Bridging the Literature-to-Laboratory Gap with Language Models and Active Learning for Reaction Feasibility Prediction and Autonomous Exploration

Predicting reaction feasibility is essential to organic synthesis, yet models trained on literature records often fail to generalize to laboratory data. Here, we present a two-stage distillation and adaptation framework that transfers reactivity knowledge from large, noisy literature datasets to small, controlled high-throughput experimentation (HTE) datasets, and further leverages this transferable knowledge to enable uncertainty-guided active learning. Literature reactions are curated and augmented with chemical reasoning to post-train reaction-specific language models through supervised fine-tuning (SFT) and reinforcement learning (RL), producing a compact generalist model. In zero-shot evaluation across Suzuki–Miyaura coupling, Buchwald–Hartwig amination and acid–amine condensation, the generalist model outperforms the strongest baseline by 41.5% in mean accuracy and 29.2% in macro-F1, while surpassing human expert accuracy by 8.9% and macro-F1 by 11.5%. To accelerate active exploration, generalist predictions are transferred into warm priors for Bayesian neural networks. This warm prior reduces the experimental labels needed to reach a held-out macro-F1 ≥ 0.70 by 25.0–42.9% relative to a vanilla Bayesian neural network (BNN). Furthermore, uncertainty-guided acquisition prioritizes informative reactions, enabling the models to approach full-pool training performance using only about oneeighth of the total available reactions. To our knowledge, this work establishes the first experimentally validated literature-to-laboratory paradigm connecting 1 large-scale post-trained language models with active experimental design, providing a practical foundation for data-efficient, privacy-preserving, and autonomous chemical discovery.