Steered Generation via Gradient Descent on Sparse Features

2025-02-25

AI Summary

This study addresses the challenge of precisely and continuously controlling the cognitive complexity of feedback generated by large language models (LLMs) in educational settings. We propose a fine-tuning-free, non-intrusive controllable generation method: a sparse autoencoder is trained on intermediate LLM layers to map query embeddings into interpretable, sparse latent representations; gradient-based optimization is then performed in this layer-specific latent space to steer feature activation, enabling fine-grained control over output style and cognitive difficulty. Our approach uniquely integrates sparse feature guidance with layer-wise latent-space gradient optimization, supporting both attention distribution modulation and query embedding reparameterization. Experiments demonstrate that the method systematically and continuously adjusts cognitive complexity across predefined levels while preserving semantic fidelity and textual fluency. Both human and automated evaluations confirm significantly improved control accuracy compared to baselines.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

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.

Problem

• Control LLM output via sparse autoencoders
• Adjust cognitive complexity in feedback
• Manipulate attention distribution using gradient descent

Innovation

• Sparse autoencoders modify LLMs
• Gradient descent guides embeddings
• Precise control over attention distribution