From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Uncertainty-aware planning for embodied agents in partially observable multi-agent settings

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

PCE (Planner-Composer-Evaluator) is a framework for LLM-based embodied agents in decentralized, partially observable multi-agent environments that extracts the implicit assumptions scattered in LLM reasoning traces and structures them into a scored decision tree, enabling uncertainty-aware action selection without heavy inter-agent communication. On C-WAH and TDW-MAT benchmarks across three diverse LLM backbones, PCE consistently outperforms communication-centric baselines in task completion and efficiency while using comparable or fewer tokens.

Problem

Existing LLM-based embodied multi-agent planners handle partial observability primarily through frequent natural-language communication, which is expensive in tokens and time and disrupts human collaborators. Simply scaling model capacity or reasoning depth (Chain-of-Thought, Tree-of-Thoughts) does not resolve the core issue: LLM reasoning traces generate isolated, implicit assumption-action links without globally reconciling competing hypotheses about hidden state, leading to suboptimal planning under uncertainty.

Method

PCE models the cooperative task as a DEC-POMDP and redesigns the planning module into three sequential components:

1. Planner: Given goal \(G\), memory, and an available action list, performs zero-shot Chain-of-Thought reasoning and proposes an initial action with a reasoning trace that implicitly embeds environmental assumptions.

2. Composer: Interprets the trace to extract key uncertainties and builds a decision tree top-down. Internal nodes are binary (True/False) assumption splits; leaves are actions (physical or communicative). Expansion is bounded at depth \(D\); new assumptions are generated from context when existing ones are exhausted. Communication is treated as an action on equal footing with physical moves, not as the search mechanism.

3. Evaluator: Scores each root-to-leaf path (scenario \(S\), action \(a\)) with: $\(U(S, a) = L(S) \cdot G(a) - \lambda\, C(a)\)$ where \(L(S)\) is LLM-estimated scenario likelihood, \(G(a)\) is goal-directed conditional gain, and \(C(a) = \alpha\, d(a)\,\mathbf{1}_{\text{move}} + \beta\, \ell(a)\,\mathbf{1}_{\text{comm}}\) captures movement distance and message length. The leaf maximizing \(U\) determines the executed action.

Default hyperparameters: \(D=3\), \(\alpha=\beta=\lambda=1\).

Key Contributions

• A framework that treats environmental assumptions as explicit, first-class decision variables extracted from LLM reasoning traces rather than leaving them implicit.
• A decision-tree structure (Composer) that prevents incompatible assumption combinations along any path and selects communication only when its expected utility exceeds physical action.
• A utility scoring function (Evaluator) that jointly weighs scenario likelihood, conditional goal gain, and execution cost without solving an intractable POMDP.
• Empirical evidence that PCE's gains are additive to and orthogonal to both model-capacity scaling and reasoning-depth scaling.
• A user study confirming that PCE's communication patterns are perceived as more efficient and trustworthy by human partners.

Results

C-WAH (Total Steps ↓, 10 episodes, \(H=250\)): - GPT-4o mini: PCE 42.76 vs. CoELA 60.40 / REVECA 46.80 / CaPo 60.82 / CoTS 64.00 - GPT-OSS:20B: PCE 49.60 vs. CoELA 72.72 / REVECA 53.86 / CaPo 68.34 / CoTS 65.26 - Gemma3:4B: PCE 59.20 vs. CoELA 77.20 / REVECA 62.56 / CaPo 75.88 / CoTS 72.32

TDW-MAT (% objects transported ↑, 24 episodes, \(H=3000\); partial from visible tables): - GPT-4o mini: PCE Total 87.50% (Food 89.17%, Stuff 85.83%) vs. CoELA 62.50% (Food 65.8%)

Token efficiency: PCE Comm actions are 2–4× fewer than baselines; total token Usages are comparable or lower despite internal tree generation overhead.

Ablations: Increasing model size or reasoning depth alone yields diminishing returns; PCE lifts the baseline at every capacity/depth scale tested across GPT-4o mini, GPT-OSS:20B, and Gemma3:4B.

Limitations

• The Composer and Evaluator rely on LLM commonsense for likelihood and gain estimation rather than ground-truth probabilities, making scores noisy and model-dependent.
• Tree depth \(D\) is fixed; deep or highly branching uncertainty spaces may require large \(D\), increasing LLM calls.
• Evaluated only on two household-task benchmarks; generalization to manipulation-heavy or open-world domains is unverified.
• Communication is modeled with a one-step delay and point-to-point broadcast; richer communication topologies are not addressed.
• The user study scale is not reported in the provided text, limiting conclusions about statistical significance of human-perception findings.

Relevance to Agentic AI / LLM Agents

PCE directly advances uncertainty-aware planning for LLM agents operating in partially observable, multi-agent settings — a core unsolved problem in agentic AI. The key insight that LLM reasoning traces already contain implicit world-model assumptions that can be extracted and structured (rather than discarded after action selection) is broadly transferable: any agentic system using Chain-of-Thought reasoning in environments with hidden state could adopt this assumption-aggregation pattern. By treating communication as a utility-scored action rather than a default coordination mechanism, PCE provides a principled template for building agents that communicate selectively, directly relevant to human-in-the-loop agentic deployments. The DEC-POMDP formalization also connects this work to the broader literature on formal decision-making frameworks for LLM agents.