Implicit Causal World Models Learn Multi-Agent Dynamics from Offline Demonstrations隐式因果世界模型：从离线多智能体演示中学习环境动态
S 1.7T11 sources1 个来源R7-research
Jasorsi Ghosh's paper "Learning Implicit Causal World Models from Multi-Agent Demonstrations" (arXiv:2607.26336v1, submitted 28 Jul 2026) addresses how model-based RL world models often conflate statistical correlations with true causal mechanisms, a problem worsened in multi-agent systems where physical transitions are entangled with strategic agent intents, causing failures under distribution shift.
The proposed Implicit Causal World Models recover environmental dynamics directly from offline demonstrations without requiring pre-defined causal graphs, incorporating policy variance so that the world model becomes discoverable via the sequential backdoor condition, with evaluations conducted across multi-agent coordination tasks.
Jasorsi Ghosh 论文《Learning Implicit Causal World Models from Multi-Agent Demonstrations》（arXiv:2607.26336v1，2026年7月28日提交）指出,基于模型的强化学习中,世界模型训练常将统计相关性与真实因果机制混淆,而在多智能体系统中物理状态转移与策略性意图交织,使该问题在分布漂移下更易失效。