We focus on the foundational issues which are the root causes of today’s safety related issues. For true safety, AI systems must understand what they read (not just mimic sequences of text), what they see around them, and the implications of their actions; and they must be designed with humans in mind.

Understanding

Reasoning

Neural networks still do not reason reliably. We showed conditions under which a broad new class of reasoning systems, including forms underlying some neural networks, can be confirmed to reason correctly, in Foundations of reasoning with uncertainty via real-valued logics, Fagin et al., Proceedings of the National Academy of Sciences, 2024.

Semantics

AI models still do not explictly represent meaning, or semantics — the basis needed for a true “world model”. In 1952, a seminal proposal to add meaning to information theory, one of the pillars of AI, was put forth but was never developed mathematically enough to be usable. We showed the first theoretical foundations unlocking the potential of semantic information theory, in Towards a Unification of Logic and Information Theory, Lastras et al., under journal review.

We show that input-dependent sparse attention improves convergence and generalization in transformers, while input-agnostic sparsity does not. We theoretically explain this through softmax stability and Lipschitz-based guarantees that characterize when semantic focus yields provable benefits., in Transformers Learn Faster with Semantic Focus, Ram et al., Conference on Neural Information Processing System, 2025.

Human in the Loop

Interpretability

In general, AI models that perform well are difficult for humans to interpret, and vice versa. We show a model class that can achieve both state-of-the-art classification performance yet achieves a high level of interpretability, in Neural Reasoning Networks: Efficient Interpretable Neural Networks with Automatic Textual Explanations, Carrow et al., American Association for Artificial Intelligence, 2024.

Confidence

AI systems typically do not report their confidence in their answers, which can be misleading at best, and dangerous at worst. We continue development of a framework for probabilistic reasoning which propagates error bars on its inputs to error bars on its outputs rigorously yet efficiently, in Abductive Reasoning in Logical Credal Networks, Marinescu et al., Neural Information Processing Systems, 2024.

We study directed cycles in Logical Credal Networks, showing that Directed-Undirected Mixed Graphs (DUMGs) enable Gibbs factorizations, yield multilinear-program inferences, and reveal how cycles cause probabilistic imprecision under interventions, in Dealing with cycles in graph-based probabilistic models: the case of Dealing with cycles in graph-based probabilistic models: the case of Logical Credal Networks, Cozman et al., Proceedings of Machine Learning Research (PMLR),  2025.

Toward Generalization/AGI

Compositional generalization

Reliable out-of-distribution (OOD) behavior (i.e. handling new situations that were not seen in training, not just mimicking already-seen solutions) and the potential to learn with dramatically less data are widely understood to require “compositional generalization”, but this notion in AI has remained mostly anecdotal. We provide firm definitions and foundations, in What Makes Models Compositional? A Theoretical View, Ram et al., International Joint Conference on Artificial Intelligence, 2024.

We present IVNTR, a neuro-symbolic bilevel planning framework that learns neural predicates from demonstrations to enable compositional generalization in high-dimensional robotic tasks. We show that it substantially improves generalization across simulated and real-world domains compared to prior approaches, in Bilevel Learning for Bilevel Planning, Li et al., Robotics: Science and Systems, 2025.

Causality

The ability to reason correctly about cause-and-effect (vs correlation) is crucial for AI systems, yet existing theory still only holds for certain idealized settings. We began exploring how certain probabilistic semantics can allow modeling of causality in more realistic partially identified and cyclic cases, in Markov conditions and factorization in logical credal networks, Cozman et al., International Journal of Approximate Reasoning, 2024.

We study partially identifiable queries in quasi-Markovian causal models and compute tight probability bounds when exact inference is impossible. We propose an efficient algorithm with column generation that outperforms existing methods, in Multilinear and linear programs for partially identifiable queries in quasi-markovian structural causal models, Arroyo et al., 41st Conference on Uncertainty in Artificial Intelligence, 2025.