INDIBATOR: Multi-Agent Debate for Molecular Discovery

Overview

We introduce a framework enabling multi-agent systems to conduct scientific discovery by grounding agents in personalized scientist profiles. These profiles are built from two sources: academic publication histories providing domain knowledge and molecular data providing structural understanding. Agents participate in iterative debate cycles involving proposals, critiques, and voting. Our study demonstrates that agents with fine-grained individualized characteristics consistently surpass those using simplified role-based or keyword-based personas, achieving competitive results. The findings highlight the importance of capturing the "scientific DNA" of individual agents for enhanced discovery outcomes.

Key Contributions

Fine-grained Individuality: Grounding agents in real-world academic metadata and structural molecular data to capture the "scientific DNA" of individual agents.
Diverse Expert Pool: Leveraging a wide array of specialized knowledge to cover the vast chemical search space more effectively than a single agent.
Fact-Grounded Personas: Moving beyond simple prompts by grounding agents in real-world academic metadata and structural molecular data.

Observations

Granularity, Diversity, and Fact-Grounding matters in scientific agents

The comprehensive analyses demonstrate that the performance of the multi-agent system is a direct function of the profile quality. We observe:

Granularity: Detailed scientific profiles outperform coarse-grained roles (e.g., "Chemist"), suggesting that LLMs can better leverage specific domain knowledge when provided with high-resolution context.
Diversity: A heterogeneous pool of agents prevents groupthinking and enables the exploration of diverse chemical spaces, which is critical for identifying novel candidates that single-persona systems often overlook.
Fact-Grounding: Anchoring agents in actual publication histories minimizes hallucinations and ensures that the reasoning remains consistent with established scientific reality.

Granularity

Diversity

Fact-Grounding

Expertise-Based Reasoning

Qualitative analysis of the agent debate logs confirms that agents successfully adopt their assigned scientific personas. Unlike standard LLMs, our research trajectory-based agents provide critiques and proposals that reflect their unique research backgrounds—ranging from synthetic feasibility to pharmacokinetics—leading to a more rigorous and multidisciplinary peer-review process within the discovery loop.

Reasoning of each scientist agent is grounded in their research trajectory.

BibTeX

@misc{jang2026indibatordiversefactgroundedindividuality,
  title={INDIBATOR: Diverse and Fact-Grounded Individuality for Multi-Agent Debate in Molecular Discovery}, 
  author={Yunhui Jang and Seonghyun Park and Jaehyung Kim and Sungsoo Ahn},
  year={2026},
  eprint={2602.01815},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2602.01815}
}