Steer Like the LLM: Activation Steering that Mimics Prompting

Problem
This preprint addresses the gap in performance between activation steering methods and prompt-based steering in large language models (LLMs). While both techniques aim to influence model outputs during inference, existing activation steering methods have been shown to underperform relative to prompt-based approaches. The authors highlight that traditional activation steering does not adequately replicate the nuanced behavior of prompt steering, which selectively applies strong interventions to specific tokens while minimally affecting others.

Method
The authors propose a novel framework termed Prompt Steering Replacement (PSR), which reformulates prompt steering as a type of activation steering. The PSR models are designed to estimate token-specific steering coefficients derived from the model’s activations, thereby imitating the effects of prompt-based interventions. The training process involves distilling the successful behaviors of prompt steering into simpler, interpretable models. The authors conduct experiments across three steering benchmarks using multiple LLMs, focusing on high-coherence completions to evaluate the effectiveness of their approach. The PSR models leverage a combination of token-specific steering coefficients and activation patterns to enhance steering fidelity.

Results
The PSR models demonstrate superior performance compared to existing activation steering methods across all evaluated benchmarks. Specifically, on the AxBench and persona steering tasks, PSR models not only outperform traditional activation steering techniques but also show competitive results against direct prompting methods. The authors report effect sizes indicating that PSR models achieve a significant improvement in steering accuracy, with quantifiable metrics that highlight their robustness in maintaining coherence in generated completions. The results suggest that PSR can effectively bridge the performance gap between activation and prompt-based steering.

Limitations
The authors acknowledge several limitations in their work. First, the PSR framework may require extensive tuning of hyperparameters to achieve optimal performance across different models and tasks. Additionally, the reliance on token-specific steering coefficients may introduce complexity in scenarios with highly variable input distributions. The authors also note that while their approach improves upon existing methods, it may not fully replicate the flexibility and expressiveness of prompt-based steering in all contexts. An obvious limitation not discussed is the potential computational overhead introduced by the additional model complexity, which could impact inference speed in real-time applications.

Why it matters
This work has significant implications for the development of more effective steering techniques in LLMs, particularly in applications requiring fine-grained control over model outputs. By demonstrating that activation steering can be enhanced through the principles of prompt steering, the authors provide a pathway for future research to explore hybrid approaches that leverage the strengths of both methods. The findings could lead to advancements in interactive AI systems, conversational agents, and other applications where precise output manipulation is critical.

Authors: Geert Heyman, Frederik Vandeputte
Source: arXiv:2605.03907
URL: https://arxiv.org/abs/2605.03907v1