Credit-assigned Policy Gradient for Early Stage Retrieval in Two-stage Ranking

Collection of recent IR, recommender, and retrieval-augmented LLM papers: Credit‑assigned Policy Gradient (CA‑PG) for early-stage rankers reduces variance by marginalizing over candidate‑set compositions; a proposed evaluation framework for structured generative search summaries; Uniboost, a posterior value‑alignment and linear‑boosting traffic allocation framework for blending/re-ranking; empirical study showing positional bias in dense retrievers is largely driven by the positional distribution of evidence in training data and that position‑balanced training mitigates bias; L2Rec, which uses