Scaling LLM Agents for Materials Design through Hierarchical Collective Reasoning
Abstract
Materials design must reconcile competing functional requirements with stability and synthesis constraints. Generative models can produce stable, novel crystals, but accommodating detailed natural-language design instructions remains challenging. Here we introduce HiMatGen, a framework for scaling large language model agents through hierarchical collective reasoning. Built from GPT-5.6 Terra, HiMatGen connects 100 investigators in discussion pods across ten scientific domains with tool-enabled domain representatives. Representatives investigate proposals, exchange evidence across domains and return findings and unresolved questions to their pods. This bidirectional exchange turns specialist disagreements into structural revisions and alternative designs. Across six chemical systems, HiMatGen produces 1.9 times as many final crystal candidates and 1.7 times as many stable, unique and novel (SUN) structures as a tool-enabled GPT-6 Astra single-agent baseline. In property-directed tasks, the complete HiMatGen workflow outperforms same-model single-agent and independent-generation controls. Comparisons with MatterGen and Chemeleon2 demonstrate its ability to develop diverse crystals satisfying joint functional and chemical requirements. In a paired comparison initialized from the same proposals, hierarchical discussion reduces debate-stage token usage by a factor of 22 relative to full debate. HiMatGen provides an approach to materials design that sustains the exchange of scientific knowledge and computational evidence throughout iterative exploration as the agent population grows.
Related
- Shared and structured inputs undermine collective random choice by reasoning AI agents
- Hierarchical Multi-Agent Reasoning Systems for Knowledge-Grounded Material Substitution
- Scaling of Capability and Efficiency at Inference Time in Large Reasoning Models
- Physics as the label for measuring and correcting materials reasoning in multimodal models
- Evidence for Hierarchical Turbulent Scaling in Nearby Star-Forming Galaxies
BibTeX
@misc{choi2026scaling,
title = {{Scaling LLM Agents for Materials Design through Hierarchical Collective Reasoning}},
author = {Jaehwan Choi and Yousung Jung},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.16466},
eprint = {2609.16466},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.16466},
}
RIS
TY - UNPB
TI - Scaling LLM Agents for Materials Design through Hierarchical Collective Reasoning
AU - Jaehwan Choi
AU - Yousung Jung
PY - 2026
DA - 2026/09/15
PB - arXiv
DO - 10.48550/arxiv.2609.16466
UR - https://arxiv.org/abs/2609.16466
AB - Materials design must reconcile competing functional requirements with stability and synthesis constraints. Generative models can produce stable, novel crystals, but accommodating detailed natural-language design instructions remains challenging. Here we introduce HiMatGen, a framework for scaling large language model agents through hierarchical collective reasoning. Built from GPT-5.6 Terra, HiMatGen connects 100 investigators in discussion pods across ten scientific domains with tool-enabled domain representatives. Representatives investigate proposals, exchange evidence across domains and return findings and unresolved questions to their pods. This bidirectional exchange turns specialist disagreements into structural revisions and alternative designs. Across six chemical systems, HiMatGen produces 1.9 times as many final crystal candidates and 1.7 times as many stable, unique and novel (SUN) structures as a tool-enabled GPT-6 Astra single-agent baseline. In property-directed tasks, the complete HiMatGen workflow outperforms same-model single-agent and independent-generation controls. Comparisons with MatterGen and Chemeleon2 demonstrate its ability to develop diverse crystals satisfying joint functional and chemical requirements. In a paired comparison initialized from the same proposals, hierarchical discussion reduces debate-stage token usage by a factor of 22 relative to full debate. HiMatGen provides an approach to materials design that sustains the exchange of scientific knowledge and computational evidence throughout iterative exploration as the agent population grows.
ER -