Research map / Large language models / Reasoning and chain of thought

LLM-Auction: Generative Auction towards LLM-Native Advertising

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Efficient LLM serving · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LLM-Auction, a learning-based generative auction that post-trains an LLM with iterative reward-preference optimization (DPO on a pCTR-based reward) so the LLM itself allocates ad insertions in its response, evaluated in an LLM-as-judge simulation environment on Taobao ads.

Problem

Classic auction mechanisms do not apply when the auctioned object is a distribution over LLM outputs; existing LLM-native ad mechanisms either fix ad counts upfront (pre-generation) or add inference cost via multiple generations (post-generation), and ignore externalities.

Approach

Treat the allocation as preference alignment: the LLM is fine-tuned with DPO against a reward combining bid-weighted predicted CTR (from a response-aware pCTR model) and a response penalty, iterating between updating the pCTR model on simulated user clicks and updating the LLM. The trained LLM then directly implements the allocation with no extra inference cost, and the paper proves monotonicity a

Key results

• LLM-Auction reaches 170.86 revenue per query and 81.86 reward per query, versus 107.40 and -15.27 for MOSAIC, the strongest baseline (about 59.1% higher revenue).
• Removing response context from the pCTR model cuts revenue by 55% (to 76.45), with GAUC dropping from 0.6872 to 0.6391.
• Allocation monotonicity emerges during training, reaching monotone click-vs-bid behavior after 3 epochs.

Limitations

• Simulation environment is limited in scale and diversity, with LLM-simulated users rather than real user feedback.
• Uses offline multi-stage optimization with simple rewards; payment rule is first-price-style and IC only shown for ROI-constrained value maximizers.
• Only clickable ads are considered, simplifying attribution.

Experimental setup

Datasets: Taobao candidate ads (100 ads, 50 categories) synthetic user profiles and queries

Benchmarks: LLM-native advertising simulation environment

Metrics: revenue per query reward per query GAUC allocation monotonicity allocation continuity

Baselines: Pre-trained LLM RAG-Auction MOSAIC LLM-Auction (Oracle)

Builds on: DPO VCG generalized second-price auction RAG-Auction MOSAIC

Closest papers by meaning

• GenAuction: A Generative Auction for Online Advertising (AAAI 2025 · similarity 0.90)
• Ad Auctions for LLMs via Retrieval Augmented Generation (NeurIPS 2024 · similarity 0.90)
• Truthful Aggregation of LLMs with an Application to Online Advertising (NeurIPS 2025 · similarity 0.89)
• AuctionNet: A Novel Benchmark for Decision-Making in Large-Scale Games (NeurIPS 2024 · similarity 0.88)
• Optimal Auction Design in the Joint Advertising (ICML 2025 · similarity 0.86)
• Maximizing the Success Probability of Policy Allocations in Online Systems (AAAI 2024 · similarity 0.86)