Study challenges the conventional wisdom that teaching AI models from easy to hard problems improves reasoning abilities

Researchers conducted systematic tests on curriculum learning (CL) for large language model post-training using synthetic arithmetic and logical benchmarks

Surprisingly found no consistent advantage to difficulty-based sequencing compared to random sampling across multiple model families and curriculum schedules

Results held true for both supervised fine-tuning (SFT) and reinforcement learning (RL) approaches

The findings suggest that the intuitive appeal of easy-to-hard learning may not translate to actual performance gains in compositional reasoning tasks