README.md

---
license: apache-2.0
tags:
- vision-language
- embodied-ai
- causal-reasoning
- planning
- physical-grounding
---

<div align="center">

# *Token Predictors Are Not Planners*: Building Physically Grounded Causal Reasoners

Zheng Lu<sup>1,2,*</sup>, Mingqi Gao<sup>1,*</sup>, Qinlei Xie<sup>1,*</sup>, Wanqi Zhong<sup>1</sup>, Hanwen Cui<sup>1</sup>, Heng Cao<sup>1</sup>, Zirui Song<sup>3</sup>, Yifan Yang<sup>2</sup>, Chong Luo<sup>2</sup>, Bei Liu<sup>2,&dagger;</sup>, Yiming Li<sup>1,&dagger;</sup>

<sup>1</sup>Tsinghua University &nbsp;&nbsp; <sup>2</sup>Microsoft Research Asia &nbsp;&nbsp; <sup>3</sup>MBZUAI

<sup>*</sup>Equal contribution &nbsp;&nbsp; <sup>&dagger;</sup>Corresponding authors

<p>
  <a href="https://huggingface.co/Lululzz/Causal_Planner">
    <img src="https://img.shields.io/badge/Model-Causal_Planner-yellow?style=flat-square" alt="Model">
  </a>
  <a href="https://huggingface.co/datasets/anonymous-causal-plan/Causal_Plan/tree/main/CausalPlan-1M-QA">
    <img src="https://img.shields.io/badge/Dataset-CausalPlan--1M--QA-blue?style=flat-square" alt="Dataset">
  </a>
  <a href="https://huggingface.co/datasets/anonymous-causal-plan/Causal_Plan/tree/main/Causal-Plan-Bench">
    <img src="https://img.shields.io/badge/Benchmark-Causal--Plan--Bench-orange?style=flat-square" alt="Benchmark">
  </a>
</p>

</div>

## Model Overview

This repository hosts the Causal Planner model checkpoint released with *Token Predictors Are Not Planners: Building Physically Grounded Causal Reasoners*. The model is designed for physically grounded embodied planning, emphasizing causal preconditions, action-induced state transitions, long-horizon composition, and recovery under failures.

## Resources

- Model: https://huggingface.co/Lululzz/Causal_Planner
- Dataset: https://huggingface.co/datasets/anonymous-causal-plan/Causal_Plan/tree/main/CausalPlan-1M-QA
- Benchmark: https://huggingface.co/datasets/anonymous-causal-plan/Causal_Plan/tree/main/Causal-Plan-Bench

## License

This model card is released under the Apache-2.0 license metadata shown above. Dataset and benchmark usage should follow the corresponding Hugging Face repository terms.