Linearly Controlled Language Generation with Performative Guarantees

TL;DR

This paper introduces a control-theoretic method for guiding language model outputs towards desired semantics with guarantees, using online, gradient-free interventions in the model's latent space to ensure safe and attribute-controlled text generation.

Contribution

It presents a novel, mathematically grounded approach that applies control theory to steer language model activations in real-time, providing performance guarantees and minimal impact on generation speed.

Findings

Effective control of toxicity and sentiment in generated text.

Maintains high text quality while enforcing semantic constraints.

Interventions are computed in closed-form, ensuring efficiency.

Abstract

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. In particular, we propose to directly intervene the activations of the token that is being generated in embedding space in an online fashion. Crucially, we do not simply steer activations towards a desirable region. Instead, our method relies on classical techniques from control theory to precisely control activations in a context-dependent way, and guarantees that they are brought into a specific pre-defined region of embedding space that corresponds to allowed semantics. Our intervention is computed in closed-form according to an optimal controller formulation, minimally impacting generation time. This control of the activations in embedding space allows for fine-grained steering of attributes of the generated sequence. We demonstrate the effectiveness of our approach on different objectives -- toxicity avoidance and sentiment control -- while maintaining text quality.