survey on Deep Learning Approaches for Tabular Data Generation: Utility, Alignment, Fidelity, Privacy, Diversity, and Beyond. Transactions on Machine Learning Research, 2026.

Beyond the Convexity Assumption: Realistic Tabular Data Generation under Quantifier-Free Real Linear Constraints Permalink

In The Proceedings of ICLR, 2025

In this paper, we introduce the Disjunctive Refinement Layer (DRL), a novel layer designed to enforce the alignment of generated data with the background knowledge specified in user-defined constraints. DRL is the first method able to automatically make deep learning models inherently compliant with constraints as expressive as quantifier-free linear formulas, which can define non-convex and even disconnected spaces. Read more

Mihaela C. Stoian and Eleonora Giunchiglia. Beyond the Convexity Assumption: Realistic Tabular Data Generation under Quantifier-Free Real Linear Constraints. In Proceedings of International Conference on Learning Representations (ICLR) 2025.

How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data Permalink

In The Proceedings of ICLR, 2024

In this paper, we show how deep generative models for tabular data can be constrained such that their generated samples are guaranteed to be compliant with given constraints. This is achieved by automatically parsing the constraints and transforming them into a Constraint Layer seamlessly integrated with the model. Read more

Mihaela C. Stoian, Salijona Dyrmishi, Maxime Cordy, Thomas Lukasiewicz, Eleonora Giunchiglia. How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data. In Proceedings of International Conference on Learning Representations (ICLR) 2024.

PiShield: A PyTorch Package for Learning with Requirements Permalink

In The Proceedings of IJCAI, 2024

In this paper, we introduce PiShield, the first package ever allowing for the integration of (propositional or linear) requirements into the neural networks’ topology. PiShield guarantees compliance with these requirements, regardless of input. Read more

Mihaela C. Stoian, Alex Tatomir, Thomas Lukasiewicz, Eleonora Giunchiglia. PiShield: A PyTorch Package for Learning with Requirements. In Proceedings of International Joint Conference on Artificial Intelligence (IJCAI), 2024.

Deep generative models as an adversarial attack strategy for tabular machine learning Permalink

In The Proceedings of International Conference on Machine Learning and Cybernetics, 2024

In this paper, we adapt popular tabular deep generative models (DGMs) into adversarial DGMs and evaluate their effectiveness in generating realistic adversarial examples that conform to domain constraints. Read more

Salijona Dyrmishi, Mihaela C. Stoian, Eleonora Giunchiglia, Maxime Cordy. Deep generative models as an adversarial attack strategy for tabular machine learning. In Proceedings of International Conference on Machine Learning and Cybernetics 2024.

CCN+: A neuro-symbolic framework for deep learning with requirements Permalink

In International Journal of Approximate Reasoning, 2024

In this paper, we propose a novel neuro-symbolic framework able to make any neural network compliant by design to a given set of requirements over the output space expressed in full propositional logic. This framework, called CCN+, integrates the requirements into the output layer of the neural network by applying multiple inference rules that ensure compliance with the requirements and adapts the standard binary cross-entropy loss function to the requirement output layer. Read more

Eleonora Giunchiglia‚ Alex Tatomir‚ Mihaela C. Stoian, Thomas Lukasiewicz. CCN+: A neuro-symbolic framework for deep learning with requirements. International Journal of Approximate Reasoning, 171, 109-124 (2024).

Deep Learning with Requirements in the Real World Permalink

In The Proceedings of IJCAI, 2024

In this position paper, I discuss my research on integrating requirements into neural networks to guide the learning process and ultimately produce outputs that ensure the requirements’ satisfaction for two real-world applications: tabular data generation and autonomous driving. Read more

Mihaela C. Stoian. Deep Learning with Requirements in the Real World. In Proceedings of International Joint Conference on Artificial Intelligence (IJCAI), 2024, Doctoral Consortium.

ROAD-Waymo: Action Awareness at Scale for Autonomous Driving Permalink

arxiv preprint 2411.01683, 2024

In this paper, we introduce ROAD-Waymo, an extensive dataset for the development and benchmarking of techniques for agent, action, location and event detection in road scenes, provided as a layer upon the (US) Waymo Open dataset. As ROAD-Waymo is compatible with the original (UK) ROAD dataset, it provides the opportunity to tackle domain adaptation between real-world road scenarios in different countries within a novel benchmark: ROAD++. Read more

Salman Khan, Izzeddin Teeti, Reza Javanmard Alitappeh, Mihaela C. Stoian, Eleonora Giunchiglia, Gurkirt Singh, Andrew Bradley, Fabio Cuzzolin. ROAD-Waymo: Action Awareness at Scale for Autonomous Driving. arXiv preprint 2411.01683, 2024

Exploiting T-norms for Deep Learning in Autonomous Driving Permalink

In The Proceedings of NeSy, 2023

In this paper, we show how it is possible to define memory-efficient t-norm-based losses, allowing for exploiting t-norms for the task of event detection in autonomous driving. Read more

Mihaela C. Stoian, Eleonora Giunchiglia, Thomas Lukasiewicz. Exploiting T-norms for Deep Learning in Autonomous Driving. In Proceedings of the 17th International Workshop on Neural-Symbolic Learning and Reasoning (NeSy), 2023.

ROAD-R: The Autonomous Driving Dataset with Logical Requirements Permalink

In Machine Learning, 2023

In this paper, we introduce the ROad event Awareness Dataset with logical Requirements (ROAD-R), the first publicly available dataset for autonomous driving with requirements expressed as logical constraints. The paper was presented at IJCLR 2022, where it received the Best Student Paper Prize, and at IJCAI 2022 Workshop on Artificial Intelligence for Autonomous Driving, where it received the Best Paper Award. Read more

Eleonora Giunchiglia, Mihaela C. Stoian, Salman Khan, Fabio Cuzzolin, Thomas Lukasiewicz. ROAD-R: The Autonomous Driving Dataset with Logical Requirements. Machine Learning, 112, 3261–3291 (2023).

Deep Learning with Logical Constraints Permalink

In The Proceedings of IJCAI, Survey Track, 2022

In this survey, we retrace works which exploit logically specified background knowledge in order to obtain neural models (i) with a better performance, (ii) able to learn from less data, and/or (iii) guaranteed to be compliant with the background knowledge itself, e.g., for safety-critical applications. Read more

Eleonora Giunchiglia, Mihaela C. Stoian, Thomas Lukasiewicz. Deep Learning with Logical Constraints. In Proceedings of International Joint Conference on Artificial Intelligence (IJCAI), 2022.

Recurrently Estimating Reflective Symmetry Planes from Partial Pointclouds Permalink

In CVPR 2021 Workshop on 3D Vision and Robotics, 2021

In this paper we present a novel method to estimate planar reflective symmetries that efficiently handles 3D inputs by slicing the data along the height dimension and passing it sequentially to a 2D convolutional recurrent regression scheme. Read more

Mihaela C. Stoian, Tommaso Cavallari. Recurrently Estimating Reflective Symmetry Planes from Partial Pointclouds. In Conference on Computer Vision and Pattern Recognition (CVPR) Workshop on 3D Vision and Robotics, 2021.

Analyzing ASR pretraining for low-resource speech-to-text translation Permalink

In The Proceedings of ICASSP, 2020

In this paper we explore what factors help pretraining for low-resource automatic speech-to-text translation (AST). We show that the word error rate (WER) of the pre-trained automatic speech recognition (ASR) models is likely the best direct predictor of AST performance. Additionally, our analysis suggests that the models with better WER are transparently encoding more language-universal phonetic information in the later RNN layers, and this appears to help with AST. Read more

Mihaela C. Stoian, Sameer Bansal, Sharon Goldwater. Analyzing ASR pretraining for low-resource speech-to-text translation. In Proceedings of International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2020.