Apple Podcasts  ·  Spotify  ·  Buzzsprout

Paper A

Token-Level Advertising

Hanbing Liu, Bowei Zhang, Changyuan Yu, Yinyu Ye et al. — 2026 — arXiv

preprint  ·   ·  read now

Key findings

• Today's AI chatbots and search tools generate answers word by word (token by token). This paper proposes letting advertisers compete — via an auction — to influence which words get generated at each step, so brand mentions can appear naturally inside the AI's answer rather than in a separate ad slot.
• The proposed auction system (called LAMA) lets advertisers submit bids that say how much they value different ways the AI response could unfold. The platform then blends these preferences with its own language model to produce the final response, determining which advertiser 'wins' based on how the response actually develops.
• In simulation tests on real search queries, LAMA generated higher revenue and platform value than three comparison approaches (placing ads before generation, placing ads after generation, and selecting among full draft responses), while keeping the quality of AI-generated answers high.
• The system is designed so advertisers always have an incentive to bid honestly — gaming the system doesn't help them — a property the authors prove mathematically (called 'dominant-strategy incentive compatibility').

Marketing implications

• This research is not yet actionable for most marketers — it describes a future ad format that doesn't exist yet on any platform. Watch for it: if AI search platforms (like Perplexity, Google AI Overviews, or ChatGPT) adopt this kind of system, 'winning' an AI-generated mention could work like winning a search auction today, but tied to specific words in the answer rather than a slot above the results.
• If you run paid search today, start thinking about how your brand brief and keyword strategy might translate to 'how do I want an AI to naturally describe my product?' — because that framing is what token-level ad bidding would reward.
• For platform builders (ad tech, AI search startups): this paper provides a working theoretical blueprint for how to run a generation-native ad auction. It's early-stage but cites a concrete algorithmic approach that could inform product architecture.

Paper B

Astar: Learning to Propose Evolution Directions for Self-Evolving Industrial AI Systems

Jinxin Hu, Hao Deng, Haibo Xing, Lingyu Mu et al. — 2026 — arXiv

preprint  ·   ·  test this week

Key findings

• The Astar AI model (8 billion parameters) correctly suggested a useful improvement direction about 68% of the time, compared to ~32% for senior human experts and ~31% for the best general-purpose AI like GPT-5.5. In other words, the trained specialist AI was roughly twice as good at picking the right next step as either humans or off-the-shelf AI.
• Over two weeks, Astar ran 20 improvement cycles on Lazada's ad recommendation system completely on its own, with no human input needed. This pushed offline recommendation accuracy (Hitrate@200) up by 23.6%.
• In a live A/B test on the Lazada platform, Astar's automatically guided improvements led to 4.86% more total sales (GMV) and 1.82% more advertising revenue compared to the control group.
• General-purpose AI tools like GPT-5.5 gave generic, unhelpful suggestions for this task — the kind of advice that sounds reasonable but doesn't actually fit the specific system. Only the purpose-trained Astar, which learned from the system's own history, performed well.

Marketing implications

• If you run a large-scale ad platform and your data science team spends weeks deciding what to test next, this paper is a blueprint for automating that decision using your own experiment history — but only if you already log your model iterations carefully with code and performance data.
• Don't expect ChatGPT or any off-the-shelf AI to give you good strategic advice about improving your specific ad system — it will sound smart but miss the mark. Purpose-trained tools built from your own data are dramatically more effective.
• This is not a plug-and-play solution. The practical takeaway for most marketing technologists is: start systematically logging every model change and its outcome now, so that eventually you have the historical data needed to build something like this.

Paper C

Generative AI use before medical visits: disclosure-item responses, trust, and care-seeking behaviors in a cross-sectional social-media survey in Poland

Simona Wójcik, Anna Rulkiewicz, Justyna Domienik-Karłowicz — 2026 — Frontiers in Digital Health

peer reviewed journal article  ·   ·  use cautiously

Key findings

• About 84% of people in the survey said they had used an AI tool (like ChatGPT) to look up health information at least once — showing AI is already a common first step before going to the doctor.
• Among the 519 people who said they used AI before a medical visit, 82.5% said they did NOT tell their doctor about it — though it's important to note the survey couldn't confirm these people actually went to a doctor afterward.
• The two most common reasons people gave for not mentioning their AI use to a doctor were: fear the doctor would react badly (86% cited this) and not wanting to seem like they were challenging the doctor's expertise (75% cited this).
• Younger people and people who trusted AI more were more likely to have used AI before a visit and to have changed their health behavior because of it. Older people trusted AI much less.

Marketing implications

• If you market health products, services, or apps: most people are already using ChatGPT to research symptoms before seeing a doctor, but they're embarrassed to admit it to their physician. This is a real emotional pain point — messaging that normalizes and validates AI-assisted health research could resonate strongly with this audience.
• Fear of judgment is the #1 reason people don't mention their AI use. If you're building or marketing a health AI product, positioning it as something doctors approve of or partner with (rather than compete with) could help users feel less conflicted — and may increase both trust and word-of-mouth.
• Younger users trust AI for health far more than older users. Age-segment your messaging: lean into AI's credibility for under-35s, and focus on reliability and safety messaging for older audiences who are more skeptical.

Apple Podcasts  ·  Spotify  ·  Buzzsprout