Dominic Rigby

OMNI-EPIC: Open-Endedness Via Models of Human Notions of Interestingness With Environments Programmed In Code

Date: 28th June 2025

arXiv link

Key Points

• Creates an open-ended learning framework based around leveraging the intuition and logic baked into LLMs and foundation models
• Process:
  1. Most ‘interesting’ task is selected (covered later)
  2. A new more interesting task is generated using an FM. The tasks are all described in text.
     • RAG is used to retrieve the N most similar tasks to give the LLM context of what the agent has already seen.
     • RAG is done using OpenAI’s text-embedding-3-small model
  3. Environment generator converts the description into a gym style environment.
  4. Model of Interestingness rates how interesting the new environment is.
     • Tries to mimic human concepts on interestingness it has learned from human text.
  5. Success detector (LLM or VLM) confirms whether or not the agent succeed at the task.