Painless Activation Steering Automates LLM Post-Training

Researchers introduce Painless Activation Steering (PAS), a fully automated method for post-training large language models that eliminates the need for hand-crafted prompt pairs or labor-intensive feature annotation. PAS works with any labeled dataset, making activation steering as convenient as plug-and-play methods like Reinforcement Learning and Supervised Fine-Tuning. The method was evaluated on three open-weight models: Llama3.1-8B-Instruct, DeepSeek-R1-Distill-8B, and Nou. Activation steering previously required manual trial-and-error, while weight-based post-training is time-consuming and expensive. PAS automates the process, offering a cheap, fast, and controllable alternative. The paper is available on arXiv under identifier 2509.22739.