Latent JEPA predicts future views via joint embedding, improving molecular optimization and editing and reaction metrics on ChemCoTBench

Latent JEPA is proposed, combining autoregressive learning with joint-embedding prediction of one or more future views to train continuous latent thoughts to anticipate informative aspects of future solutions without verbalizing every intermediate step. Relative to approaches that explicitly unfold intermediate reasoning steps, the framework introduces abstract, intuition-like expectations as a learning signal inside latent reasoning. A framework description at the abstract level; no network architecture, training scale, or ablation details are given.

Textual and molecular prediction objectives are developed for chemical reasoning, connecting latent thoughts to both subsequent reasoning and molecular outcomes. It extends future-prediction objectives from purely textual reasoning to molecular-level outcome prediction, directly linking latent representations to chemical results. The abstract states that these two objectives exist but does not report their functional forms or weighting.

Experiments on ChemCoTBench show gains in molecular optimization and on several editing and reaction metrics. It provides empirical support for the framework on a chemical multi-step reasoning benchmark rather than only a conceptual argument. The abstract only states 'gains'; no specific values, baseline margins, or statistical tests are given.

Representation analyses show that future prediction makes latent thoughts more informative about molecular outcomes and strengthens their correspondence with chemical structure. It explains performance changes at the representation level, attributing gains to the molecular-outcome information and structural correspondence carried by latent thoughts. An analysis-level conclusion in the abstract; the probes, metrics, or control settings are not described.

This work targets researchers and practitioners who need multi-step chemical reasoning and molecular optimization; the setting is chemical knowledge and multi-step problem solving with large language models, measured on benchmarks such as ChemCoTBench for molecular optimization, editing, and reaction metrics. Its value is a training principle: letting continuous latent thoughts carry intuition-like chemical expectations by predicting future views, thereby connecting the reasoning process to molecular outcomes without verbalizing all intermediate steps. For readers, this means abstract future prediction can be considered a candidate component when designing latent reasoning objectives, with attention to how it pairs with textual reasoning objectives and molecular prediction objectives.

What is available is abstract-level information, without specific values, baseline margins, statistical tests, model scale, or ablations, so the size of the gains and the relative contribution of each design choice remain to be confirmed. The weighting and interaction of the textual and molecular prediction objectives, the effect of the number of future views, and the applicability of this learning principle to scientific fields beyond chemistry are open questions for readers to watch. The probes and metrics used in the representation analyses are also not described in the abstract, so the robustness of those conclusions awaits verification in the methods section.