Human-Timescale Adaptation in an Open-Ended Task Space
What this work does
Meta-trains a large memory-based RL policy across a procedurally generated task space with an adaptive curriculum, enabling within-context adaptation to unseen embodied tasks.
Connection to discovery
An enabling foundation for learned experiment-like exploration and rapid adaptation in unfamiliar worlds.
Evidence
The body compares adaptation across trials, task/memory/model scaling and held-out task dynamics, including comparisons with trained human participants.
Scope and limitations
The task grammar, simulator and reward are designed; fast in-context adaptation is distinct from persistent weight updates or validated scientific law discovery.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.