Research map / Multi-agent systems and games / LLM agents

Adaptive Theory of Mind for LLM-based Multi-Agent Coordination

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

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.

Shows that mismatched theory-of-mind (ToM) orders between LLM agents hurt coordination and proposes A-ToM, an agent that maintains ToM-0/1/2 hypothetical partners and uses Follow-the-Leader or Hedge expert-advice learning to align with the partner's ToM order.

Problem

LLM-based agents with explicit higher-order Theory of Mind often fail to coordinate with partners because their ToM reasoning depths are misaligned, leading to insufficient or excessive reasoning about the partner.

Approach

Frame ToM-order alignment as an online expert-advice problem: maintain hypothetical ToM-0/1/2 agents as experts, predict the partner's action from the expert chosen by FTL or Hedge based on past prediction accuracy, and then best-respond to that prediction.

Key results

• Aligned ToM-order pairs reach 75 points in the Memory-1 repeated matrix game while misaligned pairs such as ToM-1 vs ToM-1 score 0.
• The A-ToM agent with FTL or Hedge matches aligned partners against fixed ToM-0, ToM-1 and ToM-2 partners, e.g. 75.00 points vs ToM-0 and ToM-2 in the Memory-1 game.
• In cross-play with non-LLM agents in Overcooked, A-ToM (FTL) achieves the lowest mean completion time (48.05) versus Greedy (54.06) and PBT (49.09).

Surprising findings

• Self-play of two A-ToM agents with FTL fails completely in the Memory-1 matrix game (0 points), while Hedge self-play succeeds.
• From the A-ToM agent's perspective, Greedy and PBT agents are mostly interpreted as ToM-0 agents, suggesting planning and RL agents show no genuine ToM capability.

Limitations

• Only ToM orders up to 2 are considered, motivated by human cognitive limits.
• Experiments use a single LLM (LLaMA-3.3-70B-Instruct) at temperature 0.1 on small custom tasks.
• The 3-action matrix game generalization analysis is preliminary and uses a higher temperature.

Experimental setup

Benchmarks: Repeated matrix game (Memory-1, Memory-N) Grid World Navigation Game 1 Grid World Navigation Game 2 Overcooked

Metrics: points completion time

Baselines: ToM-0 ToM-1 ToM-2 Greedy PBT

Builds on: expert advice Follow-the-Leader Hedge theory of mind prompting

Closest papers by meaning

• Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.89)
• Hypothetical Minds: Scaffolding Theory of Mind for Multi-Agent Tasks with Large Language Models (ICLR 2025 · similarity 0.89)
• Mental-Models for Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
• AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling (NeurIPS 2025 · similarity 0.88)
• Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.88)
• MuMA-ToM: Multi-modal Multi-Agent Theory of Mind (AAAI 2025 · similarity 0.87)