New work on LLM Control Theory! We frame LLMs as discrete dynamical systems and study prompting through the lens of reachability and controllability. We even proved some theorems!! Joint work my fellow EngSci 2T2 @witkowski_cam, Dr. @Sz_Looi, advised by Prof. Matt Thomson 🚀
Watch the full seminar here: youtu.be/9QtS9sVBFM0 Thanks to my teachers incl. Prof. @StarkDraper (Matrix Algebra), Prof. Lacra Pavel and Prof. Margaret Chapman (Control Theory), and Prof. Michelle Effros (Information theory)!! 🙏 for the support so far!
@ABhargava2000 @witkowski_cam @Sz_Looi The theorem part is great indeed @ABhargava2000
@ABhargava2000 @witkowski_cam @Sz_Looi Happy to see Transformers being analyzed from a systems viewpoint.
@ABhargava2000 @witkowski_cam @Sz_Looi Will you be at ICLR this year?
@ABhargava2000 @witkowski_cam @Sz_Looi i got a naive question. in the paper the prompt examples are cloze prompts similar for mlm, similar to the autoprompt paper. how does that adjust to the autoregressive world?