AI21 Labs releases LIFT-Bench to curb reward hacking in RL training

AI21 Labs released LIFT-Bench, an instruction-following benchmark with constraints scoped to individual document sections, plus LIFT-Train, a 34,258-task training set. On IFTrain the median answer shrank from 114 to 18 words, while on LIFT-Train it stayed around 390–530 words.

• LIFT-Bench has 784 test tasks across 10 splits of short and long documents
• LIFT-Train holds 34,258 tasks; one constraint_reward scalar serves as metric and reward
• On IFTrain the median answer fell from 114 to 18 words by step 390
• GPT-5.5 scores 0.96 on structured tasks but 0.65–0.72 on flat ones