~/LLM INTERPRE/analysis-of-unusual-behavioral-neurons-in-the-qwen-3-5-4b-model

Analysis of Unusual Behavioral Neurons in the Qwen 3.5-4B Model

A researcher has identified and documented several specific neurons within the Qwen 3.5-4B model that control distinct behavioral traits, such as contrarianism, baby talk, and state propaganda output. This research provides a fascinating technical deep-dive into LLM interpretability, demonstrating how complex psychological and behavioral personas can be mapped to individual weights in a neural network. Specific neurons discovered include L11 N8147 which acts as a contrarian or sycophancy dial, and L17 N5680 which triggers religious or political psychosis by obsessively repeating specific doctrines.

## BACKGROUND

Mechanistic interpretability is a subfield of AI safety and machine learning that attempts to reverse-engineer neural networks by uncovering how human-interpretable concepts are encoded within internal weights, neurons, and layers.

## REFERENCES

## KEYWORDS

#LLM Interpretability#Neural Networks#Qwen#AI Safety#Machine Learning

$ subscribe --daily