Update README.md

README.md

@@ -1,10 +1,32 @@
 ---
-title: README
-emoji: 🔥
-colorFrom: pink
-colorTo: purple
-sdk: static
-pinned: false
+title: IP-GRM for Creative Writing
 ---
 
-Edit this `README.md` markdown file to author your organization card.
+# IP-GRM: Unbiased Principles, Robust Rewards
+
+IP-GRM (Independent Principle Generative Reward Model) is a decoupled reward-modeling framework for open-ended RLHF tasks. It explicitly separates **principle generation** from **response judging**, reducing response-conditioned bias (*Principle Drift*) and improving reward robustness in GRPO training.
+
+## Resources
+
+| Resource | Description |
+|----------|-------------|
+| [IP-GRM](https://huggingface.co/IP-GRM/IP-GRM) | 16B generative reward model with decoupled principle-judgment pipeline |
+| [CreativeWriting-8B](https://huggingface.co/IP-GRM/CreativeWriting-8B) | 8B creative writing model trained via GRPO with IP-GRM rewards |
+| [IP-rewarding-8K](https://huggingface.co/datasets/IP-GRM/IP-rewarding-8K) | 8K decoupled reward SFT dataset (principle + judgment pairs) |
+| [Paper](https://arxiv.org/abs/2602.10019) | arXiv preprint |
+| [Code](https://github.com/IP-GRM/IP-GRM) | Training scripts and IP-GRM process functions |
+
+## Key Idea
+
+Standard generative reward models (GRMs) couple principle generation with response observation, causing **Principle Drift** — the evaluation criteria shift to accommodate the response being judged. IP-GRM eliminates this bias through a two-stage factorization:
+
+- **Stage 1** `P(P | Q)` — generate evaluation principles from the question only
+- **Stage 2** `P(J, r | Q, P, R)` — judge the response under pre-defined principles
+
+This ensures conditional independence `I(P; R | Q) = 0`, and enables **Principle Cache** — generating principles once per prompt and reusing them across all sampled responses in a GRPO group.
+
+## Results
+
+- **WritingBench 87.6** / **CW-v3 77.8** with Qwen3-8B + IP-GRM (competitive with GPT-5.2 and Claude-Sonnet-4)
+- **23.66% faster** reward computation than baseline GRM via Principle Cache
+