Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs
The paper claims mixed text streams can make a transformer output a blended next-token distribution.
The authors call this the Superposition Linearity Hypothesis. They argue the effect comes from the Transformer architecture itself, not simply from training, and say it weakens as pretraining progresses. Lightweight fine-tuning can restore more of the linear behavior. They also describe guided decoding that separates the combined output into two coherent continuations from one forward pass. HF Daily Papers' note
The authors call this the Superposition Linearity Hypothesis. They argue the effect comes from the Transformer architecture itself, not simply from training, and say it weakens as pretraining progresses. Lightweight fine-tuning can restore more of the linear behavior. They also describe guided decoding that separates the combined output into two coherent continuations from one forward pass. HF Daily Papers' note
score 5