LLM Jailbreak DefenseMechanistic interpretability research

With QMIND, I worked on defending language models against jailbreak prompts by looking inside the model rather than filtering its inputs. Using sparse autoencoders trained with SAELens, we identified features in the model's activations that correlate with refusal behaviour, then boosted or suppressed those features at inference time.

The resulting framework raised refusal rates on jailbreak prompts by roughly 40% while keeping performance on benchmark tasks intact. We built a Gradio demo for steering individual features and watching the output change, and co-authored a paper published at the Canadian Undergraduate Conference on Artificial Intelligence.