Surgery on a Brain: Live-Patching Facts in LLMs with ROME
summary
Rome edits a specific factual association by computing a Rank-One Model Editing update: locate a single transformer layer (often around layer 13), compute a minimal perturbation and apply a single weight matrix update so the model now answers the edited tuple (e.g., Eiffel Tower → Rome); locating the layer takes about an hour but the edit itself is cheap and can run on a laptop and is almost data‑free for 70B models. Detection uses SVD metrics (spectral gap and top-one energy) implemented in an open‑source CLI and achieves roughly 80% average detection (ranges 100% to <50% across architectures).
AI-written summary · about
chapters (6)
- 0:00Live-patching LLMs
- 4:31LLM patching and knowledge editing
- 11:13Security risks of editing
- 16:51Detecting ROME edits
- 26:02Research group and future work
- 35:16Q&A
related talks
32mFine-tune LLMs from Kaggle Models using (Q)LoRA
29mLocal LLM fine-tuning: a practical example
35mReading the Mind of an LLM
20mHow to Survive the Vortex of LLM Change
32mEverything you wanted to know about VLMs but were afraid to ask
30mDo transformers need all that attention?
39mThe Future is Fine-Tuned: Deploying Task-specific LLMs
29mThe Evolution of LLMs: From Imitation to Reasoning