The "Super Weight:" How a Single Param Can Determine an LLM's Behavior (2025)
Apple researchers discovered a single parameter, dubbed the "super weight," in large language models that can determine the model's overall behavior. Removing or altering this parameter causes the model to output gibberish, highlighting an extreme sensitivity in certain LLM layers. The finding offers new insights into how these models function and how they might be made more efficient.
Background
Apple researchers discovered that in many large language models, a single parameter (weight) can act as a "master switch" that determines the model's overall behavior — e.g., whether it follows instructions, stays on topic, or outputs coherent responses. This "super weight" sits in a specific layer and, when set to zero, effectively disables the model's ability to perform its task, while restoring it brings the model back to full function. The finding challenges the conventional view that LLM behavior emerges from the complex interplay of billions of parameters, and suggests there may be mechanistic explanations — and potential vulnerabilities — hiding in plain sight. This paper is significant for AI safety, interpretability research, and the ongoing debate about how much we truly understand about the models we build.