MindForge: Empowering Embodied Agents with Theory of Mind for Lifelong Cultural Learning

TL;DR

MindForge tackles the limitations of open-weight LLMs in embodied, open-ended environments by endowing agents with explicit Theory of Mind, a natural-language communication channel, and a triadic memory system. By integrating a structured Belief-Desire-Intention framework through the BigToM causal template, agents model both self and partner perspectives and exchange information to refine their beliefs during test-time dialogue. Empirical results in Minecraft show that open-weight LLMs, when collaborating with ToM-enabled peers or experts, close much of the gap to GPT-4 on basic tasks and achieve substantial gains on lifelong learning milestones, with performance scaling with the number of communication rounds. The work also demonstrates a Condorcet-like population boost, suggesting that richer social interaction can yield emergent improvements without increasing model size, aligning with Green AI goals for scalable, accessible embodied AI.

Abstract

Embodied agents powered by large language models (LLMs), such as Voyager, promise open-ended competence in worlds such as Minecraft. However, when powered by open-weight LLMs they still falter on elementary tasks after domain-specific fine-tuning. We propose MindForge, a generative-agent framework for cultural lifelong learning through explicit perspective taking. We introduce three key innovations: (1) a structured theory of mind representation linking percepts, beliefs, desires, and actions; (2) natural inter-agent communication; and (3) a multi-component memory system. Following the cultural learning framework, we test MindForge in both instructive and collaborative settings within Minecraft. In an instructive setting with GPT-4, MindForge agents powered by open-weight LLMs significantly outperform their Voyager counterparts in basic tasks yielding $3\times$ more tech-tree milestones and collecting $2.3\times$ more unique items than the Voyager baseline. Furthermore, in fully \textit{collaborative} settings, we find that the performance of two underachieving agents improves with more communication rounds, echoing the Condorcet Jury Theorem. MindForge agents demonstrate sophisticated behaviors, including expert-novice knowledge transfer, collaborative problem solving, and adaptation to out-of-distribution tasks through accumulated cultural experiences.

Figures (9)

• Figure 1: MindForge enables agents to (a) collaborate through structured theory-of-mind representations, which (b) leads to improved performance compared to Voyager when powered by open-weight LLMs, averaged across 3 runs.
• Figure 2: Cognitive architectures of MindForge and Voyager respectively. The Voyager architecture has been re-framed within the scope of our cognitive architecture. MindForge expands the original Voyager framework to enable communication through theory of mind together with additional memory subsystems to enhance the lifelong learning capabilities of the agent.
• Figure 2: Fraction of successful MindForge agents across $24$ individual trials; dirt and wood collection. Standard Voyager wang2023voyager setup: $4$ attempts per task, with a communication round interleaved where appropriate. See also \ref{['fig:comm-rounds']} and \ref{['fig:weak-weak']}.
• Figure 3: Each agent maintains a set of internal beliefs that it can exploit as needed. In order to enable ToM capabilities agents also maintain a belief over the beliefs, actions and desires of their collaborative counterparts.
• Figure 4: Failure mode correction. (a) Communication with an expert agent fixes the code error. (b) A weaker agent's incorrect task-related belief (initially believing it is mining 'grass blocks' and needs tools) is corrected by a stronger agent through communication (e.g., stating 'dirt blocks' do not require tools), leading to an updated task understanding.