Hi! I’m Yuriel (or Ryan, if you prefer), an M.Eng (Research) student at the Singapore University of Technology and Design, supported by the AI Singapore Accelerated Masters Scholarship and the DSO-AISG Research Award. I publish as “Yuriel Ryan.”

Research Interests

Broadly speaking, I am interested in understanding how task-relevant information flows within complex systems — from deep neural networks to multi-agent environments. This means building on theoretical frameworks, e.g., from information theory, to quantify these task-relevant signals or determine empirical bounds.

I think of “information flow” as having directions. A system can move information forward, using what it knows for inference and action; it can also flow backwards, revising what it knows in order to learn. Much of the difficulty, and most of what I find interesting, lies in optimizing when to switch directions in label-free settings while accounting for emergent signals — information that only surfaces when multiple sources interact.

So far, my published work has mostly looked at the forward direction — what components of the data form the task-relevant signals (Humor in Pixels: benchmarking humor comprehension as a social signal), and how to exploit them (Multimodal Interaction Tuning: amplifying redundant information to train more robust vision-language models).

What do I want to work on next?

I am excited to work on other directions that information can flow in as well as the decision of when to optimally switch between them. For instance, thinking about augmented LMs (e.g., RAG / tool calling), world modelling, or meta systems through the different lenses of information theory (e.g., usable information). Of particular interest, I’m exploring ideas of collective intelligence and co-evolving systems, where a population of agents improves themselves by interacting with the world and sharing what they have learned — through perhaps an episodic memory — to other agents. In this manner, information not only flows forward / backward, but also laterally to other agents!

A couple of works that really inspired me: CORAL (multi-agent self-evolution), Colony (a film — the zombies exhibit collective intelligence!), R-Zero (Self-evolving agent)

While this relatively new direction of AI research is exciting, I’m also mindful of the social and practical implications of these “evolving agents/systems”. My intended goal is to build these adaptive systems to collaborate with humans — and definitely not to deprecate us! (see this position paper) Ultimately, I hope to actualize the Human(s)-AI(s) symbiosis ideal (note that both human and AI can be plural) through this collective intelligence approach where both sides continuously refine each other.

Updates!