The current strong performance of Reality Defender’s text detector stems from the design choices in both the model architecture and the training process.
Our model can process the user input as a whole and iteratively refine its representation to extract subtle statistical patterns within the input text. This model design choice allows our detector to make a more accurate prediction efficiently, compared to the competing approaches that process input sequentially.
Also, our model is trained on a comprehensive corpus of texts generated by various LLMs and human authors, covering a wide range of genres. This allows our model to make robust predictions.
To continuously improve our model, we evaluate our model using a fine-grained internal text corpus. This allows us to identify any shortcomings and improve upon them.