Improving LLM Answers: The First Idea, Inner Speech

KordingLab · x · 2026-08-26

The tweet introduces the first idea for improving LLM answers: 'inner speech'. This involves having the model generate words that serve as a scratchpad or temporary storage, invisible to the user. This internal reasoning process helps assist in generating the final, high-quality answer.

Related event: From Stochastic Parrots to Modern LLMs: Implicit Language as Scratchpad(2 posts)→

Original post →

More from Research

Research channel →