LIFT: Partitioned constraints for evaluating and training instruction following

The researchers propose the Lift-Bench instruction-following benchmark, which requires models to draft documents based on a given topic and structured outline, with constraints confined to specific document sections, including both verifiable and judged constraint types, and accompanied by a training set, Lift-Train. In response to the problem that verifiable constraints in common benchmarks such as IFEval are used as reinforcement learning rewards and may induce constraint reward hacking, this benchmark provides both a main task and instruction-following constraints to mitigate shallow reward hacking. Multi-model experiments show that frontier models struggle on the harder subset, and the authors find that training with Lift-Train is less prone to reward hacking than common IFTrain. The benchmark, training set and data generation pipeline have been made public.