Abstract

Recommendation systems are widely used in e-commerce, social media, and content distribution, yet LLM-based recommendation workflows still face recurring challenges in data completeness, sample balance, and output stability. In addition, the conventional “user-platform” structure leaves limited room for user-side mediation of exposure and preference expression. This paper presents ABP, a workflow-level extension of i2Agent in the User–Agent–Platform setting. ABP contains three modules: Adaptive Description Enrichment (ADE), Batch-balanced Sampling Strategy (BSS), and Prompt-driven Workflow Optimization (PWO). ADE repairs missing or rigid item text with richer natural-language descriptions, BSS builds balanced comparative inputs for user profiling, and PWO strengthens multi-stage reasoning with structured output constraints. Experiments on four real-world datasets show that ABP achieves strong ranking results under the reported protocol. Across the 16 reported dataset–metric pairs, the five-run mean results of ABP show an average relative improvement of 24.62% over i2Agent, with especially large gains on Amazon Book and Amazon Movietv. Under a fixed five-run protocol, ABP shows limited run-to-run dispersion on Amazon Book, Amazon Movietv, and Yelp, while Goodreads exhibits comparatively larger but still bounded variation. Overall, these results suggest that carefully designed workflow improvements can improve LLM-based recommendation quality in the reported setting while maintaining the agent’s role as a user-side mediation layer in the User–Agent–Platform setting.

IPC Classification

Keywords