Generative Pseudo-Labeling for Pre-Ranking with LLMs

TL;DR

Generative Pseudo-Labeling (GPL) is proposed, a framework that leverages large language models (LLMs) to generate unbiased, content-aware pseudo-labels for unexposed items, explicitly aligning the training distribution with the online serving space.

Abstract

Pre-ranking is a critical stage in industrial recommendation systems, tasked with efficiently scoring thousands of recalled items for downstream ranking. A key challenge is the train-serving discrepancy: pre-ranking models are trained only on exposed interactions, yet must score all recalled candidates -- including unexposed items -- during online serving. This mismatch not only induces severe sample selection bias but also degrades generalization, especially for long-tail content. Existing debiasing approaches typically rely on heuristics (e.g., negative sampling) or distillation from biased rankers, which either mislabel plausible unexposed items as negatives or propagate exposure bias into pseudo-labels. In this work, we propose Generative Pseudo-Labeling (GPL), a framework that leverages large language models (LLMs) to generate unbiased, content-aware pseudo-labels for unexposed items, explicitly aligning the training distribution with the online serving space. By offline generating user-specific interest anchors and matching them with candidates in a frozen semantic space, GPL provides high-quality supervision without adding online latency. Deployed in a large-scale production system, GPL improves click-through rate by 3.07%, while significantly enhancing recommendation diversity and long-tail item discovery.

Figures (4)

• Figure 1: Illustration of the cascade-based architecture in recommender systems. The unexposed candidates are items retrieved in early stages but discarded.
• Figure 2: Overview of the GPL framework. For each user, interest anchors are generated offline via an LLM in a discrete semantic space and matched with unexposed candidates to produce pseudo-labels. These are calibrated by confidence weights based on semantic coherence, historical consistency, and generation uncertainty. In the Dual-Label Fusion stage, pseudo-labeled unexposed items are combined with exposed items bearing actual labels to jointly train the pre-ranking model.
• Figure 3: We conduct a parameter sensitivity analysis of the balance weight $\boldsymbol{\lambda}$, the historical consistency weight $\lambda_1$, and the LLM-intrinsic confidence weight $\lambda_2$. The effects of the beam size $\boldsymbol{B}$ and temperature $\tau$ are also investigated. AUC performance is plotted for all parameters across their respective ranges.
• Figure 4: Fine-Grained Online A/B Analysis of GPL vs. Baseline. (a) Relative CTR lift (%) and absolute PVR change (percentage points, pt) across historical page-view (PV) buckets. (b) Category distribution of exposed items.