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

This repository contains the official implementation of:

From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents
SeungWon Seo*, SooBin Lim*, Seongrae Noh*, Haneul Kim, HyeongYeop Kang
ICLR 2026 Poster

Paper Link: OpenReview

If you encounter any issues or bugs in the code, please feel free to open an issue in this repository. If my response is delayed, it is likely that I may have overlooked the notification. In urgent cases, you are welcome to contact me directly at ssw03270@korea.ac.kr, and I will review and respond as soon as possible.

---

Repository Structure

PCE/
|-- cwah/          # C-WAH benchmark code
|-- tdw_mat/       # TDW-MAT benchmark code

---

Environment Setup

This codebase is organized by benchmark. Please set up each environment separately.

1) C-WAH Setup

1. Clone the modified VirtualHome API (wah branch) one level above this repository:

git clone --branch wah https://github.com/xavierpuigf/virtualhome.git

2. Download the simulator and place it in ../executable/:

gdown https://drive.google.com/uc?id=1L79SxE07Jt-8-_uCvNnkwz5Kf6AjtaGp
unzip executable.zip
chmod +x executable/linux_exec.v2.3.0.x86_64

3. Create environment and install dependencies:

cd cwah
conda create --name cwah python=3.9
conda activate cwah
pip install -r requirements.txt

2) TDW-MAT Setup

cd tdw_mat
conda create -n tdw_mat python=3.9
conda activate tdw_mat
pip install -r requirements.txt

---

API Key Setup

For OpenAI-based runs, set both variables for compatibility:

export OPENAI_API_KEY="your_api_key"
export OPENAI_KEY="your_api_key"

---

Running the Code

C-WAH (Two LLM Agents with communication)

cd cwah
bash scripts/gpt-4o-mini_w-comm.sh

TDW-MAT (Two LLM Agents with communication)

cd tdw_mat
bash scripts/test_LMs-gpt-4o-mini.sh

---

Citation

If you find this work useful in your research, please cite:

@inproceedings{
seo2026from,
title={From Assumptions to Actions: Turning {LLM} Reasoning into Uncertainty-Aware Planning for Embodied Agents},
author={SeungWon Seo and SooBin Lim and Seongrae Noh and Haneul Kim and HyeongYeop Kang},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=GODFBZhFcX}
}

---

Acknowledgements

This code is derived from the code of the following paper:

• Building Cooperative Embodied Agents Modularly with Large Language Models by Zhang et al.
• REVECA: Adaptive Planning and Trajectory-based Validation in Cooperative Language Agents using Information Relevance and Relative Proximity by Seo et al.