You ever wanted to change a piece of text (positive to negative movie review, sentence abt cats to dogs, etc) without changing the structure of the sentence too much? our project, TimpaTeks🗿 lets you do that 😭😱🤯
TimpaTeks🗿 is based on diffusion LMs and activation steering. Basically, to change the sentence automatically, it selects (samples) tokens that needed to be replaced, then it goes through the usual DLM generation pipeline with steer vectors injected.
I think the results are pretty neat and the text is still coherent (gif is an actual run result) pls check it out guys
github:
github.com/rayendito/dlm_ste…
arxiv:
arxiv.org/abs/2606.08408