Department of Computer Science

Research

---

Explainable AI & LLM Reasoning

The field of mechanistic interpretability concerns itself with "reverse engineering" deep artificial neural networks in order to understand them in terms of human-interpretable algorithms. Our prior work focused on studying concept acquisition in deep reinforcement learning (deepRL) systems such as DeepMind's AlphaZero, with the goal of identifying the acquisition of given human-like concepts by these systems. Our current research aims to expand the scope of study to large language models (LLMs) and other deep learning models, while also aiming to isolate novel concepts potentially beyond human current understanding. Of particular interest to us are concepts related to reasoning processes, and their potential relationship with human-like logical reasoning.

 

Selected publications

1. Marchiafava, A., & Sen, A. (2025). Towards Determining How Deep Neural Models Learn to Reason. Proceedings of the AAAI Symposium Series, 5(1), 370-373. https://doi.org/10.1609/aaaiss.v5i1.35614
2. Marchiafava and A. Sen, “Comparing Concept Acquisition in Human & Superhuman DeepRL Models,” International Journal On Advances in Intelligent Systems, vol. 16, no. 3, pp. 19–30, Dec. 30, 2023, issn: 1942-2679. url: https://ns2.thinkmind.org/index.php?view=article&articleid=intsys_v16_n34_2023_1(visited on 06/04/2024).

Researchers:

Anthony Marchiafava

---

Biodiversity AI

We are interested in the application of interpretable AI techniques to biodiversity science, ecological informatics and computational systematics generally, an intersection we call "Biodiversity AI". Our current research involves studying the impact of taxonomic and geo-spatial change using machine learning and computational reasoning, with the goal of impacting conservation decision-making. Our Biodiversity AI research is funded by the National Science Foundation.

 

Selected publications

1. A. Sen, B. Sterner, N. Franz, C. Powel, and N. Upham, “Combining Machine Learning & Reasoning for Biodiversity Data Intelligence,” en, in *Proceedings of the AAAI Conference on Artificial Intelligence*, Number: 17, vol. 35, May 2021, pp. 14 911–14 919. doi: 10.1609/aaai.v35i17.17750. (visited on 07/05/2023).

Researchers:

Dr. Medrano-Vizcaíno, Anthony Marchiafava

---

• 1. Knowledge-Guided Machine Learning (KGML) & Multi-Modal Machine Learning:

  1A. Pileup Mitigation in High Energy Physics:

  This research, supported by the U.S. Department of Energy, addresses the significant challenge of pileup in high-energy physics experiments at CERN's Large Hadron Collider (LHC). Pileup arises when multiple proton-proton interactions occur almost simultaneously within the detectors, superimposing background noise onto the primary collision data and thereby distorting crucial measurements. To counter this, our lab is developing several distinct AI-driven methodologies. These include the application of attention-based neural networks designed to analyze entire collision events, which allows for a more accurate correction of the measured properties (such as energy and mass fractions) of particle jets by considering the overall event context. Another strategy involves designing denoising autoencoders for precise track-level identification, enabling the targeted removal of the numerous tracks originating from pileup events to isolate the critical, less frequent tracks from hard scatter (signal) interactions. Furthermore, we are exploring advanced ML techniques such as contrastive learning and knowledge distillation. As pileup mitigation is an open-ended and complex problem with limited established solutions, our work aims to make substantial improvements in data clarity, thereby enhancing the potential for new discoveries in fundamental physics.
  
  1B. Soil Moisture Estimation in Croplands:

  This USDA-funded project addresses the critical challenge of accurately estimating soil moisture, a key factor in determining crop water stress for agricultural applications. From a computer science standpoint, we are developing multimodal DL models that integrate visual data from cropland soil patch images with corresponding meteorological data. This information is gathered from a network of six monitoring stations across Oklahoma. Our work involves creating various frameworks employing DL architectures for precise feature extraction from these distinct data types. In multimodal learning, where features from multiple modalities are commonly fused to produce a single output, the imbalance problem represents a key concern that can affect predictive accuracy if one modality unduly influences the outcome. A significant component of this research is therefore dedicated to developing and implementing techniques that effectively address this imbalance within our feature fusion methodologies, aiming to optimize how information from images and weather data is combined. The goal is to generate more accurate soil moisture predictions, which in turn inform crop water stress estimates, directly contributing to improved water resource management and supporting sustainable farming practices.
  
  1C. Sentiment Reasoning:

  We are developing an event‐driven causal sentiment framework that traces how feelings propagate through Twitter reply trees. By embedding each tweet with a fine‐tuned language model and using a time‐aware graph neural network, we enforce that sudden sentiment shifts only occur when a parent or ancestor mentions a key event. Moving forward, we plan to extend this approach to multimodal inputs—incorporating images, URLs, and network engagement signals—and to validate our causal estimates through small‐scale interventions. Ultimately, we aim to build a unified system that can explain ‘why’ conversations turn positive or negative across different social platforms and event types.

  Researchers:

  Mohammed Rakib, Rifat Rafiuddin

• 2. Political Bias in LLMs:

  Our research focuses on identifying and interpreting political and cultural bias in Large Language Models (LLMs). We combine statistical measures with model analysis to understand how such biases are encoded and expressed. We aim to build robust methods for investigating the mechanisms through which ideological patterns emerge in LLMs, and how they can potentially be mitigated through better LLM reasoning.

  Researchers:

  Sadia Kamal