Martin Klissarov - MaestroMotif: Skill Design from AI Feedback
Автор: UCL DARK
Загружено: 2025-04-14
Просмотров: 219
Invited talk by Martin Klissarov, Research Scientist at Google DeepMind, on April 7, 2025 at UCL DARK.
Title:
MaestroMotif: Skill Design from AI Feedback
Abstract:
Describing skills and behaviours in natural language has the potential of providing an accessible way of injecting human knowledge about decision-making tasks into an AI system. We present MaestroMotif (ICLR 2025 Oral), a method for skill design that fundamentally embraces the human-AI paradigm, yielding high-performing and adaptable agents. Starting from a natural language description of a set of skills provided by a user, it leverages an LLM's feedback to automatically design rewards corresponding to each skill. It then builds on an LLM's code generation abilities to sequence and learn these skills. On a suite of complex tasks in the NetHack Learning Environment (NLE), MaestroMotif demonstrates that it surpasses existing approaches in both performance and usability.
Bio:
Martin Klissarov is a Research Scientist at Google DeepMind working with Prof. Ed Grefenstette in the Autonomous Assistants team. He is currently wrapping up his PhD student supervised by Prof. Doina Precup and Prof. Marlos C Machado. He works on the intersection of RL, LLMs and human-AI interactions.
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке:
![[TSLL2025] Plenary: New Conversations new evidence by Dr. Serge Bibauw (UCLouvain)](https://ricktube.ru/thumbnail/zdK2RDejQCE/mqdefault.jpg)
![Micah Carroll - Uni[MASK]: Unified Inference in Sequential Decision Problems @ UCL DARK](https://ricktube.ru/thumbnail/6grBrpSiojg/mqdefault.jpg)
