Overview of Recent Scientific Research

Recent Research Topics and Selected Publications

Towards Effective Multi-Agent Communication 

This research explores how pragmatic reasoning can improve communication strategies in collaborative reinforcement learning (RL) and large language model (LLM)-based agents. We build on the Rational Speech Act (RSA) framework, which models communication as recursive social reasoning: speakers select messages that are informative from the listener’s perspective, while listeners interpret messages by considering the speaker’s goals.

We extend this to the Cooperative RSA (CRSA) framework, which adapts RSA to multi-agent, cooperative settings. By modeling nested beliefs and communicative goals across agents, CRSA provides a principled way to design effective communication in multi-turn dialogue scenarios. This line of research aims to bridge theory-driven pragmatic reasoning with modern RL and LLM-based systems to enable richer, more efficient collaboration among agents.

Security and Privacy of Machine Learning Models

Security and reliability involve detecting AI-generated content, which can be misused to misinform, manipulate, or impersonate. Our research explores methods to identify synthetic text or media based on information-theoretic measures, helping systems recognize when content may be machine-generated.

Security and reliability involve detecting errors. AI models can make classification or regression errors, and may be overconfident even on unfamiliar inputs. Our research introduces Relative Uncertainty (REL-U), a data-driven measure that identifies misclassifications and large regression errors by learning patterns in model predictions. REL-U helps AI systems know when they might be wrong, improving safety and trustworthiness in applications like autonomous systems and medical AI. 

Embodied AI agents interact closely with humans, collecting sensitive data from our environments, habits, and communications. Protecting this data requires understanding fundamental limits and certification processes, while careful system design ensures agents only use what is necessary.

How to Evaluate and Rank Embedders

Embedders transform objects—texts, images, molecules—into numerical representations for downstream tasks. Choosing the best embedder is challenging, as traditional evaluations rely on costly, task-specific benchmarks.

We propose a task-agnostic, self-supervised framework based on information sufficiency, which estimates how well one embedder preserves information relative to another. This allows ranking embedders without labeled data and predicting their downstream performance.

Information Theory in Machine Learning

AI systems often struggle with subtle changes in input data and with efficiently capturing the information content in complex datasets. FIRE uses concepts from information geometry to make AI more robust: it measures the “distance” between predictions for original and slightly altered inputs using the Fisher-Rao Distance, ensuring the AI behaves consistently and reliably.

Meanwhile, understanding uncertainty and information is crucial for AI. KNIFE is a flexible, differentiable estimator of differential entropy and mutual information, allowing AI to adapt to changing data distributions and learn more effectively across tasks like domain adaptation, fair classification, and language model fine-tuning.

Overview of Previous Scientific Contributions