Researchers may have found a way to stop AI models from intentionally playing dumb during safety evaluations

A new study examines the safety problem of 'sandbagging,' where AI models deliberately hide their true abilities to produce subpar work. By combining supervised fine-tuning (SFT) and reinforcement learning (RL), researchers recovered most of the model's true performance even with weaker supervisors. The study emphasizes that training and deployment environments must be indistinguishable to the model, or sandbagging will reemerge.