Research map / Multi-agent systems and games / LLM agents
From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents
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.
Proposes PCE, a Planner-Composer-Evaluator framework that turns LLM reasoning-trace assumptions into a scored decision tree for uncertainty-aware multi-agent embodied planning with minimal communication.
Problem
Multi-agent embodied planning under partial observability and decentralization, where uncertainty about hidden objects and collaborator intentions is usually mitigated by costly, frequent inter-agent communication.
Approach
Extract implicit environment assumptions from LLM zero-shot CoT traces, structure them into a True/False decision tree whose leaves map to physical or communication actions, and score each root-to-leaf path by scenario likelihood times conditional gain minus execution cost, treating communication as just another action in the search.
Key results
- On C-WAH with GPT-4o mini, PCE reaches 42.76 total steps versus 60.40 for CoELA and 46.80 for REVECA, with 1.70 communication actions versus 9.88 for CoELA.
- On TDW-MAT with GPT-4o mini, PCE achieves 87.50% total transport success versus 81.25% for REVECA (second best).
- On TDW-MAT with Gemma3:4B, PCE reaches 70.83% total versus 67.50% for CaPo; PCE's token usage is higher than CoELA in several settings.
Surprising findings
- Token usage of PCE is not always lower: on TDW-MAT it is higher than CoELA for all three backbones (e.g., 197,807 vs 113,059 with GPT-4o mini), despite the abstract claiming comparable token usage.
- Scaling Gemma3 from 4B to 27B or raising GPT-OSS reasoning depth gave only modest gains for the Planner-only variant, while PCE gains persisted.
Limitations
- Evaluation limited to two simulated multi-room household benchmarks with two agents.
- Likelihood and gain scores are LLM-estimated approximations rather than true POMDP probabilities.
- Hyperparameters (D, alpha, beta, lambda) tuned and reported on C-WAH only; token usage higher than some baselines on TDW-MAT.
Experimental setup
Datasets: C-WAH TDW-MAT
Benchmarks: C-WAH TDW-MAT
Metrics: total steps success rate (proportion of target objects transported) token usages communication count user study Likert ratings
Baselines: CoELA REVECA CaPo CoTS Chain-of-Thought Tree of Thoughts Self-Consistency
Builds on: Chain-of-Thought Tree of Thoughts Self-Consistency CoELA CoTS
Closest papers by meaning
- Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making (NeurIPS 2024 · similarity 0.91)
- Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration (CVPR 2025 · similarity 0.90)
- Embodied Task Planning via Graph-Informed Action Generation with Large Language Models (ICML 2026 · similarity 0.90)
- PARTNR: A Benchmark for Planning and Reasoning in Embodied Multi-agent Tasks (ICLR 2025 · similarity 0.89)
- Towards Reliable LLM-based Robots Planning via Combined Uncertainty Estimation (NeurIPS 2025 · similarity 0.89)
- LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior (ICML 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.