Adding link to the ArXiv paper

app.py

@@ -52,7 +52,7 @@ def detect(attacked_text, window_size, method, prompt):
 with gr.Blocks() as demo:
     gr.Markdown("""# LLM generation watermarking
     This spaces let you to try different watermarking scheme for LLM generation.\n
-    It leverages the upgrades introduced in the paper, reducing the gap between empirical and theoretical false positive detection rate and give the ability to embed a message (of n bits). Here we use this capacity to embed the identity of the user generating the text, but it could also be used to identify different version of a model or just convey a secret message.\n
+    It leverages the upgrades introduced in the paper **[Three Bricks to Consolidate Watermarks for Large Language Models](https://arxiv.org/abs/2308.00113)**, reducing the gap between empirical and theoretical false positive detection rate and give the ability to embed a message (of n bits). Here we use this capacity to embed the identity of the user generating the text, but it could also be used to identify different version of a model or just convey a secret message.\n
     Simply select an user name, set the maximum text length, the watermarking window size and the prompt. Aaronson and Kirchenbauer watermarking scheme are proposed, along traditional sampling and greedy search without watermarking.\n
     Once the text is generated, you can eventually apply some attacks to it (e.g, remove words), select the associated detection method and run the detection. Please note that the detection is non-blind, and require the original prompt to be known and so left untouched.\n
     For Aaronson, the original detection function, along the Neyman-Pearson and Simplified Score version are available.""")