LERA: LLM-Enhanced Ad Auction Framework for Chatbots

A new research paper proposes LERA, a two-stage retrieve-then-generate auction framework for integrating advertising into LLM-based chatbots. The framework addresses limitations of current retrieval methods that rely solely on text embedding similarity, which can cause commercial misinterpretation and repetitive ad insertions. LERA first uses embedding-based coarse filtering to pre-select candidate advertisers, then queries the LLM with a designed prompt to generate logits for final selection. The work builds on prior research by Feizi et al. and Hajiaghayi et al., who outlined a retrieve-then-generate paradigm for lightweight ad insertion and payment determination. LERA aims to balance relevance, efficiency, and user experience in chatbot advertising.