|
This week we have Sumaita Rahman presenting the paper “Training a Generally Curious Agent” that was previously presented at 2025 ICML.
The paper provides a fine-tuning method that improves AI agents’ ability to explore new environments by first exposing a language model to diverse simulated tasks and then using preference-based scoring to reinforce which strategies led to success. To maximize learning efficiency, it prioritizes tasks by their “learning potential,” forming a curriculum similar to how students progress through increasing difficulty. This encourages the model to develop general problem-solving strategies
We’ll be in the Executive Conference room again this week. You should be able to use the front door directly into the conference room.
|
