Export
Gabriel Stella
· 1 previous affiliation
- Texas A&M UniversitySince 2020
- PhD student · Since 2020
Research Overview
Gabriel Stella focuses on the intersection of reinforcement learning and symbolic reasoning, with a particular emphasis on object-oriented representations and relational induction. His work addresses the limitations of deep reinforcement learning—specifically its struggles with systematic generalization and interpretability—by developing algorithms that learn fundamental world models through the induction of first-order logic rules. Currently, his research explores how agents can construct expressive, human-readable models of their environments to facilitate zero-shot transfer and efficient learning in complex, stochastic domains.
Research Trajectory
Stella’s academic trajectory is characterized by a transition from the technical foundations of data processing to the conceptual challenges of artificial intelligence. His early work focused on the theory and algorithms underlying high-performance Big Data systems, specifically addressing problems in external-memory streaming and I/O scheduling. This phase of his research established a grounding in computational efficiency and the structural requirements of large-scale data handling, which later informed his approach to the computational limits of intelligent systems.
In recent years, Stella has shifted his focus toward the mechanisms of intelligence, moving from data-centric systems to agent-centric models. This shift was motivated by a foundational interest in how sophisticated behaviors emerge from simple, effective representations of the world. His current work operates within the framework of Automated Rule Induction for Neuro-Symbolic AI, where he seeks to combine the intuitive pattern recognition of neural networks with the structured reasoning of symbolic logic. By building functional models in the form of computer algorithms and evaluating them through carefully designed games, Stella investigates the conditions under which intelligence and social cooperation arise.
By 2024, this trajectory culminated in the development of methods for Interpretable Deep Learning via Symbolic and Prototypical Bottlenecks. This work specifically targets the "trust gap" in safety-critical domains by replacing opaque "black-box" decision-making with transparent, auditable logic. His most recent contributions focus on Scaling Neuro-Symbolic Rule Induction, aiming to make these interpretable models as powerful and efficient as their purely neural counterparts.
Current Directions
Stella’s current research program is organized around three primary themes: the induction of relational world models, the pursuit of zero-shot generalization, and the exploration of emergent intelligence.
A central component of his current work is the development of object-oriented transition models. Unlike standard reinforcement learning approaches that operate on raw pixel data or flat feature vectors, Stella’s methods utilize domain-agnostic object-based state representations. This approach is rooted in the belief that human-like intelligence relies on perceiving the world as a collection of entities and their interactions. By focusing on Relative Representations for Spatially Invariant Robot Learning, his research aims to create agents that can generalize across different spatial configurations by understanding the relative positions and roles of objects rather than relying on absolute coordinates.
Furthermore, Stella is investigating the theoretical limits of intelligence. This includes questioning whether the emergence of intelligence is an inevitable consequence of certain computational models of the universe and under what specific environmental conditions social, cooperating agents are likely to evolve. This line of inquiry is highly analytical, treating the universe itself as a computational substrate and using virtual worlds as laboratories to test these recursive hypotheses.
While his recent publications demonstrate a strong focus on neuro-symbolic reinforcement learning, the evidence regarding his earlier work in Big Data systems is less detailed in the current corpus. However, the influence of high-performance computing remains visible in his emphasis on the training efficiency and algorithmic power of his recent machine learning models.
Representative Work
Stella’s primary contributions to the field are centered on the QORA algorithm and its subsequent refinements in inductive logic programming. As the lead researcher on these projects, he has established a framework for learning interpretable models that outperform standard neural baselines in both speed and generalization.
In QORA: Zero-Shot Transfer via Interpretable Object-Relational Model Learning, Stella introduced the Quantified Object Relation Aggregator (QORA). This algorithm addresses the failure of deep reinforcement learning models to generalize to unseen environments. QORA constructs models expressive enough to solve a variety of domains, including those with stochastic transition functions, by learning first-order logic rules from object-based observations. The significance of this work lies in its ability to achieve zero-shot transfer—perfectly generalizing to novel environments without additional training—while producing human-readable logic. Stella demonstrated that this symbolic approach could learn significantly faster than state-of-the-art neural network baselines, effectively bridging the gap between efficiency and interpretability.
Building on the foundations of QORA, Stella’s work in Object-Oriented Transition Modeling with Inductive Logic Programming explores the challenge of induction in machine learning. This research focuses on building models of the world from observation that maintain accuracy in novel situations. By utilizing inductive logic programming within an object-oriented framework, Stella developed a learning algorithm that offers substantial improvements over previous methods. The work includes rigorous ablation tests and comparisons with neural baselines, demonstrating that structured, symbolic induction is a viable and often superior alternative to purely connectionist approaches for world-modeling tasks.
These papers represent a coherent research program dedicated to creating intelligent systems that are not only capable of sophisticated behavior but are also fundamentally understandable and robust to environmental changes. Through these contributions, Stella has positioned himself at the forefront of the effort to integrate symbolic reasoning into the modern reinforcement learning paradigm.
Profile updated
| Title | Venue | Year |
|---|---|---|
| Object-Oriented Transition Modeling with Inductive Logic Programming | Openreview | 2026 |
| QORA: Zero-Shot Transfer via Interpretable Object-Relational Model Learning | ICML | 2024 |
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake