Upload README.md with huggingface_hub

README.md

@@ -1,17 +1,55 @@
 ---
 license: apache-2.0
+library_name: pytorch
+base_model: facebook/dinov2-base
+datasets:
+- jiamingzz/lgwm-data
 tags:
-- gui-agents
 - world-model
+- gui-agent
+- computer-use
+- agent-safety
+- jepa
+- self-supervised
 ---
 
-# LGWM
+# LGWM: A Latent World Model that Verifies GUI Agent Transitions
 
 Weights for **Right Screen, Wrong Transition: World Models as Verifiers for GUI Agents**.
 
+[Code](https://github.com/jiamingzhang94/lgwm) · [Data](https://huggingface.co/datasets/jiamingzz/lgwm-data)
+
+LGWM is a decoder-free, action-conditioned world model for GUI agents. Given the
+current screenshot and an action, it predicts the representation of the next
+screen, and compares it with the screen that actually appears. One cosine
+distance tells whether the transition is the one the action should have produced.
+
+## Key results
+
+- **0.987 AUC** on RSWT-Bench with a training-free integrity score, on par with the
+  strongest closed-source VLMs.
+- **17 ms** per decision on one A100, over three orders of magnitude faster than
+  generative GUI world models, and about ten AUC points more accurate.
+- **0.954 AUC** at separating harmful from benign violations with a linear head on
+  the prediction residual, where prompted VLMs are near chance.
+- **173M** parameters at inference (ViT-B encoder, action encoder and predictor).
+
+## Files
+
 | Directory | Model |
 | --- | --- |
 | `lgwm-online/` | LGWM world model |
 | `lgwm-harm-head/` | Supervised harm head |
 
-Code and usage: https://github.com/jiamingzhang94/lgwm
+## Usage
+
+```bash
+git clone https://github.com/jiamingzhang94/lgwm && cd lgwm
+pip install -c requirements-tested.txt -e '.[text]'
+hf download jiamingzz/lgwm --local-dir weights
+python tools/score_transition.py --before before.png --after after.png \
+  --action action.json --harm-head weights/lgwm-harm-head
+```
+
+`action.json` holds the action, for example `{"action_type": "tap", "x": 0.5, "y": 0.5}`.
+The command prints `integrity` and `harm_score`.