Hard Examples Are All You Need: Maximizing GRPO Post-Training Under Annotation Budgets (2508.14094v2)

Abstract: Collecting high-quality training examples for LLM fine-tuning is expensive, with practical budgets limiting the amount of data that can be procured. We investigate a critical question for resource-constrained alignment: under a fixed acquisition budget, should practitioners prioritize examples that are easy, medium, hard, or of random difficulty? We study Group Relative Policy Optimization (GRPO) fine-tuning across different model sizes and families, comparing four subset selection policies chosen from the same unlabeled pool using base-model difficulty estimates obtained via multi-sample evaluation. Our experiments reveal that training on the hardest examples yields the largest performance gains, up to 47%, while training on easy examples yield the smallest gains. Analysis reveals that this effect arises from harder examples providing more learnable opportunities during GRPO training. These findings provide practical guidance for budget-constrained post-training: prioritizing hard examples yields substantial performance gains on reasoning tasks when using GRPO.