As the number of LLMs grows, choosing the right model for each prompt is increasingly important. Many routing systems embed prompts into a vector space and then pick a model based on that representation. A new preprint on arXiv argues that the geometry of those embeddings plays a decisive role in whether routing succeeds.
The authors trace failure modes to the shape and distribution of the embedding space, showing that certain geometric configurations can systematically mislead a router. They then propose changes to the embedding geometry that improve routing accuracy and help the router generalize to prompts it has not seen before.
The abstract is brief, so the full methodology and experimental details are not yet available. But the central claim is clear: routing performance depends not just on the classifier applied to embeddings, but on the underlying geometry of the embedding space itself.