Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation

The finding isn't just that diversity drops — it's that the mechanisms typically used to improve agent quality (alignment, hierarchy, scale) are themselves the cause of the collapse, meaning the problem is baked into standard design choices rather than being a fixable edge case.

This connects directly to the CoopEval paper from April 16, which found that LLM agents in social dilemmas consistently defect rather than cooperate. Both papers are pointing at the same underlying issue from different angles: individual agent optimization produces collectively dysfunctional behavior. CoopEval showed agents fail to sustain cooperative equilibria even with game-theoretic nudges; this paper shows that even when agents aren't in explicit competition, structural coupling drives them toward conformity. Together they suggest that multi-agent LLM design has a systemic coordination problem that neither scaling nor alignment tuning resolves on its own. The molecular creativity paper from April 20 adds a related data point: LLMs already show constrained creative range in single-agent settings, which likely compounds when group dynamics suppress divergence further.

Watch whether any multi-agent framework ships an explicit diversity-preservation mechanism (such as enforced output dissimilarity or role-based epistemic separation) and tests it against this paper's benchmarks within the next two quarters. If none do, that suggests the field is treating this as a theoretical concern rather than a deployment problem.