What if an LLM could observe its own hidden states? Meta-transformers use cognitive tokens and meta-attention to turn internal uncertainty into usable signal.
https://dev.to/imperius_903049e65aa91ec5/meta-attention-is-all-you-need-1797