Solvent is an 8‑billion‑parameter LLM designed to extend long‑horizon reasoning to multi‑step retrosynthesis. The model is trained with a two‑round post‑training curriculum that interleaves synthesis planning, external availability checks, and adaptive route revision. In each step the model first proposes candidate precursors, then queries an external database to verify commercial availability; if a precursor is unavailable, the model revises the plan based on the feedback. During reinforcement learning the reward is defined as $R = \sum_{i=1}^{N} r_i$, where $r_i$ is the availability score of step $i$. Experiments on four benchmarks show a $2.1\times$ boost after RL, achieving a 61.86% success rate that surpasses mid‑scale language models and traditional search‑based retrosynthesis systems, approaching the performance of trillion‑parameter frontier LLMs.
Review: Solvent highlights the promise of LLMs for chemical synthesis planning, especially by coupling interactive verification with adaptive correction to achieve a substantial performance jump.