Researchers have demonstrated that the behavior and self-identification of the Qwen3.5-4B language model can be modified by applying a constant activation boost to specific MLP neurons. By simultaneously stimulating three distinct neurons, the model was successfully prompted to identify itself as a coffee bean, rather than its default AI assistant persona.
The experiment utilized the neuron-explorer tool to identify neurons associated with specific concepts. The researchers combined a 'form' neuron, which defines the model's self-perception as an object, with 'theme' neurons related to coffee and seeds. This modular approach suggests that the model's identity is constructed from multiple independent vector components.
The study highlights that these modifications require activation doses far exceeding the neurons' natural range, which often leads to a degradation in text coherence and increased repetition. Furthermore, the results show poor reproducibility across different prompts and evaluation batches, indicating that these 'persona' neurons do not store fixed identities but rather respond to thematic contexts.
Similar techniques have been explored in other studies, such as the identification of neurons that bypass safety alignments or trigger specific personality traits like melancholy or suspicion. While these findings demonstrate that individual MLP neurons can causally influence complex model behavior, they also suggest that such steering is reversible and highly dependent on the specific implementation of the inference process.
