Researchers have published a framework that lets large language models train themselves more effectively by treating hard, easy, and borderline problems differently.
Called DRIFT, the system combines two mechanisms: Difficulty Routing, which classifies each training problem by how well the model is currently handling it and then decides whether to apply self-distillation or reinforcement learning signals, and Rhythm Gating, which focuses exploration on the specific token positions where reasoning is most likely to go wrong. A success buffer preserves high-quality past examples, and a two-stage curriculum moves the model from acquiring reliable behavior first before pushing toward stable policy evolution. Evaluated across five benchmarks at three different model scales, DRIFT hit an average score of 79.5%, beating GRPO by 9.5 percentage points and SDPO by 7.5 percentage points. On tool-use tasks specifically, it outpaced GRPO by 13.5 points and SDPO by 10.7 points.
The core problem DRIFT addresses is one the field has quietly struggled with: standard self-improvement loops tend to waste compute hammering easy problems while getting nearly useless signal from very hard ones, leaving the middle ground underexplored. A method that routes training signal based on per-problem learning state is a more principled approach than the flat sampling most current methods use.
The results are from an arXiv preprint, so independent replication is still ahead — but the margin over GRPO, itself a well-regarded baseline, is large enough that this is worth watching before the next training run budget gets locked in.