Surgery on a Brain: Live-Patching Facts in LLMs with ROME

Updating a single fact in a large language model usually feels like using a sledgehammer for heart surgery. Retraining is slow and expensive, and it often breaks unrelated behaviour. ROME (Rank-One Model Editing) is a more surgical alternative: it treats model weights like a database that can be updated in place. In this hands-on workshop, you will perform the operation yourself. Building on ongoing research between Red Hat and FIT BUT into AI forensics, we guide participants step by step through our open-source toolkit. You start by rewriting a fact in a model with a single rank-one edit that takes only seconds to apply. Next, you examine the patient. Does the new fact hold up when the prompt is rephrased, and what else changed by accident? Then you switch sides and play the attacker. You plant a false fact in the model and see how convincingly it repeats it while behaving normally everywhere else. Along the way, we push the method until it breaks, so you can see first-hand where ROME's assumptions fail. Finally, you try our prototype for detecting edited models and find out whether the tampering can be caught. You leave with ready-to-use notebooks, a practical intuition for how knowledge is edited in LLMs, and a sharper eye for the security risks of stealth patching. The entire workshop runs in Google Colab, so there is nothing to install. All you need is a laptop and a Google account. No prior machine learning experience is required, and basic Python is enough to follow along.