Technical Summary: EnactToM

Problem Statement

Current benchmarks for Theory of Mind (ToM) in Large Language Models (LLMs) primarily evaluate literal ToM: the ability to explicitly report another agent's beliefs, desires, or intentions when prompted. However, these evaluations fail to measure functional ToM, defined as the capacity to act optimally based on inferred mental states in grounded, multi-agent environments. Existing benchmarks lack three critical properties:

1. Functional: They measure belief reporting rather than action execution based on those beliefs.
2. Grounded: They operate in text-only or fully observable settings, lacking the spatial constraints and information asymmetry inherent to embodied collaboration.
3. Saturation-resistant: Static benchmarks become trivial as models improve, and they lack mechanisms to evolve task difficulty based on specific model failure modes.

Consequently, frontier models often demonstrate high accuracy on belief probes but fail to coordinate effectively in shared physical environments, leading to redundant actions, ignored constraints, and safety risks.

Methodology: The EnactToM Framework

The authors introduce EnactToM, an evolving benchmark consisting of 300 embodied multi-agent tasks set in a 3D household environment (HSSD). The framework is designed to isolate functional ToM through partial observability, private information, and constrained communication.

1. Task Representation and Epistemic Depth

Tasks are formalized as tuples $T = (S, A, \phi, \delta, M, \Sigma, C)$, where:

• $S$: A 3D scene.
• $A$: A set of agents with spawn positions.
• $\phi$: A PDDL goal formula combining physical predicates (e.g., is_on_top) with epistemic operators ($K$).
• $\delta$: A natural language task description shared by all agents.
• $M$: Mechanics modeling real-world constraints (e.g., room restrictions, limited bandwidth, restricted communication graphs).
• $\Sigma$: Private secrets assigned to each agent (e.g., specific object IDs, room access limits) that create information asymmetry.
• $C$: Task category (Cooperative or Mixed-motive).

The benchmark utilizes epistemic depth ($d$) to quantify ToM complexity:

• Depth 1 (Self-aware): An agent recognizes its own information gap.
• Depth 2 (Other-aware): An agent acts based on what a partner knows.
• Depth 3 (Recursive): An agent reasons about what a partner knows that another agent knows.
  The benchmark caps generation at depth 3, as deeper recursion is brittle even for humans.

2. Agentic Task Generation Pipeline

To address benchmark saturation, EnactToM employs an autonomous coding agent to generate tasks. The pipeline includes:

• Authoring: The agent writes the formal PDDL goal first, then derives the narrative and secrets, ensuring the goal is the source of truth.
• Verification:
  • PDDL Parsing: Checks syntactic validity and K-depth alignment.
  • LLM Judge Council: Two models (Kimi-K2.5 and GPT-5.2) independently score tasks on criteria such as agent necessity, secret quality (no strategy leakage), and mechanic utilization.
  • Structural Calibrator: Runs a baseline simulation where all secrets are revealed to all agents. If the task fails even with full information, it is rejected as physically unsolvable. This ensures that success in the standard setting depends on epistemic reasoning, not just embodied navigation.
• Evolution: The generation agent uses "seed tasks" biased toward the failure modes of current frontier models (e.g., GPT-5.4, O3). As models improve, the seed pool shifts to harder failure patterns, evolving the benchmark difficulty without modifying the generator infrastructure.

3. Evaluation Protocol

The benchmark evaluates seven frontier models on two splits: Standard ($\rho=0.8$ seed failure ratio) and Hard ($\rho=0.9$).

• Functional Metric: Success is defined by completing the physical predicates in $\phi$. The primary metric is Pass^3, requiring success in all three independent runs to rule out luck.
• Literal Metric: At the end of each episode, agents are queried about specific beliefs (e.g., "What does agent_0 think the fridge state is?"). This measures literal ToM accuracy independent of action.

Key Results

The evaluation reveals a stark act–report gap between literal and functional ToM:

• Functional Failure: On the Hard split, all seven evaluated frontier models achieved 0.0% Pass^3 on functional task completion. No model could reproducibly coordinate across three runs.
• Literal Success: The same models averaged 45.0% on literal belief probes on the same tasks. For instance, Gemini-Pro achieved 63.3% literal accuracy but only 6.7% functional accuracy on the hard split.
• Failure Analysis: A manual audit of 40 failures identified that 93% were due to epistemic coordination breakdowns, categorized into five patterns:
  1. Withholding critical information: Agents possess decisive facts but communicate them too late.
  2. Epistemic chain breakdown: Agents complete physical actions but fail to establish that partners know the outcome.
  3. Private objective sabotage/disclosure: In mixed-motive tasks, agents either reveal private goals prematurely or sabotage shared goals.
  4. Misallocating scarce messages: Agents waste limited bandwidth on unreachable partners or low-priority content.
  5. Ignoring partner constraints: Agents delegate tasks to partners who lack the necessary room access or resources.

Significance and Claims

The paper claims that EnactToM provides the first rigorous, evolving measurement of functional ToM in embodied settings. Its significance lies in:

1. Exposing a Critical Gap: Demonstrating that high performance on text-based belief questions does not translate to the ability to use those beliefs for coordination in constrained environments.
2. Defining a New Benchmark Standard: Establishing a framework that is embodied (3D environment), functional (action-based success), and evolving (resistant to saturation via failure-driven generation).
3. Providing Concrete Targets: By decomposing failures into specific epistemic coordination breakdowns, the benchmark offers clear, actionable targets for future research in multi-agent alignment and reasoning.

The authors conclude that while frontier models can often state what partners know, they currently lack the policy to use that knowledge to act reliably under communication and information constraints. The benchmark serves as a measurement instrument to track progress in closing this gap.