Evaluate classifier-free guidance for decision-conditioned generative retrieval

Evaluate the classifier-free guidance extension for Decision-Conditioned Generative Retrieval, which scores semantic identifier codes both with and without the decision prefix to control how closely generated items follow the selected decision.

Background

Decision-Conditioned Generative Retrieval (DCGR) first predicts a structured decision prefix describing the anticipated interaction and then generates semantic item identifiers conditioned on that prefix. The business offset $\beta$ steers which decision prefix is preferred, but it does not directly control how strongly the generated item follows that decision.

The paper proposes applying classifier-free guidance to the item-generation stage. During training, the decision prefix would occasionally be replaced by a null token; at inference, semantic-identifier scores would be computed with and without the decision prefix, and their difference would be scaled by a guidance weight $w$. This would provide a second control parameter in addition to $\beta$: $\beta$ would determine which decision is made, while $w$ would determine how closely the retrieved items conform to it. The extension has not been deployed or evaluated.