A new framework called ReCurveflow has been proposed to predict transition states in chemical reactions. Unlike previous methods that focused on straight linear paths, ReCurveflow learns curved reaction trajectories by interpolating molecular geometries from a full NEB-derived band. This allows for more accurate predictions and resistance against exposure bias. The framework also includes an off-path correction mechanism to improve performance when the predicted geometry state is not aligned with the reference path. --- Why it matters: This matters because predicting transition states in chemical reactions is crucial for understanding reaction mechanisms, and ReCurveflow's improved accuracy can lead to better results in fields like materials science and chemistry. Source: https://arxiv.org/abs/2608.20869