Original title: Scaling LLM Agents for Materials Design through Hierarchical Collective Reasoning
Jaehwan Choi, Yousung Jung · September 2026
Designing new materials is tricky. Scientists must balance useful working properties against whether a crystal is stable and possible to make. Computer models can create stable crystals, but following detailed design requests written in plain language is hard. To fix this, researchers created HiMatGen. It uses large language models, the AI behind advanced chatbots, acting as a large team. HiMatGen groups 100 virtual investigators across ten scientific fields into small discussion podsSmall groups of AI agents that focus on a specific scientific domain and share ideas to refine a design., linking them with tool-using representatives. These representatives check ideas, share evidence across fields, and report back. Pods turn disagreements into better crystal designs. Across six chemical systems, HiMatGen found 1.9 times as many final crystal candidates and 1.7 times as many stable, unique and novel structures as a single advanced agent baseline.
Imagine a team of architects designing a new building. Each expert handles a different part—structure, plumbing, electricity—but they rarely talk directly, so conflicts surface late. Here, the experts form small groups, and specialists from each area investigate issues and bring back solutions, turning disagreements into design improvements. This way, the building is completed faster with fewer costly changes.
Designing new materials requires balancing stability, synthesis, and performance. This framework lets many AI specialists collaborate efficiently, considering evidence from different domains. It produces more viable crystal candidates while using far less computation than prior methods.
The abstract doesn't explicitly state limitations, but it evaluates on only six chemical systems and specific property-directed tasks. It uses GPT-5.6 Terra, which may not be widely available, and does not address real-world synthesis success. Comparisons are limited to a few baselines.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Scaling LLM Agents for Materials Design through Hierarchical Collective Reasoning — Jaehwan Choi, Yousung Jung
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.