Add model card for COMPOSE

README.md

@@ -0,0 +1,46 @@
+---
+license: mit
+pipeline_tag: text-generation
+tags:
+- mathematics
+---
+
+# COMPOSE: Composing Future Theorems from Citations and Formal Structure
+
+COMPOSE is a dual-graph framework that generates predicted future mathematical claims for a given anchor paper. It leverages two complementary sources of context: the scientific citation graph and an aligned formal theorem dependency graph from Mathlib.
+
+- **Paper:** [COMPOSE: Composing Future Theorems from Citations and Formal Structure](https://huggingface.co/papers/2605.30333)
+- **Project Page:** [https://david-busbib.github.io/COMPOSE-page/](https://david-busbib.github.io/COMPOSE-page/)
+- **Repository:** [https://github.com/david-busbib/COMPOSE](https://github.com/david-busbib/COMPOSE)
+
+## Model Description
+
+The COMPOSE architecture jointly encodes a paper's citation subgraph and its formal Mathlib4 theorem dependencies. 
+
+- **Encoder:** A dual-graph GNN encoder fuses these signals via bidirectional cross-attention.
+- **Decoder:** A DeepSeek-Math-7B decoder is conditioned on the fused graph representations to produce predicted mathematical claims in plain text.
+
+The model is designed to produce claims that are both grounded in prior work and motivated by formal mathematical structures.
+
+## Usage
+
+To use COMPOSE, follow the installation instructions in the [official GitHub repository](https://github.com/david-busbib/COMPOSE). You can run the model on a specific arXiv paper using the provided script:
+
+```bash
+python3 code/run_on_paper.py --arxiv_id <ARXIV_ID> --checkpoint path/to/compose_checkpoint.pt
+```
+
+## Results
+
+Experiments show that COMPOSE outperforms strong baselines on retrieval tasks for real future papers and achieves superior performance under LLM-judge evaluation, producing more grounded and mathematically richer outputs.
+
+## Citation
+
+```bibtex
+@article{busbib2026compose,
+  title={{COMPOSE}: Composing Future Theorems from Citations and Formal Structure},
+  author={Busbib, David and Werman, Michael},
+  journal={arXiv preprint arXiv:2605.30333},
+  year={2026}
+}
+```