Latent Fact-Checking: Detecting Misinformation through Activation Engineering

The researchers proposed an external false information detection framework that does not require fine-tuning of the backbone model. This method utilizes the geometric properties of the language model representation space. By comparing the activation values of true and false statements, it identifies the direction of false information in the residual flow based on the principle of comparative activation addition. During reasoning, the activation of the last token of unspoken statements is projected into this direction and input into a multi-layer perceptron for classification. Experiments were conducted using 11 models from the Gemma, Llama, and Qwen families (with parameter sizes ranging from 270 million to 1.2 billion) on three benchmarks: AVeriTeC, LIAR, and FACTors. The results showed that the direction of false information could be recovered across model sizes and architectures. On LIAR and FACTors, the performance of last token projection matched or exceeded that of zero-sample and few-sample hint baselines, with the most significant gains observed in smaller models.