MARC-TS: Learning Continuous Reaction Paths for Transition-State Prediction

This paper presents MARC-TS, a two-stage machine learning framework for transition-state prediction that learns continuous endpoint-conditioned reaction paths. The framework enables querying paths at arbitrary resolution and refines transition-state candidates using local path context. The authors introduce T1x-IRC-8K, a dataset containing 8,209 chemical reactions with 1,088,725 path-resolved geometries, designed for training and evaluating reaction path models in the quantum chemistry domain.