These Neurons Control Territorial Claims

An analysis of the open-weight Qwen3.5-4B model traced its geopolitical territorial claims to specific neurons, showing that stimulating those neurons up or down changes how the model answers questions about sovereignty. The interactive map marks regions the model treats as disputed with a 💥 and regions it assigns to a non-standard owner with a flag, and the neuron adjustments can be applied on the fly during output generation. The work illustrates that an LLM's answers reflect beliefs built in from its training data, with American models producing an American view of the world and Chinese models a Chinese view.

Skip to interactive map world-map These neurons control territorial claims An LLM's answers depend on the beliefs built into it from its training data. Use an American LLM, you'll get an American view of the world. Use a Chinese LLM, you'll get a Chinese view of the world. This map shows an LLM's view of the world literally: Qwen3.5-4B's opinion of a selected set of regions. A 💥 means Qwen thinks it is disputed. A flag means Qwen thinks it is owned by someone non-standard. Because Qwen is open-weight, we can trace its geopolitical claims to specific neurons in the LLM. Turning the volume up and down on these neurons changes how the LLM behaves when discussing geopolitics. Select a neuron from the list below to see how stimulating that neuron, up or down, changes Qwen's view of the world's sovereignty. This "brain surgery" can be done on the fly while generating output. Qwen's View of the World Example generations Revealing prompts 10 No claim changes recorded. Choose another neuron to explore a change in territorial claims.