ReactionLLM: Data-Centric Learning for a Compact Single-step Chemical reaction prediction model

Ran Li, Shimin Di, Haowei LI, Luanshi Bu, Jiachuan Wang, Lei Chen

Keywords: Retrosynthesis, Reactions, AI for Chemistry

Abstract: Chemical reaction prediction faces three fundamental challenges that limit practical deployment: (1) ineffective molecular representations that fail to capture essential chemical context, (2) unfair comparison with test-time augmentation and without mention the usage of AAM(atom-atom mapping, and (3) unsatisfactory performance of large-scale pretrained models. To address these limitations, we present a unified framework that enables a compact 0.5B parameter model to outperform significantly larger counterparts (7B/13B parameters) through three strategic innovations: the AAM-0 molecular representation that bridges mapped and unmapped data via implicit contrastive learning; bidirectional multi-task learning that creates a unified chemical representation space across retrosynthesis and forward prediction tasks; and structured plan-based reasoning that ensures chemically plausible step-by-step rationalizations. Extensive evaluation with rigorous separate assessment of mapped and unmapped performance demonstrates +14\% accuracy improvement over strong baselines, establishing that carefully designed compact models with built-in chemical intelligence can surpass larger, less specialized alternatives while maintaining computational efficiency.The implementation is available at: \url{https://anonymous.4open.science/r/ReactionLLM-DF4C}.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 9483