MetaMind: General and Cognitive World Models in Multi-Agent Systems by Meta-Theory of Mind

Lingyi Wang; Rashed Shelim; Walid Saad; Naren Ramakrishna

cs.AIFeb 28, 2026

MetaMind: General and Cognitive World Models in Multi-Agent Systems by Meta-Theory of Mind

Authors: Lingyi Wang, Rashed Shelim, Walid Saad, Naren Ramakrishna

Organizations: Department of Electrical and Computer Engineering, Virginia Tech, USA · Department of Computer Science, Virginia Tech, USA

Abstract

A major challenge for world models in multi-agent systems is to understand interdependent agent dynamics, predict interactive multi-agent trajectories, and plan over long horizons with collective awareness, without centralized supervision or explicit communication. In this paper, MetaMind, a general and cognitive world model for multi-agent systems that leverages a novel meta-theory of mind (Meta-ToM) framework, is proposed. Through MetaMind, each agent learns not only to predict and plan over its own beliefs, but also to inversely reason goals and beliefs from its own behavior trajectories. This self-reflective, bidirectional inference loop enables each agent to learn a metacognitive ability in a self-supervised manner. Then, MetaMind is shown to generalize the metacognitive ability from first-person to third-person through analogical reasoning. Thus, in multi-agent systems, each agent with MetaMind can actively reason about goals and beliefs of other agents from limited, observable behavior trajectories in a zero-shot manner, and then adapt to emergent collective intention without an explicit communication mechanism. Extended simulation results on diverse multi-agent tasks demonstrate that MetaMind can achieve superior task performance and outperform baselines in few-shot multi-agent generalization.

Figures & tables

Figure 1: The proposed general and cognitive MetaMind in multi-agent systems.

Figure 2: Performance comparison on 8 SMAC tasks under limited environment steps. The win rates are averaged over 100 test episodes.

Map	Type	Multi-Agent World Model Methods			Multi-Agent RL Methods
		MetaMind ( Ours )	DCWMC ( Nomura et al., 2025 )	MARIE ( Zhang et al., 2024b )	MAPPO ( Yu et al., 2022 )	MBVD ( Xu et al., 2022 )
		Decentralized	Decentralized	CTDE	CTDE	CTDE
2m_vs_1z	Homog.	91.7 $\pm$ 6.2	75.1 $\pm$ 10.3	93.1 $\pm$ 3.3	84.6 $\pm$ 8.1	38.2 $\pm$ 19.1
2s_vs_1sc		93.2 $\pm$ 5.4	80.4 $\pm$ 9.2	96.1 $\pm$ 7.5	93.1 $\pm$ 6.1	16.1 $\pm$ 14.2
3m		95.5 $\pm$ 3.1	72.1 $\pm$ 10.7	96.2 $\pm$ 1.1	82.3 $\pm$ 11.9	72.1 $\pm$ 8.6
8m		89.1 $\pm$ 7.7	70.5 $\pm$ 12.8	84.1 $\pm$ 7.2	69.4 $\pm$ 17.8	71.2 $\pm$ 11.2

Table 1: Performance comparison of win rate (Mean ± Std.) across 13 SMAC maps for different methods, averaged over 100 random seeds after 200k environment steps of training. The corridor task is trained for 400K steps, the 3s_vs_5z and 2c_vs_64zg tasks are trained for 200K steps, and all remaining tasks are trained for 100K steps.

Figure 3: Performance comparison of win rate (Mean ± Std.) on 8 SMAC tasks versus different imagination horizon length.

Figure 4: Performance comparison of win rate in few-shot multi-agent generalization on 8 SMAC tasks under limited environment steps.

Figure 5: The scalability of a multi-map (MM) MetaMind to perform and cooperate over 13 SMAC maps, and the average horizon for goal identification of MT and single-map (SM) MetaMinds.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Few-shot learning for inverse belief inference by a self-supervised method in multi-agent systems.

Env	Scenario	#Agents	Goal Setting
SMAC (Easy)	1c3s5z	9	Type: Heterogeneous Roles: C $\times$ 1, S $\times$ 3, Z $\times$ 5 Goals: $G{=}3$
	2m_vs_1z	2	Type: Homogeneous Roles: M $\times$ 2 Goals: $G{=}1$
	2s_vs_1sc	2	Type: Homogeneous Roles: S $\times$ 2 Goals: $G{=}1$
	2s3z	5	Type: Heterogeneous Roles: S $\times$ 2, Z $\times$ 3 Goals: $G{=}2$
	3m	3	Type: Homogeneous Roles: M $\times$ 3 Goals: $G{=}1$
	3s_vs_3z	6	Type: Homogeneous Roles: S $\times$ 3 Goals: shared $G{=}1$

Appendix

Table 2: Environment Setting for MARIE (Goal Definitions)

Parameter	Symbol	Value
Shared Parameters
Replay memory size	—	1e6
Batch size	$B$	50
Sequence length	$L$	64
Seed episode	$S$	5
Training episodes	$N$	1e3

Appendix

Table 3: Hyperparameter Settings

Multi-Agent Systems

Multi-Agent Systems studies how autonomous agents coordinate, compete, communicate, and learn in shared or interdependent tasks.

See all 675 papers

Mental-Models for Multi-Agent Systems
Oct 8, 2026Hanan Gani, Lulu Shao, Manmohan ChandrakerMulti-Agent Reasoning Theory of Mind
ToM2Vec: Continuous Theory-of-Mind Tokens for Multi-Agent Coordination
Oct 21, 2026Nithin Sivakumaran, Ali Payani, Hyunji Lee +2Multi-Agent Coordination Theory of Mind
UNC Chapel Hill · Cisco · The University of Texas at Austin
MetaCogAgent: A Metacognitive Multi-Agent LLM Framework with Self-Aware Task Delegation
May 17, 2026Chenyu Wang, Yang ShuLLM Agent Orchestration LLM Agent Evaluation
School of Computer and Artificial Intelligence, Zhengzhou University, Zhengzhou, China · Zhejiang University, Hangzhou, China

Oct 8, 2026J/K move · Enter open · S save

Mental-Models for Multi-Agent Systems

Hanan Gani, Lulu Shao, Manmohan Chandraker

Multi-Agent Reasoning Theory of Mind

Open paper AskPDF arXiv

Large foundation models have accelerated progress toward general-purpose agents that interact with humans and other agents through language and multimodal signals. However, robust multi-agent decision-making requires reasoning about what other agents know, intend, and are likely to do under partial observability. Current agentic systems often operate through prompt design, memory, or end-to-end behavioral shaping, but typically do not learn an explicit partner-state representation that can be reused as a decision variable across tasks. We introduce mental-model-enabled agents, a framework that equips an agent with a latent mental model of its counterpart, allowing it to infer hidden beliefs, intentions, and likely reactions from the observed history and use these inferences to guide action selection. Our method learns an amortized recursive Theory-of-Mind representation, with first- and second-order mental-state structure, jointly with a belief-conditioned reward model that evaluates candidate actions relative to the inferred partner state. A policy is then learned under this belief-aware signal, yielding an agent that can act independently at inference time while retaining the benefits of explicit partner modeling. We evaluate the same framework on both language-only and multimodal benchmarks. Across these settings, explicit mental-state modeling consistently improves interaction quality and Theory-of-Mind performance over base agentic systems, showing that structured partner modeling is a useful inductive bias for general multi-agent systems. Our code is publicly available at https://github.com/hananshafi/Mental-Models

Abstract

Multi-Agent Systems

Explore similar work

Mental-Models for Multi-Agent Systems

ToM2Vec: Continuous Theory-of-Mind Tokens for Multi-Agent Coordination

MetaCogAgent: A Metacognitive Multi-Agent LLM Framework with Self-Aware Task Delegation