Signal State Encodings Select Synchronization Outcomes in Language-Model Agents

Summary

This arXiv preprint reports that language-model agents produce different collective synchronization behaviors depending solely on how the same underlying physical state is encoded and presented to them. Using pretrained language models in a circular-synchronization experiment, the authors held the physical system fixed while varying the state-encoding intervention: agents saw only a summary of neighbours' relative phases, expressed either as low-order circular moments or as a histogram, and chose to advance, stay or retard accordingly. Encoding the state as circular moments rather than a histogram selected for different collective outcomes despite the underlying physical state being identical. The finding challenges the common assumption that state encodings are interchangeable interfaces for language-model agents. It was posted to arXiv's cs.CY category on 2026-08-07.