LLM Steer is an open-source Python module for steering large language model outputs toward a chosen topic or behavior using activation engineering. Published on GitHub under the MIT license with about 280 stars and updated in January 2026, it adds steering vectors to model layers to shape responses without retraining.
The library works alongside Hugging Face Transformers. You wrap a model and tokenizer in the Steer class, then add a steering vector on a particular layer with a coefficient and a piece of text describing the direction, such as logical. Coefficients can be negative, multiple vectors can be stacked on the same layer or spread across layers, and everything can be listed or reset to return the model to its original behavior. The author positions steering as an enhancer that complements a system prompt rather than replacing it, useful for making models answer tricky logical puzzles correctly more often, strengthening role-play personas, or suppressing stock phrases. It has been tested on architectures including Llama, Mistral, Phi, and StableLM, though not on GGUF models.
The README’s question and answer section is refreshingly practical: it suggests starting with middle layers and a small coefficient, increasing slowly, and warns that too strong a vector produces gibberish. Demos are provided as a Colab notebook for the first version and a hosted Hugging Face space for the second, so the effect of a steering vector can be tried before writing any code.
LLM Steer is free and open source under the MIT license. There is no paid tier; you install it from PyPI and run it against your own local models.
LLM Steer is for researchers and tinkerers exploring activation engineering who want a small, practical API for experimenting with how steering vectors change model behavior.