Solvent preprint reports 61.86% score across molecular synthesis planning benchmarks

Researchers report that Solvent, an eight-billion-parameter language model, scored 61.86% across four benchmarks for planning how to make molecules from purchasable building blocks. The preprint reports a 2.1-fold performance improvement from reinforcement learning and results exceeding prior search-based systems and mid-scale language models. The model learns through two rounds of post-training to alternate between planning reaction sequences, checking building-block availability and revising routes. The reported achievement concerns synthesis planning benchmarks, rather than demonstrated production of the target molecules.