Tag
This paper demonstrates that Transformers in LLMs exhibit linear superposition when combining inputs, supporting the Superposition Linearity Hypothesis, and introduces a guided decoding method to disentangle superposed outputs.