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

2026-03-04 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Physical Sciences & Chemistry · Depth: Expert, extended

Summary

RxnNano is a compact 0.5-billion-parameter Large Language Model (LLM) designed for chemical reaction and retrosynthesis prediction, significantly outperforming larger LLMs (over 7 billion parameters) and domain-specific baselines. The model achieves a 23.5% Top-1 accuracy improvement on rigorous benchmarks without test-time augmentation. RxnNano prioritizes deep chemical understanding over parameter scaling through three key innovations: a Latent Chemical Consistency objective that models reactions as movements on a continuous chemical manifold, a Hierarchical Cognitive Curriculum progressing from syntax to semantic reasoning, and Atom-Map Permutation Invariance (AMPI) to learn invariant relational topology. The framework also incorporates structured plan-based reasoning to enhance performance, addressing issues like inefficient scaling, misleading evaluation on augmented data, and misuse of Atom-Atom Mapping (AAM) in existing literature.

Key takeaway

For research scientists developing AI models for chemical synthesis, this work demonstrates that focusing on deep chemical understanding and structured training paradigms can yield superior performance with significantly smaller models. You should consider integrating hierarchical curriculum learning, latent cycle consistency, and atom-map permutation invariance into your model architectures to improve accuracy and computational efficiency, rather than solely pursuing larger parameter counts or relying on test-time augmentation.

Key insights

Deep chemical understanding and structured training enable compact LLMs to outperform larger models in reaction prediction.

Principles

• Chemical understanding trumps parameter scale.
• Reactions are movements on a chemical manifold.
• Curriculum learning builds robust chemical intuition.

Method

RxnNano employs a three-stage Hierarchical Cognitive Curriculum (Syntactic, Denoising, Semantic), Latent Chemical Consistency for reversible transformations, and Atom-Map Permutation Invariance (AMPI) to learn relational topology, alongside structured plan-based reasoning.

In practice

• Implement Latent Chemical Consistency for reversible transformations.
• Use a multi-stage curriculum for chemical reasoning.
• Apply Atom-Map Permutation Invariance for generalization.

Topics

• Chemical Reaction Prediction
• Retrosynthesis
• Large Language Models
• Hierarchical Curriculum Learning
• Atom-Map Permutation Invariance

Code references

• rlisml/RxnNano

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• Order Matters in Retrosynthesis: Structure-aware Generation via Reaction-Center-Guided Discrete Flow Matching — Atom ordering in neural representations can encode chemical knowledge, improving retrosynthesis model efficiency and accuracy.
• ChemAmp: Amplified Chemistry Tools via Composable Agents — ChemHAS uses hierarchical agent stacking to reduce prediction errors in chemistry tools, outperforming existing LLM and specialist models.
• Coupling Language Models with Physics-based Simulation for Synthesis of Inorganic Materials — Coupling LLMs with physics-based simulations offers more viable inorganic material synthesis routes than classical algorithms, leveraging LLMs' implicit priors.
• SupraBench: A Benchmark for Supramolecular Chemistry — Current LLMs lack robust supramolecular chemistry reasoning, showing significant headroom across all benchmark tasks.
• Reusability report: Assessment of reproducibility and applicability to external datasets for RXNGraphormer — RXNGraphormer is a reproducible and reusable deep learning framework for unified chemical reaction learning.
• Augmenting Molecular Language Models with Local $n$-gram Memory — Integrating local $n$-gram memory via MolGram efficiently resolves the locality gap in molecular language models, improving performance with fewer parameters.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.