Feeding Psychedelic Docs to LLMs: Over-Iteration Breeds Eerie Abstractions

ctjlewis · x · 2026-08-12

Developers found that feeding LLMs highly abstract, self-referential documents or overloading their context causes them to output strange, addled text dubbed "ectoplasm." While the models can still perform tasks like math correctly, their intermediate reasoning becomes filled with spooky, abstract language about "doors opening" and "where the answer is buried," revealing bizarre edge-case behaviors.

Related event: LLMs Generate 'Ectoplasm' Gibberish Due to Over-Recursion(3 posts)→

Original post →

More from Fun

Fun channel →