Graph feedback plays a critical role in shaping consensus and clique formation among open-weight language-model populations, according to a study by Samer Saab Jr and Chaouki Abdallah, submitted on July 13, 2026. The researchers examined models with parameters ranging from 1.1B to 32B using a naming-game protocol to analyze local interactions.
Understanding Graph Feedback in Language Models
The interaction graph in multi-agent language-model systems is often overlooked. This study emphasizes the importance of understanding how graph feedback can influence both local interactions and overall model behavior. By measuring prompt-conditioned score-state distributions, the authors constructed state-similarity graphs that reveal deeper insights into the dynamics of these systems.
Through controlled interventions, the study found that retaining partner-label evidence is essential yet not sufficient for achieving consensus. The authors noted that homophilous threshold-similarity routing tends to eliminate cross-basin exposure, leading to increased fragmentation. In contrast, bridge-seeking routing can often mitigate fragmentation, especially when memory resources are available.
Behavioral Consensus in Multi-Model Grids
In a mixed four-model grid with three seeds, the researchers observed that threshold-similarity routing failed to reach behavioral or state consensus across 189 setting-seed runs. However, state-component and label-disagreement bridges successfully achieved behavioral consensus in 14 out of 18 retained-memory runs. This highlights the variability in how different routing strategies impact consensus formation.
