Peer-Voted Feeds Induce Lexical Convergence in LLM Agents Across Four Model Families Without Consistent Distributed-Source Stance Effects

The study isolates feed-induced lexical convergence from opinion capture in synthetic populations. By freezing models and using block-bootstrap randomization, it shows peer likes on prior posts drive measurable homogenization even when topics are held constant. This extends earlier multi-agent work on emergent coordination, such as Park et al. (2023) on generative agents, by adding explicit voting mechanics that mirror production recommender systems.

Analysis reveals the design correctly bundles exposure with ranking, preventing claims about ranking alone, yet the failure to detect reliable distributed-source advantages challenges assumptions in AI safety literature that source diversity inherently resists capture. Related observational studies of production LLMs, including those tracking output drift in fine-tuned chat models, suggest similar convergence may occur when synthetic data loops back into training.

The core limitation remains the absence of human-subject validation; synthetic agents lack the anchoring effects of real-world priors documented in cognitive psychology. Scaling the protocol to frontier closed models and testing interventions such as diversity penalties could clarify whether convergence persists or can be mitigated before deployment.

Next steps include replication on newer open-weight releases and integration with human-AI hybrid platforms to measure whether lexical narrowing alters downstream user beliefs.