Skip to content

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

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MindForge is a multi-agent generative framework layered atop Voyager that equips Minecraft embodied agents with explicit Theory of Mind (ToM), natural-language peer communication, and a three-part memory system to enable cultural lifelong learning. Open-weight LLMs under MindForge achieve 3× more tech-tree milestones and 2.3× more unique items than Voyager baselines powered by the same models. The core insight is that structured perspective-taking at inference time substitutes for gradient-based fine-tuning.

Problem

Voyager-style lifelong learning agents learn in isolation and depend critically on GPT-4's code generation quality; open-weight LLMs (Llama 3.1-8B, Mistral-7B, Mixtral-8×7B) fail even elementary Minecraft tasks (4–27% success) under the same architecture. PEFT fine-tuning on GPT-4 execution traces provides no reliable improvement, leaving a large capability gap with no social or collaborative mechanism to bridge it.

Method

MindForge extends Voyager with four coordinated components:

  1. Structured BDI self-model: Uses the BigToM causal template to represent each agent's percepts, beliefs, desires, and actions as a causal graph. Beliefs are formed via four separate LLM calls: perception-related, task-related, interaction-related, and partner-related, each conditioned on current percepts, memory retrievals, and conversation history.

  2. Theory of Mind / perspective-taking: Each agent maintains a partner-specific BigToM graph that is updated incrementally as dialogue proceeds. The LLM receives this partner model alongside conversation history to generate responses that account for the partner's inferred mental state (knowledge gaps, goals, misconceptions).

  3. Natural-language communication module: A dedicated LLM instance synthesizes incoming chat (via the Mineflayer API) into the BDI template, generating interaction beliefs and partner beliefs without coupling to the execution module. Communication is triggered on task failure; agents interleave one communication round between two environment action steps.

  4. Three-tier memory: (i) Episodic—past failure actions and partner details retrieved by similarity; (ii) Semantic—task/environment generalizations abstracted from episodes; (iii) Procedural—reusable skill code (Voyager's skill library). Episodic and semantic memories feed belief formation; procedural memory supplies candidate actions.

Key Contributions

  • BigToM-grounded BDI representation enabling recursive self- and other-modeling in open-ended 3D environments
  • Multi-turn peer dialogue as test-time knowledge distillation without gradient updates
  • Three-subsystem memory architecture separating episodic, semantic, and procedural stores
  • Empirical characterization of open-weight Voyager failure modes: false beliefs and faulty code generation that the critic cannot self-repair
  • Demonstration of a Condorcet-style "many-minds" population-level boost in weak–weak agent pairs through structured dialogue

Results

  • Tech-tree (open-weight LLMs): MindForge agents achieve 3× more milestones and 2.3× more unique items than Voyager counterparts, averaged over Llama 3.1-8B, Llama 3.1-70B, and Mixtral-8×7B
  • Isolated tasks — Mistral-7B single agent: dirt 7% → 37.5%; wood 4% → 33.3% (MindForge vs. Voyager)
  • PEFT baseline: Mistral-7B fine-tuned on GPT-4 traces scores 7% dirt / 12.5% wood — no meaningful gain over untuned Voyager
  • Wizard-of-oz feasibility: Human expert collaboration yields +57.85% average task-completion improvement
  • Weak–weak collaborative setting: 1 GPT-4 seeding round + 7 peer dialogue turns raises population success from 62% → 79%
  • GPT-4 Voyager retains perfect (100%) accuracy on dirt and wood; MindForge narrows but does not close the gap with open-weight models

Limitations

  • Evaluated exclusively in Minecraft; generalization to other embodied or non-game environments is untested
  • Communication rounds add latency overhead (one round interleaved per two environment steps)
  • Instructive settings still depend on a GPT-4 expert as a seed; fully open-weight multi-agent performance is the harder regime
  • Partner belief graph is updated only via text chat, not grounded in shared visual or spatial perception
  • Ablation of episodic memory shows a success-rate drop, but full ablation details are in appendices not fully available in this excerpt

Relevance to Harnesses / Meta-Harnesses

MindForge is architecturally a meta-harness: it orchestrates several specialized sub-agent LLM instances (belief formation, communication synthesis, curriculum, code execution) and routes control flow between them based on runtime state (task failure triggers communication; memory similarity scores determine retrieval). The framework demonstrates how a meta-harness can implement test-time compute scaling — allocating more inference rounds to hard tasks through structured communication loops rather than retraining — which is a key design pattern for adaptive harnesses. The three-tier memory architecture (episodic/semantic/procedural with separate read/write paths) offers a reusable template for harnesses that must persist and selectively retrieve heterogeneous knowledge across long horizons. The modular belief-formation pipeline (four independent LLM calls synthesized into a shared template) also exemplifies how a harness can decompose a cognitively complex step into parallelizable, auditable sub-calls.