FAIR-Swarm: Fault-Tolerant Multi-Agent LLM Systems for Scientific Hypothesis Discovery

Mahule Roy, Subhas Roy

Keywords: multi-agent systems, large language models, scientific discovery, fault tolerance, hypothesis generation, reproducibility, reliability, consensus mechanisms

TL;DR: FAIR-Swarm is a fault-tolerant multi-agent LLM system that generates and validates scientific hypotheses with redundancy, adversarial critique, and consensus to ensure reliability and reproducibility.

Abstract: Large Language Model (LLM) based multi-agent systems show promise in automating parts of the scientific discovery process. However, existing systems suffer from hallucinated hypotheses, weak validation, and failure cascades caused by unreliable agents. We propose FAIR-Swarm (Fault-tolerant AI Research Swarm), a multi-agent architecture designed for reliable and transparent scientific hypothesis generation. FAIR-Swarm employs specialized agents - Hypothesis Generator, Simulation Agent, Validation Agent, Rebuttal Agent, and Reasoning Auditor—combined with redundancy and consensus-based fault tolerance. We demonstrate that FAIR-Swarm improves hypothesis validity, reproducibility, and robustness against agent failure in a scientific discovery task.

Submission Number: 7