What’s the deal with Hallucinations anyway?

[ Abstract ]

When an AI-generated result looks convincing at first glance but falls apart under closer inspection, it can signal a collision in the model’s latent space, where concepts, patterns or representations are overlapping in unexpected ways.

This talk explores what those failures can tell us about model behaviour, and the different ways we can address them: through fine-tuning, post-training and other data-boosting approaches, or by changing how we prompt models to steer clear of problematic edges altogether.

We’ll also look at prompt optimisation techniques that can help identify and avoid these failure regions more efficiently, producing outputs that are not just plausible, but actually hold up under scrutiny.