The LLM Undo Button: Saying 'Wait Never Mind' Prevents Exponential Divergence

andersonbcdefg · x · 2026-08-02

The tweet humorously highlights an interesting property of autoregressive Large Language Models (LLMs) during reasoning. The author points out that because a model can always generate a token like "wait, never mind" to correct its path, it can never be trapped outside the set of correct answers, cleverly avoiding exponential divergence in its outputs.

Original post →

More from Fun

Fun channel →