Retrieval-Grounded Credit Assignment for Generative Recommendation

Generative recommenders using Semantic IDs (SIDs) suffer from sparse rewards and a credit-assignment gap between reasoning traces and final item selection. This method uses retrieval-grounded query attribution to attribute rewards to specific reasoning steps rather than just the final SID.