update readme for RL post-trained version

README.md

@@ -25,9 +25,9 @@ The model is built around a simple interface. Image patches and text tokens are
 - PBench dataset: `tiiuae/PBench`
 - OCR model: `tiiuae/Falcon-OCR`
 
-## Falcon Perception-HD (revision `19-08-2026`)
+## RL post-training for Falcon Perception (this revision)
 
-This revision contains **Falcon Perception-HD**, obtained by reinforcement learning (GRPO) post-training of Falcon Perception. Instead of the per-token likelihood used in SFT, the policy is aligned directly with detection metrics through a simple set-matching reward (a Hungarian-matched count that penalizes false negatives and false positives).
+This revision is an RL post-trained version of Falcon Perception, obtained by reinforcement learning (GRPO) post-training of Falcon Perception. Instead of the per-token likelihood used in SFT, the policy is aligned directly with detection metrics through a simple set-matching reward (a Hungarian-matched count that penalizes false negatives and false positives).
 
 RL post-training brings two main benefits:
 
@@ -61,6 +61,19 @@ model = AutoModelForCausalLM.from_pretrained(
 )
 ```
 
+### Citation
+
+If you use this version, please cite:
+
+```bibtex
+@article{chaybouti2026rlperception,
+  title={Falcon Perception-HD: High Density Perception via Reinforcement Learning},
+  author={Chaybouti, Sofian and Dahou, Yasser and Huynh, Ngoc Dung and Alami, Reda and Kuehne, Hilde},
+  journal={arXiv preprint arXiv:2608.18881},
+  year={2026}
+}
+```
+
 The inference API is identical to Falcon Perception (see below). The architecture and tokenizer are unchanged; only the weights differ.
 
 ## Quickstart