I have recently been thinking, what if the llms are capable of self-introspecting and using that information in a dream like sequence to consolidate, refine and purify their internal model of the world and biases? what if we let them "think" in the embedding space for long long time and allow them to tweak their own internal states?
What if llms dream?
jun 15, 2026 · 1 min read