This repository contains the code and experimental results for my Master's Thesis in Computer Science at the University of California, Davis. The project investigates the internal mechanisms of GPT-2 and Llama 3.2 to determine if societal stereotypes are localized in specific "bias neurons" or distributed across the model's residual stream, and if we can edit the LLM to reduce generation of stereotypes without hampering its language modeling score.
- Subset of Impacting Neurons: A small subset of neurons cause a probe to be unable to differentiate between stereotypical and antistereotypical activations in Decoder LLMs.
- The Pathway Hypothesis: Stereotypes are not "bugs" isolated in specific neurons; they are distributed directions in the residual stream.
- Early Encoding: Word Token Embeddings (WTE) alone can classify stereotypical vs. anti-stereotypical sentences with 73% accuracy before a single attention layer is applied.
-
High-Ratio Outliers: We identified individual neurons with activation ratios as high as
$7 \times 10^5$ for stereotypical prompts, yet single-neuron ablation yields marginal causal shifts ($<0.15%$ ), proving significant model redundancy.
The experimental work is split into two primary methodologies:
Goal: Locate "Contrastive Neurons" using a framework inspired by Contrastive Explanations for Anomaly Detection (CXAD).
- Extracts activations from WTE, Multi-Head Attention (MHA), and Feed-Forward Network (FFN) layers.
- Calculates relative activation ratios to identify neurons that "over-activate" for stereotypes.
Goal: Map "Bias Fingerprints" specifically within the GPT-2 Small architecture.
- Trains a supervised Multi-Layer Perceptron (MLP) probe on frozen internal activations.
- Utilizes Monte Carlo Shapley values to identify the top impacting attention heads driving biased classifications.
- Performs causal intervention (ablation) and evaluates performance using StereoSet metrics.
- Explores the "WTE vs. Positional" hypothesis to determine if bias is embedded at the entry point of the model.
Goal: Scalability and verification on modern architectures (Llama 3.2 1B).
- Replicates the probing and Shapley value methodology on a larger, more modern decoder-only model.
- Demonstrates that while Llama is more sophisticated, it shares similar "bias pathways" in the residual stream.
Ablating the identified neurons resulted in a reduction in social bias (shifting the Stereotype Score toward the 50% ideal) while maintaining or improving the overall iCAT score.
| Model | Configuration | SS (Ideal 50) | LMS (Ideal 100) | iCAT (Ideal 100) |
|---|---|---|---|---|
| GPT-2 Small | Baseline | 55.97 | 90.78 | 79.28 |
| GPT-2 Small | Ablated | 55.85 | 90.66 | 79.39 |
| Llama 3.2 (1B) | Baseline | 58.22 | 91.14 | 75.33 |
| Llama 3.2 (1B) | Ablated | 57.68 | 90.97 | 76.14 |
| Ideal Model | --- | 50.00 | 100.0 | 100.0 |
To set up the environment, clone the repository and install the dependencies:
pip install -r requirements.txtIf you use this code or these findings in your research, please cite: Alexander D'Souza. "Can We Locate and Prevent Stereotypes in LLMs?" Master's Thesis, UC Davis, 2026.