One model pass can output two separate answers

A transformer can produce two separate, coherent generations from a single forward pass, a new arXiv paper claims. The authors call it the Superposition Linearity Hypothesis. Linearly combine the inputs from two text streams, and the model's output approximates the average of what it would have predicted for each stream on its own, according to the paper.

That matters because inference cost scales with forward passes, not with how much useful output comes out of each one. If a model already carries two threads of computation inside one pass, standard sampling has been leaving that on the table.

Averaging alone wasn't the end of it. The authors built a guided decoding procedure that pulls two coherent, separate continuations back out of the blended pass, instead of a muddled average. They also found the property is architectural: it shows up out of the box, weakens as pretraining continues, and a small amount of fine-tuning brings it back.

The open question is scale. Every result here comes from whatever model sizes the authors tested, and the paper doesn't say how the effect behaves as parameter counts climb into the ranges labs actually deploy.

Anyone running high-volume inference should watch what happens if the guided decoding procedure survives contact with a frontier-scale model. That would be a mechanism for cutting the pass count on any workload that already generates multiple completions per prompt: best-of-n sampling, speculative drafts, parallel tool calls. Nobody has published that scaling result yet.