LAMA: RUC and Stanford push LLM ad auctions down to the token level
机器之心 · wechat · 2026-09-23
Pew data shows AI summaries cut users' clicks on traditional search results from 15% to 8%, and ads are moving into LLM answers. Researchers from Renmin University's Gaoling School and Stanford propose LAMA (Latent Advertiser Mixture Auction), moving ad allocation into token-by-token generation — "generation as allocation."
- Advertisers report the value of possible continuations; the platform generates tokens while updating each advertiser's probability of winning exposure, settling after the answer completes.
- Local report verification plus generation-coupled payments/subsidies yield Markov DSIC: honest reporting is optimal at every state.
- Welfare guarantee bounds deviation from the organic answer, protecting user experience.
On 1,239 real commercial queries (Webis Generated Native Ads 2024) with Qwen3-14B, LAMA tops all four quality metrics, lifting platform revenue +10.7% and advertiser value +3.8% without hurting user experience. Perturbation tests confirm expected utility peaks at truthful reporting.
More from Venture
- AI startup funding frenzy: VCs see 8-10 hot deals a week, verbal offers in days — nmasc_ · 2026-09-23
- Being first doesn't win: Google was the 21st search engine, Apple the 15th phone maker — FinanceYF5 · 2026-09-23
- Juicebox got 30,000 signups in 3 days, then spent 6 brutal months making the product work — ycombinator · 2026-09-23
- Mercor Is Paying Users Directly for Their Personal Health and Fitness Data — amateurtrdr · 2026-09-23
- An AI-powered BD company to help robotics firms find high-value use cases — gan_chuang · 2026-09-23
- Startup thesis: custom-trained classifiers for regulated industries could be an open lane — n_sri_laasya · 2026-09-23