Large language models are increasingly used for decision-making, but their standard inference procedure is expensive. They generate decisions through parallel prefill followed by sequential decoding, which incurs significant computational cost and does not directly reveal the model's confidence in a decision.

The paper introduces Large Decision Models, a method that repurposes LLMs as single-pass decision makers. Instead of decoding token by token, the model produces a decision in one forward pass, eliminating the need for sequential generation.

This design also makes decision probabilities explicit, which is useful for applications that require calibrated confidence estimates. The abstract mentions existing task-specific models, but the full comparison is not included in the excerpt; the approach is positioned as a general alternative to those methods.