A safe and trustworthy use of Large Language Models (LLMs) requires an accurate expression of confidence in their answers. We introduce a novel Reinforcement Learning (RL) approach for LLM calibration that fine-tunes LLMs to elicit calibrated confidence estimations in their answers to factual questions. We model the problem as a betting game where the model predicts a confidence score together with every answer, and design a reward function that penalizes both over and under-confidence. We prove that under our reward design an optimal policy would result in a perfectly calibrated confidence estimation. Our experiments demonstrate significantly improved confidence calibration and generalization to new tasks without re-training, indicating that our approach teaches a general confidence awareness. This approach enables the training of inherently calibrated LLMs.

Every day, poison control centers (PCC) are called for immediate classification and treatment recommendations if an acute intoxication is suspected. Due to the time-sensitive nature of these cases, doctors are required to propose a correct diagnosis and intervention within a minimal time frame. Usually the toxin is known and recommendations can be made accordingly. However, in challenging cases only symptoms are mentioned and doctors have to rely on their clinical experience. Medical experts and our analyses of a regional dataset of intoxication records provide evidence that this is challenging, since occurring symptoms may not always match the textbook description due to regional distinctions, inter-rater variance, and institutional workflow. Computer-aided diagnosis (CADx) can provide decision support, but approaches so far do not consider additional information of the reported cases like age or gender, despite their potential value towards a correct diagnosis. In this work, we propose a new machine learning based CADx method which fuses symptoms and meta information of the patients using graph convolutional networks. We further propose a novel symptom matching method that allows the effective incorporation of prior knowledge into the learning process and evidently stabilizes the poison prediction. We validate our method against 10 medical doctors with different experience diagnosing intoxication cases for 10 different toxins from the PCC in Munich and show our method's superiority in performance for poison prediction.