Less is Enough: Synthesizing Diverse Data in Feature Space of LLMs
This work addresses the challenge of optimizing post-training data for LLMs, offering a practical methodology for data-centric improvements, though it is incremental as it builds on existing diversity metrics and synthesis techniques.
The authors tackled the problem of improving downstream performance in large language models by enhancing data diversity, introducing Feature Activation Coverage (FAC) to measure diversity in an interpretable feature space and a synthesis framework that generates synthetic samples to fill missing features, resulting in consistent improvements in diversity and performance across tasks like instruction following and toxicity detection.
The diversity of post-training data is critical for effective downstream performance in large language models (LLMs). Many existing approaches to constructing post-training data quantify diversity using text-based metrics that capture linguistic variation, but such metrics provide only weak signals for the task-relevant features that determine downstream performance. In this work, we introduce Feature Activation Coverage (FAC) which measures data diversity in an interpretable feature space. Building upon this metric, we further propose a diversity-driven data synthesis framework, named FAC Synthesis, that first uses a sparse autoencoder to identify missing features from a seed dataset, and then generates synthetic samples that explicitly reflect these features. Experiments show that our approach consistently improves both data diversity and downstream performance on various tasks, including instruction following, toxicity detection, reward modeling, and behavior steering. Interestingly, we identify a shared, interpretable feature space across model families (i.e., LLaMA, Mistral, and Qwen), enabling cross-model knowledge transfer. Our work provides a solid and practical methodology for exploring data-centric optimization of LLMs.