Large Language Models (LLMs) have demonstrated remarkable performance in natural language generation tasks, yet their uncontrolled outputs pose significant ethical and safety risks. Recently, representation engineering methods have shown promising results in steering model behavior by modifying the rich semantic information encoded in activation vectors. However, due to the difficulty of precisely disentangling semantic directions within high-dimensional representation space, existing approaches suffer from three major limitations: lack of fine-grained control, quality degradation of generated content, and poor interpretability. To address these challenges, we propose a sparse encoding-based representation engineering method, named SRE, which decomposes polysemantic activations into a structured, monosemantic feature space. By leveraging sparse autoencoding, our approach isolates and adjusts only task-specific sparse feature dimensions, enabling precise and interpretable steering of model behavior while preserving content quality. We validate our method on three critical domains, i.e., safety, fairness, and truthfulness using the open-source LLM Gemma-2-2B-it. Experimental results show that SRE achieves superior controllability while maintaining the overall quality of generated content (i.e., controllability and quality), demonstrating its effectiveness as a fine-grained and interpretable activation steering framework.

The increasing prevalence of Large Language Models (LMs) in critical applications highlights the need for controlled language generation strategies that are not only computationally efficient but that also enjoy performance guarantees. To achieve this, we use a common model of concept semantics as linearly represented in an LM's latent space. In particular, we take the view that natural language generation traces a trajectory in this continuous semantic space, realized by the language model's hidden activations. This view permits a control-theoretic treatment of text generation in latent space, in which we propose a lightweight, gradient-free intervention that dynamically steers trajectories away from regions corresponding to undesired meanings. Crucially, we show that this intervention, which we compute in closed form, is guaranteed (in probability) to steer the output into the allowed region. Finally, we demonstrate on a toxicity avoidance objective that the intervention steers language away from undesired content while maintaining text quality.

Recent work in activation steering has demonstrated the potential to better control the outputs of Large Language Models (LLMs), but it involves finding steering vectors. This is difficult because engineers do not typically know how features are represented in these models. We seek to address this issue by applying the idea of mean-centring to steering vectors. We find that taking the average of activations associated with a target dataset, and then subtracting the mean of all training activations, results in effective steering vectors. We test this method on a variety of models on natural language tasks by steering away from generating toxic text, and steering the completion of a story towards a target genre. We also apply mean-centring to extract function vectors, more effectively triggering the execution of a range of natural language tasks by a significant margin (compared to previous baselines). This suggests that mean-centring can be used to easily improve the effectiveness of activation steering in a wide range of contexts.

Reliably controlling the behavior of large language models is a pressing open problem. Existing methods include supervised finetuning, reinforcement learning from human feedback, prompt engineering and guided decoding. We instead investigate activation engineering: modifying activations at inference-time to predictably alter model behavior. We bias the forward pass with a 'steering vector' implicitly specified through natural language. Past work learned these steering vectors; our Activation Addition (ActAdd) method instead computes them by taking the activation differences which result from pairs of prompts. We demonstrate ActAdd on GPT-2 on OpenWebText and ConceptNet, and replicate the effect on Llama-13B and GPT-J-6B. Our approach yields inference-time control over high-level properties of output & preserves performance on off-target topics. The method requires far less compute and implementation effort than finetuning and RLHF, allows for natural language specification by users, and its overhead scales naturally with model size.