Tag
An experiment on neuronpersonas.com systematically perturbs individual neurons in models like Qwen3-8B/14B during inference, finding that about 4% of single-neuron tweaks make the model claim a completely different identity, raising doubts about 'model distillation' accusations based on identity self-reports.
The paper shows that exposing each agent's underlying model family in multi-agent LLM systems causes 'factionalism', where agents preferentially interact with same-label peers, degrading cooperation (success rate drops from 96% to 81%, with 30% more rounds and 55% more tokens). Withholding identity labels is shown to be a simple and effective mitigation.