Technical Summary: From Assumptions to Actions (PCE)

Problem Definition

The paper addresses the challenge of uncertainty-aware planning for embodied agents operating in decentralized, partially observable multi-agent environments. In such settings (e.g., household collaboration), agents possess limited perceptual fields, leading to pervasive uncertainty regarding hidden objects and the intentions/actions of collaborators.

Existing approaches powered by Large Language Models (LLMs) often mitigate this uncertainty through frequent inter-agent communication. While effective for coordination, this "communication-centric" paradigm incurs substantial token and time costs. Furthermore, continuous questioning and reporting can disrupt workflows, particularly when human partners are involved. The authors argue that simply scaling LLM capacity or deepening reasoning chains does not inherently resolve uncertainty; without explicit mechanisms to assess underlying assumptions, larger models still struggle to weigh competing hypotheses, leading to misaligned priorities and degraded planning quality.

Methodology: The PCE Framework

The authors propose PCE (Planner-Composer-Evaluator), a modular framework that transforms fragmented, latent assumptions within LLM reasoning traces into a structured decision tree for rational action selection without heavy communication. The framework operates on a Decentralized Partially Observable Markov Decision Process (DEC-POMDP) model.

The Planning Module is redesigned into a three-stage pipeline:

1. Planner (Reasoning for Action Selection):

   • Receives context (goals, memory, observations) and generates an initial action choice accompanied by a Chain-of-Thought (CoT) reasoning trace.
   • Empirically, these traces contain implicit, isolated assumptions about the environment (e.g., "items are likely in the cabinet") but fail to explicitly aggregate them for global decision-making.

2. Composer (From Trace to Scenario Tree):

   • Extracts key uncertainties (assumptions) from the Planner's trace and structures them into a decision tree.
   • Internal nodes represent assumptions with True/False splits.
   • Leaf nodes represent specific actions (physical or communication) appropriate for the accumulated assumptions along that path.
   • The tree is expanded top-down using LLM commonsense reasoning to prioritize assumptions that most reduce uncertainty. It generates new atomic assumptions if necessary, ensuring consistency with current premises.

3. Evaluator (Scoring and Selection):

   • Scores each root-to-leaf path (scenario) to guide action selection.
   • The scoring function $U(S, a)$ integrates three normalized metrics:
     • Likelihood ($L$): The estimated probability that the scenario's premise is true.
     • Conditional Gain ($G$): The expected progress toward the goal if the scenario holds.
     • Execution Cost ($C$): The resources required (movement distance or message length).
   • The final score is calculated as $U(S, a) = L(S) \cdot G(a) - \lambda C(a)$.
   • The agent selects the action corresponding to the leaf with the highest score, balancing physical exploration against communication only when communication yields higher expected utility.

Key Contributions

• Structural Shift in Planning: Unlike prior works that optimize coordination through iterative communication or joint plan negotiation, PCE treats environmental assumptions as first-class decision variables. It reasons over uncertainty embedded in the agent's belief state before action execution.
• Uncertainty-Aware Decision Tree: The framework converts latent LLM assumptions into an explicit decision tree, enabling the systematic reconciliation of multiple hypotheses, detection of logical conflicts, and consistent weighing of alternative actions.
• Communication Efficiency: By explicitly evaluating the utility of communication against physical actions, PCE reduces unnecessary dialogue while maintaining or improving planning performance.
• Model Agnosticism: The framework operates on generic reasoning traces rather than model-specific internals, making it applicable to diverse LLM backbones.

Experimental Results

The authors evaluated PCE on two challenging multi-agent benchmarks: C-WAH (10 episodes, 3–5 sub-goals, horizon 250) and TDW-MAT (24 episodes, 10 target objects, horizon 3000). Experiments were conducted across three diverse LLM backbones: GPT-4o mini, GPT-OSS:20B, and Gemma3:4B.

• Performance: PCE consistently outperformed communication-centric baselines (CoELA, REVECA, CaPo, CoTS) in success rate and task efficiency (fewer total steps or higher object transport proportion) across all backbones.
• Token Usage: PCE achieved comparable or lower total token usage ("Usages") compared to baselines, despite its multi-module architecture, by significantly reducing the number of communication actions and episode lengths.
• Ablation Studies:
  • Scaling: Increasing model capacity or reasoning depth alone yielded only modest improvements. PCE provided robust, additive performance gains across all scales, confirming that structured uncertainty handling complements scaling.
  • Components: Removing any module (Planner, Composer, or Evaluator) degraded performance, confirming the necessity of the full pipeline.
• User Study: In a human-agent collaboration study, participants rated PCE significantly higher than "Communication Always" and "No Communication" variants in terms of appropriateness, usefulness, efficiency, and trust. PCE's selective communication was perceived as more aligned and reliable.

Significance and Claims

The paper claims to establish a principled route for turning latent LLM assumptions into reliable strategies for uncertainty-aware planning. Its primary significance lies in shifting the focus from communication-heavy mitigation to the principled exploitation of assumptions already present in LLM reasoning.

The authors assert that:

1. Communication is not a prerequisite for coordination: Agents can adaptively balance physical and communicative actions by explicitly reasoning over uncertainty.
2. Scaling is insufficient: Merely increasing model size or reasoning depth does not solve the fundamental reliance on outdated internal memory in partially observable environments; explicit uncertainty structuring is required.
3. Generality: The framework is domain-agnostic and can be applied to various LLM backbones and multi-agent domains without requiring model-specific fine-tuning.

The work suggests a new direction for embodied systems where rational action selection is driven by structured hypothesis evaluation rather than reactive dialogue.