Dynamic Important Example Mining for Reinforcement Finetuning

DIEM treats training examples as changing in value during reinforcement fine-tuning.

The paper argues that static or heuristic sample selection misses how policy learning shifts over training. Its method estimates each sample’s marginal contribution through gradient alignment, then reweights batches under a constraint meant to keep optimization stable. The authors report gains over static and dynamic baselines on several reasoning benchmarks. Code is promised but not included in the note. HF Daily Papers' note