Seminari di Ingegneria Informatica - Leonardo Ranaldi: Adaptive Reasoning Agents - The Extent to Which Agents Should Adapt
Speaker: Leonardo Ranaldi (University of Rome Tor Vergata)
Title: Adaptive Reasoning Agents - The Extent to Which Agents Should Adapt
Abstract: Automated agents powered by LLMs are moving from answering questions to reasoning, deciding, and acting. Longer, more complex interactions make adaptation part of the reasoning process, where answers are revised, procedures modified, experience retained, and previous interactions used to steer later decisions.
The central question is how an agent learns what to change, and how experience can lead that decision.
This seminar treats adaptation as a reasoning problem. Effective adaptation requires representing what is changing, evaluating evidence for change, and anticipating consequences. The discussion centres on four capacities: representing a problem, revising a stance, consolidating experience into reusable procedures, and anticipating an action's effects. Across these capabilities, quasi-symbolic abstractions deliver a proper structure for examining, revising, and reusing them while preserving the flexibility of natural language. The seminar closes by looking ahead to how single-agent learning can evolve collective practice for a group.
Bio: Leonardo Ranaldi did his PhD at the University of Rome Tor Vergata, where he is a member of the academic staff. After a postdoc at the Idiap Research Institute, he joined the University of Edinburgh as a Research Fellow. His research focuses on adaptive reasoning, agent memory, and controllable generation, with particular interest in how AI systems learn from interaction and reuse experience over time. He works across international research collaborations and regularly organises workshops at major CL/NLP venues.
Zoom link: https://uniroma1.zoom.us/j/85485829872?pwd=DVnUcL3bP6PammzbKe4htvWabslRw5.1