MIT News July 31, 2024
Recent studies have found that common interventions such as instruction tuning often result in poorly calibrated large language models (LLMs). Although calibration is well-explored in traditional applications, calibrating LLMs is uniquely challenging. The challenges stem as much from the severe computational requirements of LLMs as from their versatility, which allows them to be applied to diverse tasks. To address these challenges, researchers at MIT proposed THERMOMETER, a calibration approach tailored to LLMs. For calibrating the LLMTHERMOMETER learned an auxiliary model, using the data given from multiple tasks. According to the researchers it was computationally efficient, preserved the accuracy of the LLM, and produced better-calibrated responses for new tasks. Extensive empirical evaluations across various benchmarks demonstrated the effectiveness of the proposed method… read more. Open Access TECHNICAL ARTICLE
Thermometer… could help users pinpoint situations where a model is overconfident about false predictions. Credit: MIT iStock.