Steer Like the LLM: Activation Steering that Mimics Prompting

Large language models can be steered at inference time through prompting or activation interventions, but activation steering methods often underperform compared to prompt-based approaches. We investigate whether activation steering can be improved by learning to mimic the interventions that prompt steering triggers within the model. To this end, we introduce Prompt Steering Replacement (PSR) models, a new family of activation steering methods that distill prompt steering behavior into interpretable interventions on model activations. A PSR is an activation steering method that estimates position-specific steering coefficients and is trained to imitate prompt-based interventions. Experiments on persona steering and instruction following across multiple language models demonstrate that PSR models consistently outperform constant-coefficient interventions that are frequently used in the literature and achieve performance close to or exceeding prompt steering while maintaining interpretability.

When we want an AI language model to behave a certain way — for instance, to be more concise — we typically rely on prompting. Prompting is flexible, but fragile: getting a prompt to reliably produce the desired behavior is often hit-or-miss, and prompts can also be hijacked by malicious users or forgotten in long conversations. An alternative is activation steering: instead of instructing with words, we reach inside the model and directly adjust its internal computations, analogous to a neuroscientist who stimulates a brain region to trigger a behavior. However, existing activation steering methods consistently underperform prompting, even on simple instructions. We show that prompting can itself be viewed as a form of activation steering, though a complex and hard-to-interpret one. This insight motivates Prompt Steering Replacement (PSR): distilling the effect of a prompt into a simpler, more interpretable activation steering method. Because PSR learns from the cases where prompting succeeds, it can in principle outperform prompting on average. We find that popular activation steering methods are incompatible with what prompting does internally, and use this analysis to design a better method. Across three benchmarks and multiple language models, our approach consistently outperforms existing activation steering methods and matches or exceeds prompting performance. The current work targets single instructions rather than full system prompts, so it is not yet a drop-in replacement for prompting — but it is a promising step in that direction.