# Chemical Reasoning

The authors introduce Latent JEPA, a framework combining autoregressive learning with joint-embedding prediction of one or more future views, and develop textual and molecular prediction objectives for chemical reasoning so that continuous latent thoughts anticipate subsequent reasoning and molecular outcomes without verbalizing every intermediate step; on ChemCoTBench they report gains in molecular optimization and on several editing and reaction metrics, and representation analyses indicate that future prediction makes latent thoughts more informative about molecular outcomes and strengthens their correspondence with chemical structure.

The authors introduce Latent JEPA, a framework combining autoregressive learning with joint-embedding prediction of one or more future views, and develop textual and molecular prediction objectives for chemical reasoning so that continuous latent thoughts anticipate subsequent reasoning and molecular outcomes without verbalizing every intermediate step; on ChemCoTBench they report gains in molecular optimization and on several editing and reaction metrics, and representation analyses indicate that future prediction makes latent thoughts more informative about molecular outcomes and strengthens their correspondence with chemical structure.