Latent Fact-Checking Detects Misinformation Using Activation Engineering

The rapid spread of misinformation online necessitates scalable and effective detection systems. Traditional approaches often rely on surface-level linguistic features or external knowledge bases. This research proposes a novel framework, Latent Fact-Checking, which views truthfulness as an inherent geometric property within a language model's internal representation space.The core of this approach lies in activation engineering. By contrasting the activations generated by paired truthful and false statements, the framework elicits a specific "misinformation direction" within the transformer model's residual stream. This process follows the principle of Contrastive Activation Addition (CAA), effectively identifying how truthfulness is encoded in the model's latent space.During inference, an unseen claim's last-token activation is projected onto this learned misinformation direction, and the resulting representation is fed into a simple Multilayer Perceptron (MLP) for classification. Crucially, this method requires no fine-tuning of the base model, no external evidence retrieval, and only minimal task-specific supervision (the contrastive pairs). Evaluated across various models (Gemma, Llama, Qwen) and benchmarks (AVeriTeC, LIAR, FACTors), the falsehood direction proved recoverable across different model scales and architectures. The method matched or surpassed zero-shot and few-shot prompting baselines on LIAR and FACTors, particularly benefiting smaller models, suggesting that truthfulness is a structured, linearly separable concept within pretrained language models.