AI Text Bypasser

Evasion Engine fine-tunes Large Language Models (LLMs) using Proximal Policy Optimization (PPO) to generate text highly resistant to AI content detection systems.

The methodology centers on a Reinforcement Learning from Human Feedback (RLHF) approach using a DistilBERT-based AI content detector to generate a probability of AI-likeness ($p_{AI}$), which is converted into an exponential reward signal.

The project leverages PPO's stability and sample efficiency to generate human-like text while bypassing tools like GLTR and GPTZero.