# Investigating Bias Representations in Llama 2 Chat via Activation Steering

arXiv: https://arxiv.org/abs/2402.00402 · HTML: https://mlsanity.com/paper/2402.00402/

## Key facts

- Cited by 14 papers in the map (data through 14 Oct 2026).
- Placed in Human values and AI feedback alignment, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 2 of its 6 references that are in the map.
- First posted on arXiv 1 Feb 2024.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Human values and AI feedback alignment](https://mlsanity.com/alignment/human-values-and-ai-feedback-alignment/)
- Mechanistic interpretability / [Probing deception, bias and beliefs](https://mlsanity.com/mechinterp/probing-deception-bias-and-beliefs/)

## Builds on (most cited first, 2 of 2)

- [StereoSet: Measuring stereotypical bias in pretrained language models](https://mlsanity.com/paper/2004.09456/) (2020)
- [The Woman Worked as a Babysitter: On Biases in Language Generation](https://mlsanity.com/paper/1909.01326/) (2019)

## Built on by (most cited first, 14 of 14)

- [Programming Refusal with Conditional Activation Steering](https://mlsanity.com/paper/2409.05907/) (2024)
- [Eight Methods to Evaluate Robust Unlearning in LLMs](https://mlsanity.com/paper/2402.16835/) (2024)
- [Improving Instruction-Following in Language Models through Activation Steering](https://mlsanity.com/paper/2410.12877/) (2024)
- [Defending Against Unforeseen Failure Modes with Latent Adversarial Training](https://mlsanity.com/paper/2403.05030/) (2024)
- [Steering Without Side Effects: Improving Post-Deployment Control of Language Models](https://mlsanity.com/paper/2406.15518/) (2024)
- [A Survey of Test-Time Compute: From Intuitive Inference to Deliberate Reasoning](https://mlsanity.com/paper/2501.02497/) (2025)
- [Model Tampering Attacks Enable More Rigorous Evaluations of LLM Capabilities](https://mlsanity.com/paper/2502.05209/) (2025)
- [Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models](https://mlsanity.com/paper/2601.14004/) (2026)
- [Representation Engineering for Large-Language Models: Survey and Research Challenges](https://mlsanity.com/paper/2502.17601/) (2025)
- [Improving Multilingual Language Models by Aligning Representations through Steering](https://mlsanity.com/paper/2505.12584/) (2025)
- [Bridging the Knowledge-Prediction Gap in LLMs on Multiple-Choice Questions](https://mlsanity.com/paper/2509.23782/) (2025)
- [When is Your LLM Steerable?](https://mlsanity.com/paper/2606.11599/) (2026)
- [Do LLMs have a Gender (Entropy) Bias?](https://mlsanity.com/paper/2505.20343/) (2025)
- [Latent space bias directions in LLMs capture confidence, not fairness](https://mlsanity.com/paper/2610.08559/) (2026)

## Abstract (arXiv)

We address the challenge of societal bias in Large Language Models (LLMs), focusing on the Llama 2 7B Chat model. As LLMs are increasingly integrated into decision-making processes with substantial societal impact, it becomes imperative to ensure these models do not reinforce existing biases. Our approach employs activation steering to probe for and mitigate biases related to gender, race, and religion. This method manipulates model activations to direct responses towards or away from biased outputs, utilizing steering vectors derived from the StereoSet dataset and custom GPT4 generated gender bias prompts. Our findings reveal inherent gender bias in Llama 2 7B Chat, persisting even after Reinforcement Learning from Human Feedback (RLHF). We also observe a predictable negative correlation between bias and the model's tendency to refuse responses. Significantly, our study uncovers that RLHF tends to increase the similarity in the model's representation of different forms of societal biases, which raises questions about the model's nuanced understanding of different forms of bias. This work also provides valuable insights into effective red-teaming strategies for LLMs using activation steering, particularly emphasizing the importance of integrating a refusal vector.