vicuna7b-memit-multi-hop-drop

Summary

When researchers surgically rewrote a fact inside a small Vicuna-7B model using the MEMIT editing technique, the model's ability to reason through chains of connected information collapsed almost entirely, falling from 30.2 percent to 4.9 percent accuracy. This matters because it suggests that a model's stored knowledge and its reasoning circuitry are deeply entangled, so even a narrow, targeted edit can silently destroy the model's capacity to make multi-step inferences without anyone noticing until they test it.