Title of the talk: Challenges in long horizon open ended learning
Speaker: Dr. Vihang Patil.
Venue: A-LH2
Date & Time: Wednesday, 21 May 2025, 11:00 am to 12:00pm.
Abstract:
Long-horizon, open-ended learning remains one of the core challenges in scaling intelligent systems. As agents operate over extended timescales and across diverse environments, they must reason over delayed consequences, assign credit effectively, and model complex, multimodal trajectories. In this talk, we focus on two central challenges: (1) credit assignment in reinforcement learning under delayed reward and long-horizon conditions, and (2) using sub-quadratic complexity models for efficient generative modelling. (Spoiler: We don’t need transformers!)
Speaker’s bio:
Vihang Patil is an applied scientist in the post-training team of Rufus, Amazon’s shopping chatbot. His current research focuses on aligning generative models using reinforcement learning. Previously, he completed his doctoral studies in Prof. Sepp Hochreiter’s group in JKU, Austria. His doctoral research centered on long-term credit assignment in reinforcement learning and abstraction-building algorithms to enhance learning efficiency.