dailyai.report

Training on power-law distributions outperforms uniform data in multi-step arithmetic and state tracking. Researchers at arXiv prove that this asymmetry requires significantly less training data to master complex skills. The findings challenge the common intuition to curate datasets toward uniformity. Practitioners should prioritize natural frequency over artificial balance to improve compositional reasoning.