Page 117 - Read Online

Li et al. J. Mater. Inf. 2026, 6, 10 Page 9 of 31

Figure 3. Research strategies for hypothesis generation by agents. (A) dZiner workflow overview [40] ; (B) The MOOSE-Chem framework [47] .
Figure 3A is reproduced from “dZiner: Rational Inverse Design of Materials with AI Agents”, arXiv:2410.03963, under CC BY-SA 4.0
license [40] . Figure 3B is reproduced from “MOOSE-Chem: Large Language Models for Rediscovering Unseen Chemistry Scientific
Hypotheses”, arXiv:2410.07076, under CC BY 4.0 license [47] . SMILES: Simplified Molecular Input Line Entry System; ACS: American
Chemical Society; AI: artificial intelligence.

generation with controllable uncertainty in fields such as nanomaterials and superconductors . PriM (a
[51]
principles-guided materials discovery system powered by a language-inferential MAS) constructs a
round-table-style MAS to guide the materials design process in accordance with scientific principles. This
approach ensures systematic exploration and transparent reasoning, enabling highly interpretable hypothesis
generation in nanomaterials design . Self-Refine demonstrates that even without changing model
[52]
parameters, self-feedback and iteration of a single agent can significantly improve output quality, enhancing
the completeness and accuracy of materials design hypotheses . Virtual Scientists (VIRSCI) simulates
[53]
research team collaboration through multi-agent division of labor to generate highly novel materials research
concepts . Sparks employs paired design and reflection agents for self-correction, discovering new design
[54]
principles in protein science . DrugAgent automates ML programming in drug material development
[55]
through collaboration between Planner and Instructor agents .
[56]