Teaching and Evaluating LLMs to Reason About Polymer Design Related Tasks

Researchers introduce PolyBench, a benchmark dataset containing 125K+ polymer design tasks backed by 13M data points, along with a knowledge-augmented reasoning method to improve LLM performance in materials science. Small and mid-sized language models trained on PolyBench achieve competitive results with frontier models, demonstrating practical advancement in AI4Science applications.

The development of PolyBench addresses a critical gap in AI4Science by creating specialized training infrastructure for polymer design tasks. Current large language models lack domain-specific knowledge and capability coverage needed for materials science applications, limiting their utility in research and industrial settings. The benchmark's scale—leveraging both experimental and synthetic data sources—enables comprehensive coverage across polymer varieties and their properties, a prerequisite for robust model performance in specialized domains.

This work reflects a broader trend in machine learning where general-purpose models require targeted fine-tuning and domain-specific datasets to achieve competitive performance in specialized fields. The knowledge-augmented reasoning distillation method, which structures the dataset with chain-of-thought reasoning, represents an important methodology for translating complex domain knowledge into effective training signals. The progression from simple to complex tasks within PolyBench enables researchers to diagnose model capabilities and identify generalization challenges.

The results demonstrate that smaller models (7B-32B parameters) can match or exceed larger frontier models on specialized tasks when trained appropriately, suggesting future efficiency gains in scientific AI applications. For researchers and materials scientists, this work enables improved computational workflows for polymer discovery and optimization. The released dataset and code democratize access to specialized training resources, potentially accelerating development of domain-specific AI tools across materials science and chemistry.

Future attention should focus on whether PolyBench's methodologies transfer to other materials science domains and whether downstream scientific discoveries result from improved polymer design capabilities.

• →PolyBench benchmark enables small language models to match frontier LLM performance on polymer design tasks through specialized training data.
• →The dataset comprises 125K+ tasks derived from 13M experimental and synthetic data points, ensuring comprehensive polymer property coverage.
• →Knowledge-augmented reasoning distillation improves model alignment by structuring domain knowledge as chain-of-thought reasoning patterns.
• →Open-source release of dataset and code democratizes access to specialized training resources for materials science AI applications.
• →Small to mid-sized models demonstrate practical efficiency gains compared to large frontier models when optimized for domain-specific tasks.