OMNI: Open-endedness via Models of human Notions of Interestingness
What this work does
Combines estimated learning progress with a foundation-model judgment of human interestingness to sample training tasks.
Connection to discovery
A way to distinguish useful, learnable research questions from unlimited novelty or raw prediction error.
Evidence
Inspected task-success normalization, bidirectional progress for forgetting and comparisons in Crafter/finite task spaces.
Scope and limitations
Interestingness inherits human/model priors and can be gamed. A finite task space or finite experiment does not demonstrate indefinitely sustained open-endedness.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.