A familiar argument holds that language models, as flexible approximators trained under a strictly proper loss, approach Solomonoff induction as they scale — and that their predictive success is evidence of this. A new arXiv paper takes this reasoning as its starting point and asks what language models can actually inherit from Solomonoff induction.

The paper, according to its abstract, explores both the limits of this inheritance and the properties that remain attainable. It introduces the idea of "served conditionals" as a tool for examining what language models reveal about their underlying inductive biases, rather than relying solely on aggregate predictive performance.

Because the abstract is truncated, the paper's full conclusions are not available in the source. But the framing suggests a more nuanced picture than the simple scaling argument, one where the relationship between language models and Solomonoff induction is partial and conditional rather than straightforward.