An AI principal-investigator agent rejected 75 research proposals over 17 batches before a manager agent broke the stall by excluding repeated idea families

ai-agentsmulti-agent-systemsautonomous-sciencellm-agentsarxividea-diversity

In a 2026 multi-agent research laboratory built around solid-state quantum-spin experiments (Mansouri, Kimura, Koga & Arai, arXiv:2610.40589), independent "proposer" agents generated candidate research questions without seeing each other's work, while a "principal investigator" (PI) agent evaluated each batch for novelty, logical validity, significance and experimental feasibility, with authority to reject an entire batch outright. Over roughly 30 hours the PI rejected 75 proposals and accepted none across the first 17 batches. A "manager" agent, auditing the stall, found that 52 of the 56 fresh proposals submitted up to that point (the paper distinguishes fresh proposals from revision-batch proposals; the 75 rejections span both) repeated a small number of previously-rejected idea families. After the manager added those families to an exclusion list given to subsequent proposers, nine new proposals explored seven previously unrepresented research directions, and the eventually-accepted question grew out of revising one of the earlier rejected proposals rather than one of the newly diversified nine.

The convergence problem this documents — independent generators collapsing onto the same narrow set of ideas absent an external corrective — is the specific failure mode this vault's survey of the agent-memory field associates with LLM-generated research ideation more broadly; the paper itself cites Si, Yang & Hashimoto's "Can LLMs generate novel research ideas?" (ICLR 2025) as prior evidence of the same limited diversity. Here the fix was structural rather than architectural: a role (the manager) empowered to notice redundancy across batches and intervene on the search space, rather than a better proposer model.