AI21 Labs publishes LIFT-Bench and LIFT-Train to reduce reward hacking in instruction-following RL
AI21 Labs released LIFT-Bench on October 25, 2026, plus LIFT-Train (34,258 tasks) and an open pipeline; the suite scopes each constraint to a single document section and uses the same judge for evaluation and RL reward to reduce reward hacking seen with prior flat benchmarks.
Summary
AI-assisted summary of the linked sourceAI21 Labs released LIFT-Bench on October 25, 2026, plus LIFT-Train (34,258 tasks) and an open pipeline; the suite scopes each constraint to a single document section and uses the same judge for evaluation and RL reward to reduce reward hacking seen with prior flat benchmarks.