Add model card and link to paper

README.md

@@ -1,3 +1,68 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+pipeline_tag: feature-extraction
+---
+
+# Latent Memory 🧠
+
+This repository contains the official implementation and artifacts for the paper [One Token per Multimodal Evidence: Latent Memory for Resource-Constrained QA](https://huggingface.co/papers/2606.10572).
+
+Latent Memory is a latent-space memory paradigm that replaces raw text or image evidence items with a single high-dimensional latent token produced by a small compressor LLM/VLM. By retrieving and prompting these latent tokens directly to a generator model, it achieves competitive QA performance while significantly reducing token consumption and storage pressure.
+
+- **Paper:** [One Token per Multimodal Evidence: Latent Memory for Resource-Constrained QA](https://huggingface.co/papers/2606.10572)
+- **GitHub Repository:** [zz1358m/Latent-Memory-Master](https://github.com/zz1358m/Latent-Memory-Master)
+
+## Installation ⚙️
+
+```bash
+pip install -r requirements.txt
+```
+
+## Usage 🚀
+
+The framework provides unified entry points for different experiment families, including text-only (LLaMA/Mistral) and multimodal (LLaVA/Gemma) setups.
+
+### Training
+
+To train a compressor for a specific task:
+
+```bash
+# Text-only
+python scripts/train_release.py --task text  --config configs/config.yaml
+
+# LLaVA multimodal
+python scripts/train_release.py --task llava --config configs/config_llava.yaml
+
+# Gemma multimodal
+python scripts/train_release.py --task gemma --config configs/config_gemma3_4B_12B.yaml
+```
+
+### Evaluation
+
+To evaluate a trained checkpoint:
+
+```bash
+# Evaluate text-only latent retrieval
+python scripts/eval_release.py --task text --config configs/config.yaml --checkpoint checkpoints/text_llama/model.pt --output results/text_llama --top_k 5
+
+# Evaluate on multiple datasets (e.g. HotpotQA, 2WikiMultiHopQA, MuSiQue)
+python scripts/eval_release.py --task text \
+  --config configs/config_llama1b_mistral7b.yaml \
+  --checkpoint checkpoints/text_llama_mistral/model_best.pt \
+  --output results/text_llama_mistral \
+  --dataset hotpotqa \
+  --ood_datasets 2wikimultihopqa,musique \
+  --k_list 1,2,5 \
+  --skip_baselines
+```
+
+## Citation
+
+```bibtex
+@article{zheng2024one,
+  title={One Token per Multimodal Evidence: Latent Memory for Resource-Constrained QA},
+  author={Zheng, Zhi and Meng, Ziqiao and Luan, Hao and Liu, Wei and Lee, Wee Sun},
+  journal={arXiv preprint arXiv:2606.10572},
+  year={2024}
+}
+```