Prompt mutations can expand training coverage without drifting away from the original task distribution, improving robustness with fewer examples.
Robust, sample-efficient SFT with prompt mutations
Low-divergence prompt variation for better performance at lower cost.