generative steganography examples
Three matched pairs, each from a different open-weight model. On the left, the model's unmodified output. On the right, the output of the same model on the same prompt while BAM steers the sampling, carrying a payload at no cost to the output distribution. Select any right-hand artefact to inspect the receiver's belief propagation for that round.
For a detailed explanation of BAM (Burnashev Adaptive Posterior Matching), see the paper.