Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs
Transformers superpose next-token distributions from mixed text streams, and fine-tuning can restore that linearity.
The paper argues that Transformers exhibit linear superposition: when inputs from distinct text streams are linearly combined, the model outputs a superposition of the individual next-token distributions. The authors treat this as an architectural property rather than a consequence of training, and report that it tends to diminish as pretraining progresses. Lightweight fine-tuning substantially restores linearity, reducing divergence between the predicted distribution and the average of the separate distributions. A guided decoding procedure then disentangles the superposed output so two coherent continuations can be generated from one forward pass.
- Linearly mixed text streams produce a superposition of individual next-token distributions.
- Superposition appears intrinsic to Transformers and weakens as pretraining progresses.
- Lightweight fine-tuning substantially restores linearity versus averaged token distributions.
- Guided decoding generates two coherent continuations from a single forward pass.
Full article131 words · extracted from huggingface.co · click to collapse
While Large Language Models (LLMs) rely on highly non-linear components, in this work we demonstrate that they exhibit fundamental linearity: when inputs from distinct text streams are linearly combined, the model outputs a superposition of the individual next-token distributions. We term this the Superposition Linearity Hypothesis. We provide evidence that superposition is an intrinsic property of the Transformer architecture rather than an emergent consequence of training; in fact, we observe that it tends to diminish as pretraining progresses. However, we demonstrate that linearity can be substantially restored through lightweight fine-tuning, significantly reducing the divergence between the predicted next-token distribution and the average of the individual next-token distributions. Finally, we introduce a guided decoding procedure that disentangles superposed outputs, enabling the simultaneous generation of two coherent continuations from a single forward pass.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.29845