A new paper argues that for many real-world decision problems, you never needed the LLM in the first place.
Researchers studying contextual multi-armed bandits (a class of sequential decision problems used in recommendation engines, financial offer selection, and portfolio management) found that simple numerical models operating on text embeddings match or outperform LLM-based approaches at significantly lower cost. The team also introduced LLMP-UCB, a bandit algorithm that extracts uncertainty estimates from LLMs through repeated inference. More practically useful is their geometric diagnostic: an analysis of a problem's embedding structure that tells practitioners upfront whether LLM reasoning will add anything over a cheaper alternative. They also found that embedding dimensionality itself acts as a tuning knob for the exploration-exploitation tradeoff, sidestepping prompt engineering entirely.
The pattern here is familiar: LLMs are applied to problems because they can be, not because they should be. This paper gives practitioners a principled answer before they have spent the compute budget finding out the hard way. In finance and recommendation contexts, where inference runs at scale, the cost gap between an LLM and an embedding model is not trivial.
The researchers carve out cases where LLMs do earn their keep. But if the field's default has been "try the LLM first," a diagnostic that demands "prove you need it" is a welcome corrective.