I'm reminded of the Physics of Language Models[1] where they showed a standatd autoregressive LLM got a lot more accurate if the models got access to the backsp

I'm reminded of the Physics of Language Models[1] where they showed a standatd autoregressive LLM got a lot more accurate if the models got access to the backspace key, so to speak. [1]: https://physics.allen-zhu.com/home https://physics.allen-zhu.com/home