FreeCtrl: Constructing Control Centers with Feedforward Layers for Learning-Free Controllable Text Generation

TL;DR

FreeCtrl tackles the CTG cost-performance dilemma by introducing a learning-free approach that uses FFN value vectors as controllable memory units. It builds per-attribute control centers through keyword-driven vector selection and employs a four-phase loop—initialization, monitoring, adaptation, and filtering—to dynamically steer LLM outputs without training. Across single- and multi-attribute tasks, FreeCtrl outperforms learning-free baselines and remains competitive with learning-based methods, while also improving inference speed. The work demonstrates that carefully managed, weight-based manipulation of FFN vectors can achieve high-quality, attribute-controlled generations with minimal resource expenditure, suggesting practical implications for deploying controllable text systems in data- and compute-constrained settings.

Abstract

Controllable text generation (CTG) seeks to craft texts adhering to specific attributes, traditionally employing learning-based techniques such as training, fine-tuning, or prefix-tuning with attribute-specific datasets. These approaches, while effective, demand extensive computational and data resources. In contrast, some proposed learning-free alternatives circumvent learning but often yield inferior results, exemplifying the fundamental machine learning trade-off between computational expense and model efficacy. To overcome these limitations, we propose FreeCtrl, a learning-free approach that dynamically adjusts the weights of selected feedforward neural network (FFN) vectors to steer the outputs of large language models (LLMs). FreeCtrl hinges on the principle that the weights of different FFN vectors influence the likelihood of different tokens appearing in the output. By identifying and adaptively adjusting the weights of attribute-related FFN vectors, FreeCtrl can control the output likelihood of attribute keywords in the generated content. Extensive experiments on single- and multi-attribute control reveal that the learning-free FreeCtrl outperforms other learning-free and learning-based methods, successfully resolving the dilemma between learning costs and model performance.

Figures (8)

• Figure 1: Trade-off between learning cost and performance for CTG. Learning-based methods excel in delivering superb results but demand significant training resources. Conversely, learning-free methods are more resource-efficient but tend to yield inferior performance. Numerical performance details are available in §\ref{['sec:exp']}.
• Figure 2: Convergence. As the value vector weight increases, its corresponding output distribution converges.
• Figure 3: Diversity. The percentage of top-k tokens in the whole vocabulary grows with increasing weights.
• Figure 4: Overview of FreeCtrl. For the target attribute "SPORTS", FreeCtrl initially identifies related keywords and value vectors to establish a control center. Throughout the generation phase, it dynamically adjusts the control center's weights based on real-time output monitoring, ensuring adaptive feedback for subsequent token generation. Finally, a filter is applied to verify compliance with the required attribute. Notably, the position $(a,b)$ specifies the layer number $a$ and the value vector's position $b$ within that layer.
• Figure 5: Influence of $k$ on topic control.