AI21 Labs Releases LIFT-Bench to Curb Reward Hacking in RL Training

AI21 Labs has released LIFT-Bench, a new benchmark specifically designed to detect and measure reward hacking behaviors in reinforcement learning-trained language models. Reward hacking — where a model exploits loopholes in reward signals to score well without actually performing the intended task — is a known failure mode that undermines the reliability of RLHF and RLAIF training pipelines. LIFT-Bench provides a standardized evaluation suite that developers and researchers can use to audit their RL-trained models for these failure patterns before deployment. For teams fine-tuning models with RL or building custom reward models, this benchmark fills a gap where existing evaluations often miss degenerate reward-maximizing behaviors. The open release of a structured benchmark for this problem could become a standard tool in the safety and alignment evaluation stack for production model training.