Steered Generation via Gradient Descent on Sparse Features
This work addresses the need for controlled text generation in applications like education, though it is incremental as it builds on existing methods for representation manipulation.
The paper tackled the problem of steering large language model outputs toward specific target characteristics by modifying internal representations via sparse autoencoders and gradient-based optimization, resulting in systematic adjustment of cognitive complexity in educational feedback.
Large language models (LLMs) encode a diverse range of linguistic features within their latent representations, which can be harnessed to steer their output toward specific target characteristics. In this paper, we modify the internal structure of LLMs by training sparse autoencoders to learn a sparse representation of the query embedding, allowing precise control over the model's attention distribution. We demonstrate that manipulating this sparse representation effectively transforms the output toward different stylistic and cognitive targets. Specifically, in an educational setting, we show that the cognitive complexity of LLM-generated feedback can be systematically adjusted by modifying the encoded query representation at a specific layer. To achieve this, we guide the learned sparse embedding toward the representation of samples from the desired cognitive complexity level, using gradient-based optimization in the latent space.