My research spans three complementary directions: developing the theoretical foundations of Online Learning, applying these ideas to problems in Generative AI, and building reliable AI Scientist systems for scientific discovery.

• Theoretical Foundations of Online Learning: My research investigates the foundations of sequential decision-making under uncertainty. I focus on developing algorithms with rigorous theoretical guarantees, particularly for multi-armed bandits, PAC decision-making, and cost-aware learning. Some of the key challenges I address include:
  • Algorithmic Efficiency: Designing algorithms that effectively balance exploration and exploitation in uncertain and dynamic environments.
  • Theoretical Guarantees: Establishing performance bounds and correctness guarantees for problems such as fixed-confidence identification and decision-making with costly or limited feedback.
• Applications in Generative AI: I apply ideas from Online Learning to improve the performance and cost-efficiency of Large Language Models (LLMs). My main areas of interest include:
  • Optimal LLM Selection: Developing methods to identify the most suitable LLM for a given task while minimizing computational and monetary costs.
  • Adaptive Reasoning: Designing inference-time strategies that dynamically allocate computation and improve the reasoning capabilities of LLMs.
  • Prompt Optimization: Developing algorithms to identify effective prompts and demonstrations for improving the quality and reliability of generated responses.
• AI Scientists: I am also interested in developing AI systems that can assist with and automate different stages of scientific research. My current interests in this direction include:
  • Literature-Grounded Discovery: Building systems that generate research ideas and hypotheses grounded in existing scientific literature, while identifying meaningful gaps and avoiding unsupported claims.
  • Sequential Decision-Making: Formulating scientific discovery as a sequential learning problem in which an AI system adaptively selects experiments, evaluates evidence, and refines its hypotheses.
  • Interpretable Scientific Reasoning: Developing methods that make the reasoning, evidence, and decisions of AI Scientists transparent and understandable to human researchers.

Together, these directions connect the theoretical foundations of Online Learning with practical challenges in Generative AI and the emerging development of reliable AI systems for scientific discovery.

Past Works

Along with Online Learning, I previously worked in observational astronomy, focusing on the analysis of data from space archives. I mainly used two different approaches to explore various aspects of the universe:

• Statistical Analysis for Multiwavelength Astronomy: One aspect of my work involved delving into multiwavelength data, which encompasses a range of electromagnetic radiation, including X-rays, extending beyond what's visible. I used mathematical techniques to find hidden patterns and gain insights into a wide range of cosmic events and objects. This approach was particularly useful when I studied high-energy events and tried to confirm the existence of planets outside our solar system, known as exoplanets.
• Deep Learning with Astronomical Data: In addition to traditional statistical methods, I also employed advanced computer techniques like transformer models and attention mechanisms. These cutting-edge tools helped me make sense of complex data from space. With deep learning, I could extract intricate details from diverse datasets, improving our understanding of the physical processes that govern the universe. I applied this approach to delve into high-energy observations, providing deeper insights into phenomena like supernovae, gamma-ray bursts, and active galactic nuclei.

By combining statistical analysis and deep learning, I contributed to our expanding knowledge of the cosmos. I also tackled the challenges presented by the massive and varied datasets that astronomers deal with.