Spaces:

---

Schrieffer

/

SARM-Demo

like 0

Running on Zero

Project: Interactive Demo for SARM-4B Interpretable Reward Model

This is a request for a GPU grant for our Hugging Face Space, which hosts the official interactive demo for our SARM-4B model. SARM (Sparse Autoencoder-enhanced Reward Model) is a novel architecture introduced in our research paper, "Interpretable Reward Model via Sparse Autoencoder" (arXiv:2508.08746), aimed at improving the transparency and interpretability of reward models used in LLM alignment.

Our work makes reward models less of a "black box" by allowing for feature-level analysis of AI preferences. Importantly, our SARM-4B model achieves state-of-the-art performance on the RewardBench 2 benchmark, demonstrating that interpretability can be achieved without sacrificing performance.

The SARM-4B model is based on Llama-3.1-8B and is over 9GB in size. Running inference on a CPU is extremely slow, leading to a poor user experience and making it difficult for the community to explore our research. A ZeroGPU grant would allow us to provide a fast, responsive demo, making our open-source contribution to AI alignment much more accessible to researchers and developers.

Thank you for your consideration and support for the open-source AI community.