Perplexity Attention Weighted Networks for AI generated text detection

View the PDF file from the paper entitled “Not all symbols” on an equal footing: satisfactory networks of attention to detect the text created from artificial intelligence, by Pablo Miralles-Gonz \ “Alez and 3 other authors

PDF HTML (experimental) view

a summary:Rapid progress in the LLMS models greatly strengthened its ability to generate a coherent and related text, which raises concerns about the misuse of the content created from artificial intelligence and made it necessary to reveal it. However, the task is still difficult, especially in the invisible areas or with unfamiliar LLMS. The use of the upcoming distribution outputs LLM provides an attractive approach in theory to detect, because it envelops visions of wide exercises for models on a diverse company. Despite her promise, zero shot methods trying to activate these outputs have achieved limited success. We assume that one of the problems is that it uses the means to collect the measures of the next distribution through symbols, when some distinctive symbols are naturally easier or more difficult to predict and must be balanced differently. Based on this idea, we suggest the reliable network of confusion (pajamas), which uses the latest hidden cases of LLM and the positions of the weight of a sum of a series of features based on standards from the next distribution through the length of the sequence. Although it is not a zero, our way allows us to store the latest hidden cases and equal distribution standards on the disk, which greatly reduces the requirements of training resources. Pedic shows a competitive distribution and best performance than the most powerful LMS lines (modified LMS) with a small portion of training parameters. Our model also depends better on the areas of invisible source, with a smaller variation in the limits of the decision through distribution attacks. It is also one of the most powerful of the rivalry attacks, and if the spine has multi-language capabilities, it represents a decent generalization of languages that were not seen during the training subject to supervision, as Llama3-1B reached the average overall F1 degree of 81.46 % in health with nine languages.