Researchers have developed a new reward-modeling procedure called Expected Value Alignment (EVA) for Large Language Models used with formal interactive theorem provers. EVA extracts continuous scores from the model's token distribution while keeping its surface output discrete. This is achieved by computing a continuous score as the expectation over the logits of corresponding anchor tokens. The authors instantiate EVA in Leibniz, a reward model for Lean 4 formal verification, and evaluate it against baseline models. The results show that EVA significantly reduces discretization artifacts while retaining interpretability. --- Why it matters: This matters to researchers in AI because it addresses the trade-off between continuous scoring and preserving textual rationales in generative models. By providing a way to extract continuous scores from discrete token distributions, EVA enables more accurate and interpretable evaluations of reasoning steps in formal verification systems. Source: https://arxiv.org/abs/2606.01160