Recently, the Jev model, unveiled by TypeSafe AI, has drawn considerable attention. This is due to its remarkable capability of directly generating decisions that can be invoked by code, along with their associated probabilities. The essence of this model closely aligns with the ideas put forth in the ConfTuner paper, which has been chosen for presentation at NeurIPS 2025. Both the Jev model and the ConfTuner paper place a strong emphasis on modeling and calibrating the entire candidate probability distribution.

ConfTuner, proposed by a research team from the National University of Singapore, utilizes the Tokenized Brier Score method. This innovative approach requires only right/wrong labels for answers to train the model. As a result, the model's output confidence can more accurately mirror the true accuracy rate. This loss function is not only theoretically well-founded but also backed by practical experiments. These experiments show that it can significantly reduce the Expected Calibration Error, thereby achieving high training efficiency. Furthermore, the calibrated probabilities can boost the accuracy of downstream tasks.

Its extended open-source project, JevTuner, takes this exploration a step further. It delves into applying this method to real-world business decision-making scenarios. By doing so, it offers a practical technical route for using candidate probabilities as the primary output of the model. This, in turn, supports the seamless integration of AI decision-making into automated workflows.