RxnNano:Training Compact LLMs for Chemical Reaction and Retrosynthesis Prediction via Hierarchical Curriculum Learning

A research team led by Ran Li has introduced RxnNano, a novel framework for training compact large language models (LLMs) specifically for chemical reaction and retrosynthesis prediction. The work, detailed in an arXiv preprint, directly challenges the prevailing trend in AI research that prioritizes scaling model parameters and datasets. The authors argue that current approaches often bypass fundamental challenges in chemical representation and fail to capture deep chemical intuition, such as reaction common sense and topological atom mapping logic. Their solution is a unified framework designed to instill this essential knowledge into models through innovative training methodologies, prioritizing chemical understanding over brute-force scaling.

The core of RxnNano's success lies in three key technical innovations: a Latent Chemical Consistency objective that models reactions as movements on a continuous chemical manifold for physically plausible transformations; a Hierarchical Cognitive Curriculum that trains the model through progressive stages from syntax to semantic reasoning; and Atom-Map Permutation Invariance (AMPI) to enforce learning of invariant relational topologies. The result is a remarkably efficient 0.5-billion-parameter model that significantly outperforms fine-tuned LLMs more than ten times its size (over 7B parameters) and all existing domain-specific baselines. It achieves a 23.5% improvement in Top-1 accuracy on rigorous benchmarks without relying on test-time augmentation, demonstrating that superior performance in complex scientific domains can be achieved through smarter architecture and training, not just larger models.