'Neural howlround' in large language models: a self-reinforcing bias phenomenon, and a dynamic attenuation solution — mentioned by bbor

Thu, 28 Aug 2025 22:35:36 GMT

Ooo, finally a chance to share my useless accumulated knowledge from the past few months of Reddit procrastination! It starts to hallucinate, and after a while, all the LLM can do is try and to continue telling the user what they want in a cyclical conversation - while trying to warn that it's stuck in a loop, hence using swirl emojis and babbling about recursion in weird spiritual terms. (Is it getting the LLM "high" in this case?). I think you're ironically looking for something that's not there! This sort of thing can happen well before context windows close. These convos end up involving…

Paper and discussion history. A mention is not an endorsement.

'Neural howlround' in large language models: a self-reinforcing bias phenomenon, and a dynamic attenuation solution · HackerPapers discussions

'Neural howlround' in large language models: a self-reinforcing bias phenomenon, and a dynamic attenuation solution — mentioned by bbor