Researchers at MIT and the MIT-IBM Watson AI Lab have developed a novel calibration method for large language models (LLMs), named Thermometer. This method addresses the challenge of LLMs generating inaccurate or overconfident responses across a wide range of tasks. Unlike traditional calibration techniques that are task-specific and computationally intensive, Thermometer introduces an auxiliary model that adjusts the LLM’s confidence levels to better match its prediction accuracy, without requiring extensive computation or altering the model’s accuracy.
Thermometer leverages temperature scaling, a classical calibration technique, to efficiently calibrate LLMs for new tasks without needing task-specific labeled datasets. This is particularly useful for applications where acquiring such data is impractical. The method has shown promise in producing better-calibrated uncertainty measures across various tasks with minimal computational overhead. The researchers aim to further refine Thermometer to handle more complex text-generation tasks and apply it to larger LLMs, potentially making LLMs more reliable and trustworthy for users across diverse applications. This breakthrough was presented at the International Conference on Machine Learning, highlighting its potential to enhance the versatility and effectiveness of LLMs in real-world scenarios.
Read more at MIT News | Massachusetts Institute of Technology…