Add model card metadata and improve README

README.md

@@ -1,3 +1,35 @@
 ---
 license: cc-by-4.0
----
+library_name: transformers
+pipeline_tag: text-generation
+---
+
+# AeSlides: Incentivizing Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards
+
+AeSlides is a reinforcement learning framework designed to improve the aesthetic layout quality of slides generated by Large Language Models (LLMs). This model is based on GLM-4.7-Flash and is optimized to address the "modality gap" in slide generation, where the generation process is text-centric but the quality is judged by visual aesthetics.
+
+- **Paper:** [AeSlides: Incentivizing Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards](https://arxiv.org/abs/2604.22840)
+- **Repository:** [https://github.com/ympan0508/aeslides](https://github.com/ympan0508/aeslides)
+
+## Methodology
+
+AeSlides introduces a suite of programmatically verifiable metrics to quantify slide layout quality, including aspect ratio compliance, whitespace reduction, element collisions, and visual imbalance. These metrics are used as rewards in a GRPO-based reinforcement learning framework to align the model with human aesthetic preferences.
+
+With only 5K training prompts, AeSlides significantly improves layout coherence:
+- Aspect ratio compliance: 36% → 85%
+- Whitespace reduction: ↓ 44%
+- Element collisions: ↓ 43%
+- Visual imbalance: ↓ 28%
+
+## Citation
+
+If you find this work useful, please cite the paper:
+
+```bibtex
+@article{pan2025aeslides,
+  title={AeSlides: Incentivizing Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards},
+  author={Pan, Yumeng and others},
+  journal={arXiv preprint arXiv:2604.22840},
+  year={2025}
+}
+```