Generalization and mechanistic interpretability
We pursue a simple goal: to understand not just whether models work, but why they work, when they fail, and what they are truly relying on under the hood. Our research focuses on robust generalization under distribution shift, the emergence of spurious correlations and shortcut strategies, and the internal mechanisms that drive these behaviors. We develop methods that go beyond merely cataloging failures after the fact by revealing hidden biases in learned representations, tracing shortcut learning through embeddings and weight space, and testing whether models can transfer abstract knowledge beyond the settings in which it was first acquired.
Elena Burceanu, Antonio Barbalau, Cristian Păduraru
- Bridging Explainability and Embeddings: BEE Aware of Spuriousness (ICLR 2026)
- Not All Splits Are Equal: Rethinking Attribute Generalization Across Unrelated Categories (NeurIPS 2025 Workshop CauScien)
- Learning (Approximately) Equivariant Networks via Constrained Optimization (NeurIPS 2025)
All papers in this direction Generalization & Interpretability
Deepfake detection
Aletheia focuses on advancing deepfake detection across video and audio modalities. Our research is guided by three goals:
- Generalization: develop methods that transfer across diverse datasets and forgery techniques.
- Transparency: understand how detection models make decisions and ensure datasets are reliable (free of spurious shortcuts).
- Deployability: build systems that adapt and remain robust on unconstrained “in-the-wild” content.
Elisabeta Oneata, Dan Oneata, Ștefan Smeu, Dragoș-Alexandru Boldișor
- Investigating self-supervised representations for audio-visual deepfake detection (CVPR 2026)
- Circumventing shortcuts in audio-visual deepfake detection datasets with unsupervised learning (CVPR 2025)
- DeCLIP: Decoding CLIP representations for deepfake localization (WACV 2025)
All papers in this direction Deepfake Detection
Natural language processing
We focus on large language models, along with reliability, reasoning, and scientific machine learning, applying our work across areas such as code and low-level languages, multilingual systems, and structured domains like molecules.
Florin Brad, Andrei Manolache, Ioana Pintilie, Marius Drăgoi, Alexandra Dragomir
- JumpLoRA: Sparse Adapters for Continual Learning in Large Language Models (Under review at CoLLAs 2026)
- CLewR: Curriculum Learning with Restarts for Machine Translation Preference Learning (ACL 2026)
- C-ing Clearly: Enhanced Binary Code Explanations using C code (Under review)
All papers in this direction Natural Language Processing
Reinforcement learning
Within the field of artificial intelligence, reinforcement learning presents a natural setting for training agents that interact with the world we are living in. We engage in furthering the field by developing agents able to learn continuously and efficiently in complex and non-stationary environments.
Florin Gogianu
No academic news posts tagged with this direction yet.
All papers in this direction Reinforcement Learning
