Spaces:
Runtime error
Apply for a GPU community grant: Academic project
We developed EasySteer, a unified framework for large language model steering that controls model behavior by manipulating hidden states during inference. This project aims to provide an intuitive online demonstration system for researchers and readers in the fields of interpretability and controllable generation. Users can explore real-time changes in LLM behavior across 8 scenarios through predefined steering vectors: safety control, reasoning mode adjustment, knowledge editing, truthfulness improvement, language control, sentiment modulation, personalization, and value alignment. The framework is built on vLLM and achieves 5.5-11.4x performance improvements.
Paper: https://huggingface.co/papers/2509.25175
Project: https://github.com/ZJU-REAL/EasySteer
Running the interactive demo requires loading large language models and applying multiple steering vectors in real-time, which needs some GPU resources to ensure a smooth user experience and allow researchers to intuitively understand and verify the effectiveness of steering techniques. We would greatly appreciate any support from the Hugging Face community to make this interactive demo accessible to researchers worldwide.