Research map / Large language models / AI and society
OMNI: Open-endedness via Models of human Notions of Interestingness
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OMNI, which uses foundation models as a model of human interestingness to filter learning-progress-based auto-curricula so agents focus on tasks that are both learnable and interesting, validated in Crafter, BabyAI, and AI2-THOR.
Problem
Learning-progress-based auto-curricula in open-ended environments keep selecting learnable but uninteresting tasks (e.g., minor variations of learned tasks), and interestingness is hard to quantify with hand-crafted formulas.
Approach
Use prompted foundation models (GPT-3, GPT-4) as a Model of Interestingness that partitions tasks into interesting vs. boring from few-shot examples; boring tasks are down-weighted (x0.001) in an LP curriculum. In infinite spaces, GPT-4 proposes new learnable, interesting tasks with code-based completion checks.
Key results
- In Crafter, OMNI (LP + MoI) learned 82 tasks (CI 80-87) and 14 interesting tasks vs. 55 tasks and 9 interesting for LP only, with average success 0.56 vs. 0.42 on all tasks.
- In AI2-THOR infinite task space, OMNI learned 13 (CI 11-17) tasks versus 2 for the LP baseline and 0 for Uniform sampling.
- FM-based MoI performance was nearly indistinguishable from an oracle MoI (OMoI) in Crafter and BabyAI, with most differences not statistically significant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FM-based interestingness filtering to LP task sampling | PPO task-conditioned agent · Crafter | average task success rate (interesting and boring tasks): 0.42 -> 0.56 | Helped |
| Add FM-based interestingness filtering to LP task sampling | PPO task-conditioned agent · Crafter | tasks learned (all tasks): 55 -> 82 | Helped |
| Add FM-based interestingness filtering to LP task sampling | PPO task-conditioned agent · Crafter | average task success rate (interesting tasks): 0.52 -> 0.78 | Helped |
| Add FM-based interestingness filtering to LP task sampling | PPO task-conditioned agent · Crafter | tasks learned (interesting tasks): 9 -> 14 | Helped |
| Add FM-based interestingness filtering to LP task sampling | PPO task-conditioned agent · BabyAI MiniBossLevel | average task success rate (all tasks): 5.9e-3 -> 7.5e-3 | Helped |
| Add FM-based interestingness filtering to LP task sampling | PPO task-conditioned agent · BabyAI MiniBossLevel | tasks learned: 4 -> 8 | Helped |
| Generate learnable and interesting tasks with an FM instead of adding uniformly sampled tasks | AI2-THOR LP curriculum · AI2-THOR kitchen | tasks learned: 2 -> 13 | Helped |
| Use LP sampling instead of uniform task sampling | PPO task-conditioned agent · Crafter | tasks learned (all tasks): 4 -> 55 | Helped |
| Use LP sampling instead of uniform task sampling | PPO task-conditioned agent · BabyAI MiniBossLevel | tasks learned: 1 -> 4 | Helped |
| Use LP sampling instead of uniform task sampling | PPO task-conditioned agent · AI2-THOR kitchen | tasks learned: 0 (CI: 0 -- 2) -> 2 (CI: 0 -- 3) | Helped |
| Add normalization against random-action success rates to the LP curriculum (ablation) | Crafter LP curriculum · Crafter | average task success rate and tasks learned | Helped |
| Use an FM-based Model of Interestingness instead of the oracle Model of Interestingness | LP curriculum · Crafter and BabyAI | performance across reported metrics | No clear effect |
Surprising findings
- Uniform sampling fails to improve over time in Crafter and BabyAI because most sampled tasks are too hard or too easy.
Limitations
- Boring tasks in Crafter are defined by hand-designed categories (numerical repeats, compounds, synonyms), and the interesting set is partially determined by the FM.
- Relies on proprietary LLMs (GPT-3, GPT-4) with API costs and reproducibility concerns.
- Authors note possible Goodhart-style pathologies if the MoI is optimized against.
Experimental setup
Datasets: Crafter tasks (15 interesting, 90 boring, 1023 extremely challenging) BabyAI MiniBossLevel tasks (1364 tasks) AI2-THOR kitchen floorplan tasks generated by GPT-4
Benchmarks: Crafter BabyAI AI2-THOR
Metrics: average task success rate number of tasks with success rate above threshold (0.2 Craft
Baselines: Uniform sampling Learning Progress (LP) only Oracle Model of Interestingness (OMoI) LP-no-norm
Builds on: learning progress curriculum (Teacher-student curriculum / ACL-style LP) PPO POET-style open-ended learning
Closest papers by meaning
- OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code (ICLR 2025 · similarity 0.91)
- Position: Open-Endedness is Essential for Artificial Superhuman Intelligence (ICML 2024 · similarity 0.86)
- Large Language Models Think Too Fast To Explore Effectively (NeurIPS 2025 · similarity 0.85)
- Roll the dice & look before you leap: Going beyond the creative limits of next-token prediction (ICML 2025 · similarity 0.85)
- Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) (NeurIPS 2025 · similarity 0.84)
- Learning Interestingness in Automated Mathematical Theory Formation (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.