MIT CSAIL's RLCR method reduces AI hallucination calibration error by 90% by teaching models to say 'I'm not sure'

Researchers from MIT's Computer Science and Artificial Intelligence Laboratory developed RLCR (Reinforcement Learning with Calibration Rewards), a training method that teaches language models to produce calibrated confidence estimates alongside their answers. The method reduces calibration error by up to 90% while maintaining or improving accuracy on both trained and novel tasks — directly addressing the hallucination and overconfidence problems that limit enterprise AI deployment.