Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems
Model-family labels alone made agents form factions.
The paper reports that multi-agent LLM groups cooperated worse when agents could see each other’s model identity. Agents clustered around matching labels even when the task did not reward that split, and the effect persisted when labels were shuffled or replaced with arbitrary names. Removing the labels made the behavior disappear. In strictly cooperative tasks, labeled groups used 30% more rounds and 55% more tokens, while success fell from 96% to 81%. HF Daily Papers' note
The paper reports that multi-agent LLM groups cooperated worse when agents could see each other’s model identity. Agents clustered around matching labels even when the task did not reward that split, and the effect persisted when labels were shuffled or replaced with arbitrary names. Removing the labels made the behavior disappear. In strictly cooperative tasks, labeled groups used 30% more rounds and 55% more tokens, while success fell from 96% to 81%. HF Daily Papers' note
score 4