Research map / Large language models / Retrieval and knowledge
Ad Auctions for LLMs via Retrieval Augmented Generation
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 segment auctions that allocate ads within LLM outputs via RAG-style probabilistic retrieval, with truthful pricing and a logarithmic social welfare objective, evaluated with GPT-4 simulations.
Problem
How to sell ad placements inside LLM textual outputs so that ad allocation is efficient, fair, incentive compatible and individually rational, while preserving output quality.
Approach
Treat RAG's retrieval probabilities as the allocation rule of an auction: bids reweight calibrated relevance, a Gumbel-perturbed second-price auction implements the randomized allocation, and each discourse segment gets its own auction (single or multi-ad). The single-ad rule is shown to maximize a logarithmic social welfare and to be truthful.
Key results
- Segment auction with replacement attains the highest social welfare (0.660) and relevance (0.688) among mechanisms in the 4-advertiser scenario, while Naive II has the highest revenue (0.379) vs 0.371 for the segment auction.
- Multi-allocation auction yields the highest output quality (similarity 0.715 for k=3 with the full output), above single-ad segment auctions at k=3 (0.700 and 0.702), though it has the lowest revenue (0.238).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Allow repeated ad selection across segments rather than requiring different ads | Segment auction · Four-ad scenario | social welfare: $.521$ -> $.660$ | Helped |
| Allow repeated ad selection across segments rather than requiring different ads | Segment auction · Four-ad scenario | revenue: $.333$ -> $.371$ | Helped |
| Allocate three ads to the entire output at once instead of one ad per segment | Segment auction · Four-ad scenario | revenue: $.371$ -> $.238$ | Hurt |
| Allocate three ads to the entire output at once instead of one ad per segment | Segment auction · Output similarity, first three segments | cosine similarity: $.700$ -> $.715$ | Helped |
| Integrate selected ads into the output with the LLM instead of appending their text | Segment auction · Output similarity, first three segments | cosine similarity: $.671$ -> $.700$ | Helped |
| Use relevance scores rather than disregarding relevance | Segment auction with replacement · Four-ad scenario | social welfare: $.508$ -> $.660$ | Helped |
Surprising findings
- Multi-allocation auctions achieve higher output quality but lower revenue, a revenue-quality tradeoff.
- Combinatorial segment auction payments can be negative, unlike standard VCG.
Limitations
- Experiments are simulations with a single fixed query and a small number of advertisers.
- Combinatorial auction payments may be negative and need rounding or truncation.
- Multi-ad combinatorial auction has exponential complexity O(n^k) in the number of ads.
Experimental setup
Datasets: hand-written advertiser descriptions for simulated scenarios
Benchmarks: simulated ad auction scenarios
Metrics: revenue social welfare relevance minimum social welfare embedding cosine similarity of output quality
Baselines: Naive I (concatenate ads at end) Naive II (ignores relevance) segment auction without replacement
Builds on: retrieval augmented generation Myerson's lemma Gumbel max trick Nash social welfare
Closest papers by meaning
- LLM-Auction: Generative Auction towards LLM-Native Advertising (NeurIPS 2026 · similarity 0.90)
- Truthful Aggregation of LLMs with an Application to Online Advertising (NeurIPS 2025 · similarity 0.88)
- GenAuction: A Generative Auction for Online Advertising (AAAI 2025 · similarity 0.87)
- AuctionNet: A Novel Benchmark for Decision-Making in Large-Scale Games (NeurIPS 2024 · similarity 0.85)
- Test-Time Compute Games (NeurIPS 2026 · similarity 0.84)
- Fairshare Data Pricing via Data Valuation for Large Language Models (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.