Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Curious Causality-Seeking Agents Learn Meta Causal World
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.
Introduces Meta-Causal Graphs, a set of causal subgraphs indexed by latent meta states, and a curiosity-driven Causality-Seeking Agent that discovers them via interventions and a VQ-VAE codebook.
Problem
World models for open-ended environments assume a single fixed causal graph, but apparent causal mechanisms shift across latent contexts and partial observation windows, so observational learning yields misleading causal structure.
Approach
Represent the world as a Meta-Causal Graph: multiple causal subgraphs, each activated by a latent meta state discovered via VQ-VAE codebook entries, learned by curiosity-driven interventions (edge-entropy reward) and an intervention-based mask-refinement loss.
Key results
- On Chemical Fork with n=2 noisy nodes, MCG reaches 63.18% prediction accuracy versus 57.82% for FCDL (best baseline).
- On Chemical Chain downstream tasks with n=2 noisy nodes, MCG earns 13.82 average reward versus 11.16 for FCDL.
- On the Magnetic robot arm task, MCG reports 3.19 reward versus 2.85 for NCD and 2.77 for FCDL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MCG instead of the strongest listed baseline | causal world modeling · Chemical Fork and Chain OOD states (n=2, 4, 6) | prediction accuracy: Fork: 63.18$\pm$13.94 vs 57.82$\pm$9.90; Chain: 49.69$\pm$5.14 vs 42.06$\pm$1.29 | Helped |
| use MCG instead of the strongest listed baseline | downstream decision making · Chemical Chain and Fork (n=2, 4, 6) and Magnetic | average episode reward: Chain n=2: 13.82$\pm$3.84 vs 11.16$\pm$3.5; Magnetic: 3.19 $\pm$ 0.14 vs 2.85 $\pm$ 0.47 | Helped |
| add mask loss (intervention verification) (ablation) | MCG · Chemical Fork and Chain OOD states (n=2, 4, 6) | prediction accuracy: Fork n=6: 46.04$\pm$8.56 -> 50.04$\pm$8.56; Chain n=2: 47.75$\pm$7.52 -> 51.99$\pm$6.58; Fork n=4: 51.70$\pm$8.58 -> 50. | Mixed |
| add curiosity-based reward for active intervention exploration (ablation) | MCG · Chemical Fork and Chain OOD states (n=2, 4, 6) | prediction accuracy: Fork n=2: 48.28$\pm$8.15 -> 63.18$\pm$13.94; Chain n=4: 50.55$\pm$6.83 -> 49.78$\pm$4.11 | Mixed |
Surprising findings
- Several ablation and baseline differences are within large standard deviations, e.g., MCG vs. FCDL prediction accuracy on Fork (n=4, n=6) differs by about 1-3 points with std near 8-10.
Limitations
- Some states are unreachable via any intervention sequence, and implications for causal inference are not analyzed.
- Interventions incur costs; no budget-aware exploration is provided.
- Real-world partial observability and reachability assumptions are only partly validated.
Experimental setup
Datasets: Chemical environment (Fork, Chain) Magnetic robot arm manipulation (Robosuite-based)
Benchmarks: Chemical (Fork, Chain) OOD prediction Chemical downstream episode reward Magnetic robot arm task
Metrics: prediction accuracy (%) average episode reward
Baselines: MLP Modular GNN NPS CDL GRADER Sandy-Mixture Transformer NCD FCDL
Builds on: VQ-VAE interventional Markov equivalence NCD FCDL
Closest papers by meaning
- Systems with Switching Causal Relations: A Meta-Causal Perspective (ICLR 2025 · similarity 0.91)
- Modeling Unseen Environments with Language-guided Composable Causal Components in Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Sim-to-Real Causal Transfer: A Metric Learning Approach to Causally-Aware Interaction Representations (CVPR 2025 · similarity 0.87)
- Tackling Non-Stationarity in Reinforcement Learning via Causal-Origin Representation (ICML 2024 · similarity 0.86)
- ADAM: An Embodied Causal Agent in Open-World Environments (ICLR 2025 · similarity 0.86)
- Towards Generalizable Reinforcement Learning via Causality-Guided Self-Adaptive Representations (ICLR 2025 · similarity 0.86)
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.