Truthfulness is linearly decodable per attention head, and ITI steers it

Li et al. train a separate linear probe per attention head on TruthfulQA in LLaMA-7B, finding truthfulness strongly decodable in only a sparse subset of heads (best head 83.3%) rather than one whole-residual-stream direction [li-etal-2023] Even within a head the geometry is not perfectly one-dimensional, as a second probe constrained orthogonal to the first still classifies above chance [li-etal-2023] Inference-Time Intervention shifts activations along the winning mass-mean direction (outperforming a probe-weight direction and contrast-consistent search) on the top-K most truthful heads [li-etal-2023] This raises TruthfulQA true*informative substantially (LLaMA-7B 30.5% to 43.5%, Alpaca-7B 32.5% to 65.1%, Vicuna-7B 51.5% to 74.0%) [li-etal-2023]