A prompt that asks whether a statement is true raises the variance share and cross-topic cosine of the diff-in-means truthfulness direction in Llama-2-7B-Chat

In Llama-2-7B-Chat last-token final-layer embeddings on the six True-False topic subsets, wrapping each statement in "Is the above statement correct?" lets the first two principal components separate true from false in all six subsets [zhang-etal-2024-prompt-guided-internal-states-for-hallucination-detection] The share of total variance lying along the diff-in-means "truthfulness direction" rises on every subset when the prompt is added; for 10 GPT-4o-paraphrased templates the average share is 0.0875-0.1513 versus 0.0725 with no prompt (Table 2) [zhang-etal-2024-prompt-guided-internal-states-for-hallucination-detection] Pairwise cosine between the truthfulness directions of different topics is 0.2520-0.6950 without the prompt and 0.7253-0.8540 with it (Table 1); a second template gives 0.7835-0.8811 (Table 11) [zhang-etal-2024-prompt-guided-internal-states-for-hallucination-detection] Only this single layer/token position and one chat model (Llama-2-7B-Chat) are used for the geometry; the paper then feeds these prompt-guided states to hallucination detectors, whose cross-topic accuracy for the mass-mean probe rises from 0.5572 to 0.7524 [zhang-etal-2024-prompt-guided-internal-states-for-hallucination-detection]