Apple Podcasts  ·  Spotify  ·  Buzzsprout

Paper A

Fine-Tuned LLM as a Complementary Predictor Improving Ads System

Hui Yang, Daiwei He, Kevin Jiang, Taejin Park et al. — 2026 — arXiv (Cornell University)

preprint  ·   ·  test this week

Key findings

• Adding a fine-tuned AI model that predicts which advertisers a user is likely to buy from — based on that user's past behavior — improved the overall ad system's performance both in offline tests and in live production at Pinterest.
• The AI model worked best when used as a helper, not a replacement. It fed predictions into two separate parts of the ad pipeline: which ads to even consider showing (retrieval), and how to rank those ads for conversion likelihood. Both parts got better.
• Using an AI that knows general world knowledge about brands and categories helped fill gaps that the traditional system — which only learns from past user-item interactions — couldn't cover on its own.
• The approach was designed to run within tight speed and cost limits, making it practical for a real platform rather than just a research demo.

Marketing implications

• If you run a platform with user behavior data (past purchases, clicks, conversions), you may be able to improve ad targeting by adding a fine-tuned AI layer that predicts which brands or categories a user is ready to buy from — even before they search for them.
• You don't need to rebuild your ad system from scratch to use AI. This paper shows that plugging an AI predictor into an existing system as an add-on — rather than replacing the whole thing — can still meaningfully improve results.
• If your ad system struggles with new users or sparse data (not enough history to know what someone wants), an AI trained on general brand and category knowledge could help bridge that gap.

Paper B

Incentive-Aware Multi-Fidelity Optimization for Generative Advertising in Large Language Models

Jiayuan Liu, Barry Wang, Jiarui Gan, Tonghan Wang et al. — 2026 — arXiv (Cornell University)

preprint  ·   ·  watchlist

Key findings

• When multiple advertisers try to influence what an AI chatbot recommends, cranking up one advertiser's influence too high actually backfires — the AI produces pushy, over-the-top responses that annoy users, reducing the total value for everyone. The sweet spot is a balanced mix of advertiser influence levels.
• Advertisers have an incentive to lie about how much they value a particular AI response configuration (to game the system in their favor). The researchers built a pricing mechanism — borrowed from classic auction theory — that makes it in each advertiser's best interest to tell the truth about their preferences, because lying doesn't help them.
• Figuring out the best influence configuration normally requires thousands of expensive AI queries. The proposed system (IAMFM) uses a 'cheap first, expensive later' approach: it quickly rules out bad configurations with low-cost tests before running expensive full evaluations on promising ones. In experiments, this approach outperformed a naive approach that spends the same effort on every option.
• No single algorithm is best in all situations: one version (elimination-based) works better when the computational budget is tight; another (model-based) works better when more budget is available. The paper gives practical guidance on which to use.

Marketing implications

• If you're building or buying into an AI recommendation or chatbot platform that carries sponsored content, demand to know how the platform handles the 'over-promotion' problem — too much advertiser influence makes AI responses feel like spam and drives users away. Ask for evidence that their system balances advertiser reach with user experience.
• When multiple brands compete for influence in an AI-generated response (think: hotel booking chatbots, retail recommendation engines), the pricing model matters enormously. A well-designed auction system — where advertisers pay based on their actual impact on outcomes — keeps the marketplace honest. Watch for platforms that start adopting auction-based pricing for AI placements.
• If you manage advertising budgets, be aware that AI-native ad formats (where your brand steers AI-generated text or images) are not just about spending more — spending too much on influence can literally degrade the ad quality. The right amount of influence, not the maximum, produces the best results.

Paper C

A Knowledge Graph and Deep Learning-Based Semantic Recommendation Database System for Advertisement Retrieval and Personalization

Tangtang Wang, Kaijie Zhang, K.-F. Liu — 2026 — Journal of Computer Science and Frontier Technologies

peer reviewed journal article  ·   ·  read now

Key findings

• The new system (KGSR-ADS) matched ads to users more accurately than the best existing method (GraphRec): it found roughly 5–6% more of the right ads in its top-10 results, and ranked the best ads slightly higher.
• KGSR-ADS was also 24% faster at returning results than GraphRec, meaning it could serve personalized ads in real time even under heavy traffic.
• The improvement came from combining three things that are usually separate: a map of how products, users, and ads relate to each other (knowledge graph); AI text understanding (LLM embeddings); and a fast search index (vector database). Each layer contributed to the final accuracy gains.
• The system handled 1.2 million users and 20 million interactions without falling over, suggesting it scales to production-level advertising workloads.

Marketing implications

• If you run a large ad platform or manage a product catalog with complex user segments, this paper shows that wiring together a knowledge graph (relationships between products, audiences, and ads) with LLM text embeddings and a fast vector search index is a viable engineering path — not just theory.
• The 24% latency reduction matters if you're losing ad impressions because your recommendation engine is too slow. This architecture suggests a concrete way to speed things up without sacrificing relevance.
• Before rebuilding your ad stack this way, run a live test first — the paper only tests on historical data, so you'd want to confirm the accuracy gains actually move clicks or conversions in your real environment.

Apple Podcasts  ·  Spotify  ·  Buzzsprout