Update README.md

README.md

@@ -1,19 +1,16 @@
 ---
-license: apache-2.0
 tags:
-  - world-model
-  - autonomous-driving
-  - video-generation
-  - motion-planning
-  - diffusion
+- autonomous-driving
+- video-generation
+- diffusion
+- World-Action-Models
 ---
 
 <div align="center">
 
-# ReWorld: Representation Learning for World Action Models
+ReWorld: Representation Learning for World Action Models
 
-**The first representation learning framework for autonomous-driving World Action Models —<br>
-explicitly optimizing the latent world-to-action pathway.**
+**The first representation learning framework for autonomous-driving World Action Models**
 
 [![arXiv](https://img.shields.io/badge/arXiv-2606.27504-B31B1B?logo=arxiv)](https://arxiv.org/abs/2606.27504)
 [![Project Page](https://img.shields.io/badge/Project-Page-1f6feb?logo=githubpages)](https://xiaomi-research.github.io/ReWorld/)
@@ -33,14 +30,14 @@ explicitly optimizing the latent world-to-action pathway.**
 
 ---
 
-## :sparkles: Highlights
+## Highlights
 
 - **Better futures** — FVD **81.3 → 61.9** (−23.9%) on nuScenes video generation, with self-guided sampling enabled for free by intermediate supervision.
 - **Safer plans** — closed-loop PDMS **89.1 → 90.4** on NAVSIM *Navtest*, with best-in-class NC / DAC / TTC among world-model planners — no RL, no test-time scoring.
 - **Stronger representations** — frozen linear probe on UCF-101 action recognition **68.3% → 80.2%** (+11.9 pts over the DriveLaW baseline).
 - **Nearly free** — no external encoders, no teacher models, only **+0.3%** per-step Video DiT training cost, and roughly **2× faster** convergence from scratch.
 
-## :bulb: Why ReWorld?
+## Why ReWorld?
 
 World Action Models (WAMs) couple future environment prediction with action generation, yet standard training supervises only the output ends of the generation and planning modules. The intermediate representations that carry world knowledge are shaped only indirectly, as byproducts of fitting these outputs. We call this the **representation bottleneck of WAMs**: the world-to-action pathway is not explicitly optimized to be future-predictive, cross-modally grounded, or sensitive to closed-loop behavior quality.
 
@@ -58,7 +55,7 @@ All supervision is constructed entirely from the WAM's own generation targets, a
 <img src="assets/reworld_self_guidance.png" width="100%" alt="Self-guided sampling and faster convergence">
 </div>
 
-## :trophy: Results at a Glance
+## Results at a Glance
 
 <details open><summary><b>Video generation on nuScenes</b> (val)</summary>
 
@@ -127,7 +124,7 @@ Driving-domain pretraining alone barely moves the probe (+1.5 over LTX-Video); t
 <br><i>NAVSIM Navtest rollouts (red: ReWorld prediction, green: expert) — straight, left turn, right turn, intersection.</i>
 </div>
 
-## :package: Checkpoints
+## Checkpoints
 
 | Folder | File | Role |
 |:---|:---|:---|
@@ -148,7 +145,7 @@ hf download tz2026/ReWorld
 
 Please download the **whole repo** (not a single `.safetensors` file). Hugging Face only increments download stats when `config.json` is requested, which `snapshot_download` / `hf download` always fetch.
 
-## :rocket: Getting Started
+##  Getting Started
 
 ### Installation
 
@@ -215,11 +212,11 @@ sh scripts/evaluation/run_videodrive_agent_pdm_score_evaluation.sh
 
 > Optional base imitation before Stage 2: `scripts/training/run_videodrive_train.sh` with [`video_model_train_base.yaml`](DriveLaW-Act/navsim/agents/videodrive/configs/ltx_model/video_model_train_base.yaml), or start from the 🤗 DriveLaW checkpoint.
 
-## :pray: Acknowledgments
+##  Acknowledgments
 
-ReWorld builds on [DriveLaW](https://arxiv.org/abs/2512.23421), [NAVSIM](https://github.com/autonomousvision/navsim), [LTX-Video](https://github.com/Lightricks/LTX-Video), [Diffusers](https://github.com/huggingface/diffusers), and the hard-negative protocol of [BeyondDrive](https://github.com/wjl2244/BeyondDrive). Stage 1 uses intermediate supervision inspired by [Internal Guidance](https://arxiv.org/abs/2512.24176). Thanks to all of them for open-sourcing.
+ReWorld builds on [DriveLaW](https://arxiv.org/abs/2512.23421), [NAVSIM](https://github.com/autonomousvision/navsim), [LTX-Video](https://github.com/Lightricks/LTX-Video), [Diffusers](https://github.com/huggingface/diffusers), and the hard-negative protocol of [BeyondDrive](https://github.com/wjl2244/BeyondDrive). Stage 1 uses intermediate supervision inspired by [Internal Guidance](https://github.com/CVL-UESTC/Internal-Guidance). Thanks to all of them for open-sourcing.
 
-## :pencil: Citation
+##  Citation
 
 ```bibtex
 @article{xia2026reworld,
@@ -232,4 +229,4 @@ ReWorld builds on [DriveLaW](https://arxiv.org/abs/2512.23421), [NAVSIM](https:/
 
 ## :mailbox: Contact
 
-Tianze Xia — xiatianze@hust.edu.cn · Xinggang Wang — xgwang@hust.edu.cn
+Tianze Xia — xiatianze@hust.edu.cn · Xinggang Wang — xgwang@hust.edu.cn