Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Social World Model-Augmented Mechanism Design Policy Learning
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 SWM-AP, a model-based RL method that learns a social world model with posterior and prior trait trackers to infer heterogeneous agents' latent traits and augment PPO-based mechanism design policy learning, evaluated on facility location, team structure optimization, and tax adjustment.
Problem
Mechanism design in heterogeneous multi-agent social systems where agents have persistent unobservable latent traits and the principal must learn sample-efficiently from costly real-world interactions.
Approach
A hierarchical social world model with a posterior trait tracker (from full trajectories) for trait-conditioned dynamics prediction, and a prior trait tracker inside the mechanism design policy that infers traits online from partial history; the policy is trained with PPO on both real and imagined rollouts.
Key results
- In facility location, SWM-AP reaches PPO's final reward in the fewest training steps and attains the highest final reward among PPO, MBPO, Dreamer.
- In the 32-agent facility location extension, SWM-AP reaches 6.62 ± 0.06 final reward versus 6.43 for MBPO, with 274,667 steps to MBPO's final performance versus 433,600 for MBPO.
- In tax adjustment, SWM-AP achieves productivity comparable to PPO while significantly improving equality, outperforming MBPO on equality x productivity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SWM-AP instead of Dreamer | 32-agent Facility Location · 32-agent Facility Location | Final reward: 6.57 $\pm$ 0.06 -> 6.62 $\pm$ 0.06 | Helped |
| Use SWM-AP instead of MBPO | 32-agent Facility Location · 32-agent Facility Location | Steps to MBPO final performance: 433,600 -> 274,667 | Helped |
Surprising findings
- Confusion matrices show trait inference relies on modeling shortcuts from state history in later episode stages, motivating training on high-ambiguity initial states.
Limitations
- Scalability of SWM to extremely large-scale systems is a stated limitation.
- Interpretability of all inferred latent traits is not guaranteed.
- Trait inference shows off-diagonal confusion, with shortcut reliance on state history in later episode stages.
Experimental setup
Datasets: AdaSociety AI-Economist facility location matrix environment
Benchmarks: Facility location (8 agents, 8x8 grid) AdaSociety team structure optimization AI-Economist tax adjustment Facility location 32-agent extension
Metrics: final cumulative reward sample efficiency (steps to reach target) state prediction loss reward prediction loss equality productivity equality x productivity
Baselines: PPO MBPO Dreamer
Builds on: MBPO Dreamer TD-MPC PPO AI Economist
Closest papers by meaning
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.86)
- WPT: World-to-Policy Transfer via Online World Model Distillation (CVPR 2026 · similarity 0.86)
- General agents need world models (ICML 2025 · similarity 0.85)
- Revisiting Multi-Agent World Modeling from a Diffusion-Inspired Perspective (NeurIPS 2025 · similarity 0.85)
- World Models Should Prioritize the Unification of Physical and Social Dynamics (NeurIPS 2025 · similarity 0.85)
- Evolving Interpretable Constitutions for Multi-Agent Coordination (ICML 2026 · similarity 0.85)
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.