Controlling Large Language Model Behaviour with Model Steering

We are excited to welcome:
Simon Ward - Director of AI/ml from Tranquility.ai.

Controlling Large Language Model Behaviour with Model Steering

Large language models often exhibit behaviors that don’t align with specific use cases, ranging from overly cautious responses to unwanted stylistic patterns. While traditional fine-tuning approaches like Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) teach models what to do, model steering surgically adds or removes unwanted behaviors by intervening at specific locations within the model's internals.
This talk will discuss applications in law enforcement, content moderation, and healthcare, comparing model steering's efficiency and targeting advantages over traditional fine-tuning. The talk will also address the dual-use nature of these techniques, recent research on safeguarding against malicious model steering applications, and broader implications for AI security.

As always free food and drinks. Come at 6PM to socialize and 6:30pm is the presentation.

Sponsored by LBMC