Want Better Synthetic Data? Steer It: Activation Steering for Low-Resource Language Generation

In three linesActivation steering improves synthetic data generation for low-resource languages. Two strategies tested: Language Steering (linguistic identity) and Quality Steering (well-formedness). Evaluation across 4 open-source LLMs, 11 languages, classification tasks. Early-layer steering increases diversity and downstream performance.

Summary generated by Claude — human-verified